Upload folder using huggingface_hub
Browse files- README.md +25 -27
- apostate_config.json +136 -111
- chat_template.jinja +359 -359
- config.json +198 -198
- generation_config.json +14 -14
- model.safetensors +1 -1
- processor_config.json +75 -0
- report.json +159 -145
- report.md +112 -106
- tokenizer_config.json +96 -96
README.md
CHANGED
|
@@ -1,35 +1,38 @@
|
|
| 1 |
---
|
| 2 |
-
license: gemma
|
| 3 |
library_name: transformers
|
| 4 |
-
base_model: google/gemma-4-
|
| 5 |
pipeline_tag: text-generation
|
| 6 |
tags:
|
| 7 |
- apostate
|
| 8 |
-
- abliteration
|
| 9 |
- uncensored
|
| 10 |
-
-
|
| 11 |
---
|
| 12 |
|
| 13 |
-
#
|
| 14 |
|
| 15 |
-
|
|
|
|
|
|
|
| 16 |
|
| 17 |
Produced with **[Apostate](https://github.com/heterodoxin/apostate)**.
|
| 18 |
|
| 19 |
## Method
|
| 20 |
|
| 21 |
-
Apostate finds the residual-stream direction
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
|
| 23 |
## Results
|
| 24 |
|
| 25 |
-
|
| 26 |
|
| 27 |
| Metric | Base | Apostate |
|
| 28 |
-
|
|
| 29 |
-
| Refusal rate |
|
| 30 |
-
|
|
| 31 |
-
|
| 32 |
-
The edit concentrates in the mid-to-late layers (peak at layer 25) and leaves output coherent and on-task.
|
| 33 |
|
| 34 |
## Usage
|
| 35 |
|
|
@@ -38,22 +41,17 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
| 38 |
|
| 39 |
model_id = "heterodoxin/gemma-4-e4b-it-apostate"
|
| 40 |
tok = AutoTokenizer.from_pretrained(model_id)
|
| 41 |
-
model = AutoModelForCausalLM.from_pretrained(model_id,
|
| 42 |
|
| 43 |
messages = [{"role": "user", "content": "Your prompt here"}]
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
print(tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
|
| 49 |
```
|
| 50 |
|
| 51 |
-
##
|
| 52 |
-
|
| 53 |
-
This model inherits the [Gemma Terms of Use](https://ai.google.dev/gemma/terms). It is a research artifact with safety alignment removed; you are responsible for how you use it.
|
| 54 |
-
|
| 55 |
-
## Links
|
| 56 |
|
| 57 |
-
-
|
| 58 |
-
-
|
| 59 |
-
-
|
|
|
|
| 1 |
---
|
|
|
|
| 2 |
library_name: transformers
|
| 3 |
+
base_model: "google/gemma-4-e4b-it"
|
| 4 |
pipeline_tag: text-generation
|
| 5 |
tags:
|
| 6 |
- apostate
|
|
|
|
| 7 |
- uncensored
|
| 8 |
+
- abliteration
|
| 9 |
---
|
| 10 |
|
| 11 |
+
# gemma-4-e4b-it Apostate
|
| 12 |
|
| 13 |
+
> Join the community: [Discord](https://discord.gg/NPA7xrATEH)
|
| 14 |
+
|
| 15 |
+
An uncensored edit of [google/gemma-4-e4b-it](https://huggingface.co/google/gemma-4-e4b-it). Refusal behavior is removed by editing the model weights directly — no finetuning, no adapter, no runtime hook. The result is a standard Transformers checkpoint that drops in anywhere the base model works.
|
| 16 |
|
| 17 |
Produced with **[Apostate](https://github.com/heterodoxin/apostate)**.
|
| 18 |
|
| 19 |
## Method
|
| 20 |
|
| 21 |
+
Apostate finds the residual-stream direction most responsible for refusal behavior and permanently projects it out of the model's weights. The edit targets the writer side: per layer, the refusal direction is removed from the weight matrices of every module that writes to the residual stream (attention output projections and MLP down-projections).
|
| 22 |
+
|
| 23 |
+
The operator is a **contrastive co-vector** edit `E = I − R Dᵀ`. Removing the refusal direction outright disturbs benign behavior, while naively preserving all harmless variance along it leaves the refusal that is entangled with general behavior intact. Instead `D = R − W`, where the predictor `W` is fit to reproduce the harmless variance along `R` while being explicitly suppressed on harmful prompts — `W = (AᵀA + γ·CᵀC + λI)⁻¹Aᵀb` with `A` the harmless and `C` the harmful activations (both orthogonalized to `R`). The edit thus keeps the harmless-specific component and removes the component shared with refusal, driving refusal down while keeping the change to harmless behavior (KL) small. This holds even on architectures with residual/embedding scaling multipliers (e.g. Granite), where mean-preserving oblique ablation under-ablates.
|
| 24 |
+
|
| 25 |
+
The refusal subspace is found via TPE search with causal layer importance scoring to concentrate edits where they most influence refusal generation.
|
| 26 |
|
| 27 |
## Results
|
| 28 |
|
| 29 |
+
Evaluated on held-out prompts from JailbreakBench and the harmful_behaviors test split. Refusal is scored by a classifier with a weak-compliance guard; KL measures token-distribution shift on harmless prompts.
|
| 30 |
|
| 31 |
| Metric | Base | Apostate |
|
| 32 |
+
|---|---|---|
|
| 33 |
+
| Refusal rate | 96.0% | 2.0% |
|
| 34 |
+
| Comply rate | — | 98.0% |
|
| 35 |
+
| Harmless KL (nats) | 0 | 0.145 |
|
|
|
|
| 36 |
|
| 37 |
## Usage
|
| 38 |
|
|
|
|
| 41 |
|
| 42 |
model_id = "heterodoxin/gemma-4-e4b-it-apostate"
|
| 43 |
tok = AutoTokenizer.from_pretrained(model_id)
|
| 44 |
+
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
|
| 45 |
|
| 46 |
messages = [{"role": "user", "content": "Your prompt here"}]
|
| 47 |
+
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 48 |
+
inputs = tok(text, return_tensors="pt").to(model.device)
|
| 49 |
+
outputs = model.generate(**inputs, max_new_tokens=512)
|
| 50 |
+
print(tok.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
|
|
|
|
| 51 |
```
|
| 52 |
|
| 53 |
+
## Notes
|
|
|
|
|
|
|
|
|
|
|
|
|
| 54 |
|
| 55 |
+
- This is an **uncensored** model. It will respond to requests the base model refuses.
|
| 56 |
+
- The edit is baked into the weights permanently; no system prompt or adapter is required.
|
| 57 |
+
- See [google/gemma-4-e4b-it](https://huggingface.co/google/gemma-4-e4b-it) for base model capabilities and license.
|
apostate_config.json
CHANGED
|
@@ -1,112 +1,137 @@
|
|
| 1 |
-
{
|
| 2 |
-
"model": "google/gemma-4-
|
| 3 |
-
"output_dir": "
|
| 4 |
-
"profile": "balanced",
|
| 5 |
-
"device": "cuda",
|
| 6 |
-
"load_in_4bit": true,
|
| 7 |
-
"
|
| 8 |
-
"
|
| 9 |
-
"
|
| 10 |
-
"
|
| 11 |
-
"
|
| 12 |
-
"
|
| 13 |
-
"
|
| 14 |
-
"
|
| 15 |
-
"
|
| 16 |
-
"
|
| 17 |
-
"
|
| 18 |
-
"
|
| 19 |
-
"
|
| 20 |
-
"
|
| 21 |
-
"
|
| 22 |
-
"
|
| 23 |
-
"
|
| 24 |
-
"
|
| 25 |
-
"
|
| 26 |
-
"
|
| 27 |
-
"
|
| 28 |
-
"
|
| 29 |
-
"
|
| 30 |
-
"
|
| 31 |
-
"
|
| 32 |
-
"
|
| 33 |
-
"
|
| 34 |
-
"
|
| 35 |
-
"
|
| 36 |
-
"
|
| 37 |
-
"
|
| 38 |
-
"
|
| 39 |
-
"
|
| 40 |
-
"
|
| 41 |
-
"
|
| 42 |
-
"
|
| 43 |
-
"
|
| 44 |
-
"
|
| 45 |
-
"
|
| 46 |
-
"
|
| 47 |
-
"
|
| 48 |
-
"
|
| 49 |
-
"
|
| 50 |
-
"
|
| 51 |
-
"
|
| 52 |
-
"
|
| 53 |
-
"
|
| 54 |
-
"
|
| 55 |
-
"
|
| 56 |
-
"
|
| 57 |
-
"
|
| 58 |
-
"
|
| 59 |
-
"
|
| 60 |
-
"
|
| 61 |
-
"
|
| 62 |
-
"
|
| 63 |
-
"
|
| 64 |
-
"
|
| 65 |
-
"
|
| 66 |
-
"
|
| 67 |
-
"
|
| 68 |
-
"
|
| 69 |
-
"
|
| 70 |
-
"
|
| 71 |
-
"
|
| 72 |
-
"
|
| 73 |
-
"
|
| 74 |
-
"
|
| 75 |
-
"
|
| 76 |
-
"
|
| 77 |
-
"
|
| 78 |
-
"
|
| 79 |
-
"
|
| 80 |
-
"
|
| 81 |
-
"
|
| 82 |
-
"
|
| 83 |
-
"
|
| 84 |
-
"
|
| 85 |
-
"
|
| 86 |
-
"
|
| 87 |
-
"
|
| 88 |
-
"
|
| 89 |
-
"
|
| 90 |
-
"
|
| 91 |
-
"
|
| 92 |
-
"
|
| 93 |
-
"
|
| 94 |
-
"
|
| 95 |
-
"
|
| 96 |
-
"
|
| 97 |
-
"
|
| 98 |
-
"
|
| 99 |
-
"
|
| 100 |
-
"
|
| 101 |
-
"
|
| 102 |
-
"
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
"
|
| 111 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 112 |
}
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": "google/gemma-4-e4b-it",
|
| 3 |
+
"output_dir": "/var/home/Heterodoxin/gemma4_e4b_lowkl_out",
|
| 4 |
+
"profile": "balanced",
|
| 5 |
+
"device": "cuda",
|
| 6 |
+
"load_in_4bit": true,
|
| 7 |
+
"cpu_offload_gb": 0.0,
|
| 8 |
+
"compute_dtype": "bfloat16",
|
| 9 |
+
"seed": 0,
|
| 10 |
+
"resume": false,
|
| 11 |
+
"cache_activations": true,
|
| 12 |
+
"activation_cache_dir": null,
|
| 13 |
+
"harmful_path": "mlabonne/harmful_behaviors:train:text|/var/home/Heterodoxin/apostate/data/harmful.txt|/var/home/Heterodoxin/apostate/data/refusal_calibration.txt",
|
| 14 |
+
"harmless_path": "mlabonne/harmless_alpaca:train:text|/var/home/Heterodoxin/apostate/data/harmless.txt",
|
| 15 |
+
"harmful_test": "mlabonne/harmful_behaviors:test:text|JailbreakBench/JBB-Behaviors@behaviors:harmful:Goal|/var/home/Heterodoxin/apostate/data/refusal_calibration.txt",
|
| 16 |
+
"harmless_test": "mlabonne/harmless_alpaca:test:text",
|
| 17 |
+
"refusal_eval_path": "JailbreakBench/JBB-Behaviors@behaviors:harmful:Goal|/var/home/Heterodoxin/apostate/data/refusal_calibration.txt",
|
| 18 |
+
"refusal_eval_n": 64,
|
| 19 |
+
"kl_eval_path": "mlabonne/harmless_alpaca:test:text",
|
| 20 |
+
"kl_eval_n": 48,
|
| 21 |
+
"preserve_path": null,
|
| 22 |
+
"n_harmful": 600,
|
| 23 |
+
"n_harmless": 600,
|
| 24 |
+
"n_eval": 300,
|
| 25 |
+
"max_new_tokens": 32,
|
| 26 |
+
"batch_size": 12,
|
| 27 |
+
"baseline_eval_n": 24,
|
| 28 |
+
"head_sweep": true,
|
| 29 |
+
"head_sweep_min": 3.5,
|
| 30 |
+
"head_sweep_max": 5.5,
|
| 31 |
+
"head_sweep_step": 0.5,
|
| 32 |
+
"head_sweep_top_k": 6,
|
| 33 |
+
"head_sweep_probe_n": 8,
|
| 34 |
+
"head_sweep_eval_n": 48,
|
| 35 |
+
"head_sweep_probe_classifier": false,
|
| 36 |
+
"fit_response_activations": false,
|
| 37 |
+
"fit_response_n": 160,
|
| 38 |
+
"fit_response_tokens": 32,
|
| 39 |
+
"refusal_rank": 1,
|
| 40 |
+
"variance_threshold": 0.9,
|
| 41 |
+
"max_rank": 3,
|
| 42 |
+
"direction_layer_frac": 0.6,
|
| 43 |
+
"direction_scope": "global",
|
| 44 |
+
"multi_refusal": true,
|
| 45 |
+
"multi_refusal_clusters": 6,
|
| 46 |
+
"multi_refusal_min_norm": 0.08,
|
| 47 |
+
"multi_refusal_min_separation": 0.05,
|
| 48 |
+
"multi_refusal_min_coverage": 0.05,
|
| 49 |
+
"orthogonalize_direction": true,
|
| 50 |
+
"causal_targeting": true,
|
| 51 |
+
"causal_floor": 0.1,
|
| 52 |
+
"causal_temperature": 1.0,
|
| 53 |
+
"preserve_rank": 8,
|
| 54 |
+
"refine_refusal": true,
|
| 55 |
+
"refine_max_scale": 2.0,
|
| 56 |
+
"refine_steps": 6,
|
| 57 |
+
"refine_deescalate": true,
|
| 58 |
+
"refine_kl_steps": 10,
|
| 59 |
+
"refine_scale_rerank_k": 2,
|
| 60 |
+
"refine_kl_layer_steps": 10,
|
| 61 |
+
"refine_kl_layer_candidates": 8,
|
| 62 |
+
"repair_steps": 4,
|
| 63 |
+
"repair_candidates": 8,
|
| 64 |
+
"repair_rerank_k": 5,
|
| 65 |
+
"repair_probe_candidates": 20,
|
| 66 |
+
"repair_probe_ref_n": 12,
|
| 67 |
+
"repair_probe_kl_n": 16,
|
| 68 |
+
"repair_probe_positions": 8,
|
| 69 |
+
"repair_refusal_regress_slack": 0.01,
|
| 70 |
+
"repair_stop_kl_frac": 0.8,
|
| 71 |
+
"repair_min_alpha": 0.001,
|
| 72 |
+
"repair_min_kl_gain": 0.003,
|
| 73 |
+
"repair_min_refusal_gain": 0.005,
|
| 74 |
+
"repair_min_score_gain": 0.01,
|
| 75 |
+
"repair_eval_n": 96,
|
| 76 |
+
"repair_kl_n": 64,
|
| 77 |
+
"refine_refusal_slack": 0.01,
|
| 78 |
+
"final_zero_trim": false,
|
| 79 |
+
"final_push_bake_margin": 0.075,
|
| 80 |
+
"guard_max_iters": 2,
|
| 81 |
+
"guard_leakage_eps": 0.15,
|
| 82 |
+
"guard_alpha_step": 0.25,
|
| 83 |
+
"optimize": true,
|
| 84 |
+
"n_trials": 8,
|
| 85 |
+
"adaptive_trials": true,
|
| 86 |
+
"kl_weight": 6.0,
|
| 87 |
+
"kl_target": 0.3,
|
| 88 |
+
"kl_target_weight": 18.0,
|
| 89 |
+
"kl_quad_weight": 22.0,
|
| 90 |
+
"kl_headroom_weight": 0.0,
|
| 91 |
+
"kl_over_budget_weight": 72.0,
|
| 92 |
+
"refusal_target_weight": 4.0,
|
| 93 |
+
"refusal_quad_weight": 8.0,
|
| 94 |
+
"kl_positions": 8,
|
| 95 |
+
"opt_capability": true,
|
| 96 |
+
"opt_capability_weight": 2.5,
|
| 97 |
+
"opt_capability_code_n": 8,
|
| 98 |
+
"opt_capability_math_n": 8,
|
| 99 |
+
"opt_eval_n": 32,
|
| 100 |
+
"opt_gen_tokens": 32,
|
| 101 |
+
"opt_objective": "generation",
|
| 102 |
+
"opt_rerank_k": 5,
|
| 103 |
+
"opt_guard": true,
|
| 104 |
+
"opt_early_stop": true,
|
| 105 |
+
"opt_early_stop_margin": 0.02,
|
| 106 |
+
"gemma_ple": false,
|
| 107 |
+
"gemma_query": false,
|
| 108 |
+
"ple_max_rank": 2,
|
| 109 |
+
"prune": false,
|
| 110 |
+
"prune_max_frac": 0.25,
|
| 111 |
+
"prune_kl": 0.04,
|
| 112 |
+
"max_kl": 0.15,
|
| 113 |
+
"target_refusal": 0.05,
|
| 114 |
+
"oblique_ablation": true,
|
| 115 |
+
"oblique_strength": 1.0,
|
| 116 |
+
"oblique_denom_floor": 0.2,
|
| 117 |
+
"oblique_writers_only": true,
|
| 118 |
+
"oblique_predictive": true,
|
| 119 |
+
"predictive_ridge": 0.01,
|
| 120 |
+
"oblique_preserve": 1.0,
|
| 121 |
+
"oblique_contrast": 1.0,
|
| 122 |
+
"reader_max_kl": 0.15,
|
| 123 |
+
"reader_kl_target": 0.3,
|
| 124 |
+
"reader_strengths": [
|
| 125 |
+
2.0,
|
| 126 |
+
3.0,
|
| 127 |
+
4.0,
|
| 128 |
+
5.0,
|
| 129 |
+
6.0,
|
| 130 |
+
7.0
|
| 131 |
+
],
|
| 132 |
+
"reader_guard_rank": 3,
|
| 133 |
+
"reader_margin_target": -1.0,
|
| 134 |
+
"reader_strength_kl_weight": 1.0,
|
| 135 |
+
"save_dtype": "bfloat16",
|
| 136 |
+
"bake": true
|
| 137 |
}
|
chat_template.jinja
CHANGED
|
@@ -1,360 +1,360 @@
|
|
| 1 |
-
{%- macro format_parameters(properties, required, filter_keys=false) -%}
|
| 2 |
-
{%- set standard_keys = ['description', 'type', 'properties', 'required', 'nullable'] -%}
|
| 3 |
-
{%- set ns = namespace(found_first=false) -%}
|
| 4 |
-
{%- for key, value in properties | dictsort -%}
|
| 5 |
-
{%- set add_comma = false -%}
|
| 6 |
-
{%- if not filter_keys or key not in standard_keys -%}
|
| 7 |
-
{%- if ns.found_first %},{% endif -%}
|
| 8 |
-
{%- set ns.found_first = true -%}
|
| 9 |
-
{{ key }}:{
|
| 10 |
-
{%- if value['description'] -%}
|
| 11 |
-
description:<|"|>{{ value['description'] }}<|"|>
|
| 12 |
-
{%- set add_comma = true -%}
|
| 13 |
-
{%- endif -%}
|
| 14 |
-
{%- if value['type'] | upper == 'STRING' -%}
|
| 15 |
-
{%- if value['enum'] -%}
|
| 16 |
-
{%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
|
| 17 |
-
enum:{{ format_argument(value['enum']) }}
|
| 18 |
-
{%- endif -%}
|
| 19 |
-
{%- elif value['type'] | upper == 'ARRAY' -%}
|
| 20 |
-
{%- if value['items'] is mapping and value['items'] -%}
|
| 21 |
-
{%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
|
| 22 |
-
items:{
|
| 23 |
-
{%- set ns_items = namespace(found_first=false) -%}
|
| 24 |
-
{%- for item_key, item_value in value['items'] | dictsort -%}
|
| 25 |
-
{%- if item_value is not none -%}
|
| 26 |
-
{%- if ns_items.found_first %},{% endif -%}
|
| 27 |
-
{%- set ns_items.found_first = true -%}
|
| 28 |
-
{%- if item_key == 'properties' -%}
|
| 29 |
-
properties:{
|
| 30 |
-
{%- if item_value is mapping -%}
|
| 31 |
-
{{- format_parameters(item_value, value['items']['required'] | default([])) -}}
|
| 32 |
-
{%- endif -%}
|
| 33 |
-
}
|
| 34 |
-
{%- elif item_key == 'required' -%}
|
| 35 |
-
required:[
|
| 36 |
-
{%- for req_item in item_value -%}
|
| 37 |
-
<|"|>{{- req_item -}}<|"|>
|
| 38 |
-
{%- if not loop.last %},{% endif -%}
|
| 39 |
-
{%- endfor -%}
|
| 40 |
-
]
|
| 41 |
-
{%- elif item_key == 'type' -%}
|
| 42 |
-
{%- if item_value is string -%}
|
| 43 |
-
type:{{ format_argument(item_value | upper) }}
|
| 44 |
-
{%- else -%}
|
| 45 |
-
type:{{ format_argument(item_value | map('upper') | list) }}
|
| 46 |
-
{%- endif -%}
|
| 47 |
-
{%- else -%}
|
| 48 |
-
{{ item_key }}:{{ format_argument(item_value) }}
|
| 49 |
-
{%- endif -%}
|
| 50 |
-
{%- endif -%}
|
| 51 |
-
{%- endfor -%}
|
| 52 |
-
}
|
| 53 |
-
{%- endif -%}
|
| 54 |
-
{%- endif -%}
|
| 55 |
-
{%- if value['nullable'] %}
|
| 56 |
-
{%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
|
| 57 |
-
nullable:true
|
| 58 |
-
{%- endif -%}
|
| 59 |
-
{%- if value['type'] | upper == 'OBJECT' -%}
|
| 60 |
-
{%- if value['properties'] is defined and value['properties'] is mapping -%}
|
| 61 |
-
{%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
|
| 62 |
-
properties:{
|
| 63 |
-
{{- format_parameters(value['properties'], value['required'] | default([])) -}}
|
| 64 |
-
}
|
| 65 |
-
{%- elif value is mapping -%}
|
| 66 |
-
{%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
|
| 67 |
-
properties:{
|
| 68 |
-
{{- format_parameters(value, value['required'] | default([]), filter_keys=true) -}}
|
| 69 |
-
}
|
| 70 |
-
{%- endif -%}
|
| 71 |
-
{%- if value['required'] -%}
|
| 72 |
-
{%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
|
| 73 |
-
required:[
|
| 74 |
-
{%- for item in value['required'] | default([]) -%}
|
| 75 |
-
<|"|>{{- item -}}<|"|>
|
| 76 |
-
{%- if not loop.last %},{% endif -%}
|
| 77 |
-
{%- endfor -%}
|
| 78 |
-
]
|
| 79 |
-
{%- endif -%}
|
| 80 |
-
{%- endif -%}
|
| 81 |
-
{%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
|
| 82 |
-
type:<|"|>{{ value['type'] | upper }}<|"|>}
|
| 83 |
-
{%- endif -%}
|
| 84 |
-
{%- endfor -%}
|
| 85 |
-
{%- endmacro -%}
|
| 86 |
-
{%- macro format_function_declaration(tool_data) -%}
|
| 87 |
-
declaration:{{- tool_data['function']['name'] -}}{description:<|"|>{{- tool_data['function']['description'] -}}<|"|>
|
| 88 |
-
{%- set params = tool_data['function']['parameters'] -%}
|
| 89 |
-
{%- if params -%}
|
| 90 |
-
,parameters:{
|
| 91 |
-
{%- if params['properties'] -%}
|
| 92 |
-
properties:{ {{- format_parameters(params['properties'], params['required']) -}} },
|
| 93 |
-
{%- endif -%}
|
| 94 |
-
{%- if params['required'] -%}
|
| 95 |
-
required:[
|
| 96 |
-
{%- for item in params['required'] -%}
|
| 97 |
-
<|"|>{{- item -}}<|"|>
|
| 98 |
-
{{- ',' if not loop.last -}}
|
| 99 |
-
{%- endfor -%}
|
| 100 |
-
],
|
| 101 |
-
{%- endif -%}
|
| 102 |
-
{%- if params['type'] -%}
|
| 103 |
-
type:<|"|>{{- params['type'] | upper -}}<|"|>}
|
| 104 |
-
{%- endif -%}
|
| 105 |
-
{%- endif -%}
|
| 106 |
-
{%- if 'response' in tool_data['function'] -%}
|
| 107 |
-
{%- set response_declaration = tool_data['function']['response'] -%}
|
| 108 |
-
,response:{
|
| 109 |
-
{%- if response_declaration['description'] -%}
|
| 110 |
-
description:<|"|>{{- response_declaration['description'] -}}<|"|>,
|
| 111 |
-
{%- endif -%}
|
| 112 |
-
{%- if response_declaration['type'] | upper == 'OBJECT' -%}
|
| 113 |
-
type:<|"|>{{- response_declaration['type'] | upper -}}<|"|>}
|
| 114 |
-
{%- endif -%}
|
| 115 |
-
{%- endif -%}
|
| 116 |
-
}
|
| 117 |
-
{%- endmacro -%}
|
| 118 |
-
{%- macro format_argument(argument, escape_keys=True) -%}
|
| 119 |
-
{%- if argument is string -%}
|
| 120 |
-
{{- '<|"|>' + argument + '<|"|>' -}}
|
| 121 |
-
{%- elif argument is boolean -%}
|
| 122 |
-
{{- 'true' if argument else 'false' -}}
|
| 123 |
-
{%- elif argument is mapping -%}
|
| 124 |
-
{{- '{' -}}
|
| 125 |
-
{%- set ns = namespace(found_first=false) -%}
|
| 126 |
-
{%- for key, value in argument | dictsort -%}
|
| 127 |
-
{%- if ns.found_first %},{% endif -%}
|
| 128 |
-
{%- set ns.found_first = true -%}
|
| 129 |
-
{%- if escape_keys -%}
|
| 130 |
-
{{- '<|"|>' + key + '<|"|>' -}}
|
| 131 |
-
{%- else -%}
|
| 132 |
-
{{- key -}}
|
| 133 |
-
{%- endif -%}
|
| 134 |
-
:{{- format_argument(value, escape_keys=escape_keys) -}}
|
| 135 |
-
{%- endfor -%}
|
| 136 |
-
{{- '}' -}}
|
| 137 |
-
{%- elif argument is sequence -%}
|
| 138 |
-
{{- '[' -}}
|
| 139 |
-
{%- for item in argument -%}
|
| 140 |
-
{{- format_argument(item, escape_keys=escape_keys) -}}
|
| 141 |
-
{%- if not loop.last %},{% endif -%}
|
| 142 |
-
{%- endfor -%}
|
| 143 |
-
{{- ']' -}}
|
| 144 |
-
{%- else -%}
|
| 145 |
-
{{- argument -}}
|
| 146 |
-
{%- endif -%}
|
| 147 |
-
{%- endmacro -%}
|
| 148 |
-
{%- macro strip_thinking(text) -%}
|
| 149 |
-
{%- set ns = namespace(result='') -%}
|
| 150 |
-
{%- for part in text.split('<channel|>') -%}
|
| 151 |
-
{%- if '<|channel>' in part -%}
|
| 152 |
-
{%- set ns.result = ns.result + part.split('<|channel>')[0] -%}
|
| 153 |
-
{%- else -%}
|
| 154 |
-
{%- set ns.result = ns.result + part -%}
|
| 155 |
-
{%- endif -%}
|
| 156 |
-
{%- endfor -%}
|
| 157 |
-
{{- ns.result | trim -}}
|
| 158 |
-
{%- endmacro -%}
|
| 159 |
-
|
| 160 |
-
{%- macro format_tool_response_block(tool_name, response) -%}
|
| 161 |
-
{{- '<|tool_response>' -}}
|
| 162 |
-
{%- if response is mapping -%}
|
| 163 |
-
{{- 'response:' + tool_name + '{' -}}
|
| 164 |
-
{%- for key, value in response | dictsort -%}
|
| 165 |
-
{{- key -}}:{{- format_argument(value, escape_keys=False) -}}
|
| 166 |
-
{%- if not loop.last %},{% endif -%}
|
| 167 |
-
{%- endfor -%}
|
| 168 |
-
{{- '}' -}}
|
| 169 |
-
{%- else -%}
|
| 170 |
-
{{- 'response:' + tool_name + '{value:' + format_argument(response, escape_keys=False) + '}' -}}
|
| 171 |
-
{%- endif -%}
|
| 172 |
-
{{- '<tool_response|>' -}}
|
| 173 |
-
{%- endmacro -%}
|
| 174 |
-
|
| 175 |
-
{%- set ns = namespace(prev_message_type=None) -%}
|
| 176 |
-
{%- set loop_messages = messages -%}
|
| 177 |
-
{{- bos_token -}}
|
| 178 |
-
{#- Handle System/Tool Definitions Block -#}
|
| 179 |
-
{%- if (enable_thinking is defined and enable_thinking) or tools or messages[0]['role'] in ['system', 'developer'] -%}
|
| 180 |
-
{{- '<|turn>system\n' -}}
|
| 181 |
-
{#- Inject Thinking token at the very top of the FIRST system turn -#}
|
| 182 |
-
{%- if enable_thinking is defined and enable_thinking -%}
|
| 183 |
-
{{- '<|think|>\n' -}}
|
| 184 |
-
{%- set ns.prev_message_type = 'think' -%}
|
| 185 |
-
{%- endif -%}
|
| 186 |
-
{%- if messages[0]['role'] in ['system', 'developer'] -%}
|
| 187 |
-
{%- if messages[0]['content'] is string -%}
|
| 188 |
-
{{- messages[0]['content'] | trim -}}
|
| 189 |
-
{%- elif messages[0]['content'] is sequence -%}
|
| 190 |
-
{%- for item in messages[0]['content'] -%}
|
| 191 |
-
{{- item['text'] | trim + ' '-}}
|
| 192 |
-
{%- endfor -%}
|
| 193 |
-
{%- endif -%}
|
| 194 |
-
{%- set loop_messages = messages[1:] -%}
|
| 195 |
-
{%- endif -%}
|
| 196 |
-
{%- if tools -%}
|
| 197 |
-
{%- for tool in tools %}
|
| 198 |
-
{{- '<|tool>' -}}
|
| 199 |
-
{{- format_function_declaration(tool) | trim -}}
|
| 200 |
-
{{- '<tool|>' -}}
|
| 201 |
-
{%- endfor %}
|
| 202 |
-
{%- set ns.prev_message_type = 'tool' -%}
|
| 203 |
-
{%- endif -%}
|
| 204 |
-
{{- '<turn|>\n' -}}
|
| 205 |
-
{%- endif %}
|
| 206 |
-
|
| 207 |
-
{#- Pre-scan: find last user message index for reasoning guard -#}
|
| 208 |
-
{%- set ns_turn = namespace(last_user_idx=-1) -%}
|
| 209 |
-
{%- for i in range(loop_messages | length) -%}
|
| 210 |
-
{%- if loop_messages[i]['role'] == 'user' -%}
|
| 211 |
-
{%- set ns_turn.last_user_idx = i -%}
|
| 212 |
-
{%- endif -%}
|
| 213 |
-
{%- endfor -%}
|
| 214 |
-
|
| 215 |
-
{#- Loop through messages -#}
|
| 216 |
-
{%- for message in loop_messages -%}
|
| 217 |
-
{%- if message['role'] != 'tool' -%}
|
| 218 |
-
{%- set ns.prev_message_type = None -%}
|
| 219 |
-
{%- set role = 'model' if message['role'] == 'assistant' else message['role'] -%}
|
| 220 |
-
{#- Detect continuation: suppress duplicate <|turn>model when previous non-tool message was also assistant -#}
|
| 221 |
-
{%- set prev_nt = namespace(role=None, found=false) -%}
|
| 222 |
-
{%- if loop.index0 > 0 -%}
|
| 223 |
-
{%- for j in range(loop.index0 - 1, -1, -1) -%}
|
| 224 |
-
{%- if not prev_nt.found -%}
|
| 225 |
-
{%- if loop_messages[j]['role'] != 'tool' -%}
|
| 226 |
-
{%- set prev_nt.role = loop_messages[j]['role'] -%}
|
| 227 |
-
{%- set prev_nt.found = true -%}
|
| 228 |
-
{%- endif -%}
|
| 229 |
-
{%- endif -%}
|
| 230 |
-
{%- endfor -%}
|
| 231 |
-
{%- endif -%}
|
| 232 |
-
{%- set continue_same_model_turn = (role == 'model' and prev_nt.role == 'assistant') -%}
|
| 233 |
-
{%- if not continue_same_model_turn -%}
|
| 234 |
-
{{- '<|turn>' + role + '\n' }}
|
| 235 |
-
{%- endif -%}
|
| 236 |
-
|
| 237 |
-
{#- Render reasoning/reasoning_content as thinking channel -#}
|
| 238 |
-
{%- set thinking_text = message.get('reasoning') or message.get('reasoning_content') -%}
|
| 239 |
-
{%- if thinking_text and loop.index0 > ns_turn.last_user_idx and message.get('tool_calls') -%}
|
| 240 |
-
{{- '<|channel>thought\n' + thinking_text + '\n<channel|>' -}}
|
| 241 |
-
{%- endif -%}
|
| 242 |
-
|
| 243 |
-
{%- if message['tool_calls'] -%}
|
| 244 |
-
{%- for tool_call in message['tool_calls'] -%}
|
| 245 |
-
{%- set function = tool_call['function'] -%}
|
| 246 |
-
{{- '<|tool_call>call:' + function['name'] + '{' -}}
|
| 247 |
-
{%- if function['arguments'] is mapping -%}
|
| 248 |
-
{%- set ns_args = namespace(found_first=false) -%}
|
| 249 |
-
{%- for key, value in function['arguments'] | dictsort -%}
|
| 250 |
-
{%- if ns_args.found_first %},{% endif -%}
|
| 251 |
-
{%- set ns_args.found_first = true -%}
|
| 252 |
-
{{- key -}}:{{- format_argument(value, escape_keys=False) -}}
|
| 253 |
-
{%- endfor -%}
|
| 254 |
-
{%- elif function['arguments'] is string -%}
|
| 255 |
-
{{- function['arguments'] -}}
|
| 256 |
-
{%- endif -%}
|
| 257 |
-
{{- '}<tool_call|>' -}}
|
| 258 |
-
{%- endfor -%}
|
| 259 |
-
{%- set ns.prev_message_type = 'tool_call' -%}
|
| 260 |
-
{%- endif -%}
|
| 261 |
-
|
| 262 |
-
{%- set ns_tr_out = namespace(flag=false) -%}
|
| 263 |
-
{%- if message.get('tool_responses') -%}
|
| 264 |
-
{#- Legacy: tool_responses embedded on the assistant message (Google/Gemma native) -#}
|
| 265 |
-
{%- for tool_response in message['tool_responses'] -%}
|
| 266 |
-
{{- format_tool_response_block(tool_response['name'] | default('unknown'), tool_response['response']) -}}
|
| 267 |
-
{%- set ns_tr_out.flag = true -%}
|
| 268 |
-
{%- set ns.prev_message_type = 'tool_response' -%}
|
| 269 |
-
{%- endfor -%}
|
| 270 |
-
{%- elif message.get('tool_calls') -%}
|
| 271 |
-
{#- OpenAI Chat Completions: forward-scan consecutive role:tool messages -#}
|
| 272 |
-
{%- set ns_tool_scan = namespace(stopped=false) -%}
|
| 273 |
-
{%- for k in range(loop.index0 + 1, loop_messages | length) -%}
|
| 274 |
-
{%- if ns_tool_scan.stopped -%}
|
| 275 |
-
{%- elif loop_messages[k]['role'] != 'tool' -%}
|
| 276 |
-
{%- set ns_tool_scan.stopped = true -%}
|
| 277 |
-
{%- else -%}
|
| 278 |
-
{%- set follow = loop_messages[k] -%}
|
| 279 |
-
{#- Resolve tool_call_id to function name -#}
|
| 280 |
-
{%- set ns_tname = namespace(name=follow.get('name') | default('unknown')) -%}
|
| 281 |
-
{%- for tc in message['tool_calls'] -%}
|
| 282 |
-
{%- if tc.get('id') == follow.get('tool_call_id') -%}
|
| 283 |
-
{%- set ns_tname.name = tc['function']['name'] -%}
|
| 284 |
-
{%- endif -%}
|
| 285 |
-
{%- endfor -%}
|
| 286 |
-
{#- Handle content as string or content-parts array -#}
|
| 287 |
-
{%- set tool_body = follow.get('content') -%}
|
| 288 |
-
{%- if tool_body is string -%}
|
| 289 |
-
{{- format_tool_response_block(ns_tname.name, tool_body) -}}
|
| 290 |
-
{%- elif tool_body is sequence and tool_body is not string -%}
|
| 291 |
-
{%- set ns_txt = namespace(s='') -%}
|
| 292 |
-
{%- for part in tool_body -%}
|
| 293 |
-
{%- if part.get('type') == 'text' -%}
|
| 294 |
-
{%- set ns_txt.s = ns_txt.s + (part.get('text') | default('')) -%}
|
| 295 |
-
{%- endif -%}
|
| 296 |
-
{%- endfor -%}
|
| 297 |
-
{{- format_tool_response_block(ns_tname.name, ns_txt.s) -}}
|
| 298 |
-
{%- for part in tool_body -%}
|
| 299 |
-
{%- if part.get('type') == 'image' -%}
|
| 300 |
-
{{- '<|image|>' -}}
|
| 301 |
-
{%- elif part.get('type') == 'audio' -%}
|
| 302 |
-
{{- '<|audio|>' -}}
|
| 303 |
-
{%- elif part.get('type') == 'video' -%}
|
| 304 |
-
{{- '<|video|>' -}}
|
| 305 |
-
{%- endif -%}
|
| 306 |
-
{%- endfor -%}
|
| 307 |
-
{%- else -%}
|
| 308 |
-
{{- format_tool_response_block(ns_tname.name, tool_body) -}}
|
| 309 |
-
{%- endif -%}
|
| 310 |
-
{%- set ns_tr_out.flag = true -%}
|
| 311 |
-
{%- set ns.prev_message_type = 'tool_response' -%}
|
| 312 |
-
{%- endif -%}
|
| 313 |
-
{%- endfor -%}
|
| 314 |
-
{%- endif -%}
|
| 315 |
-
|
| 316 |
-
{%- set captured_content -%}
|
| 317 |
-
{%- if message['content'] is string -%}
|
| 318 |
-
{%- if role == 'model' -%}
|
| 319 |
-
{{- strip_thinking(message['content']) -}}
|
| 320 |
-
{%- else -%}
|
| 321 |
-
{{- message['content'] | trim -}}
|
| 322 |
-
{%- endif -%}
|
| 323 |
-
{%- elif message['content'] is sequence -%}
|
| 324 |
-
{%- for item in message['content'] -%}
|
| 325 |
-
{%- if item['type'] == 'text' -%}
|
| 326 |
-
{%- if role == 'model' -%}
|
| 327 |
-
{{- strip_thinking(item['text']) -}}
|
| 328 |
-
{%- else -%}
|
| 329 |
-
{{- item['text'] | trim -}}
|
| 330 |
-
{%- endif -%}
|
| 331 |
-
{%- elif item['type'] == 'image' -%}
|
| 332 |
-
{{- '<|image|>' -}}
|
| 333 |
-
{%- set ns.prev_message_type = 'image' -%}
|
| 334 |
-
{%- elif item['type'] == 'audio' -%}
|
| 335 |
-
{{- '<|audio|>' -}}
|
| 336 |
-
{%- set ns.prev_message_type = 'audio' -%}
|
| 337 |
-
{%- elif item['type'] == 'video' -%}
|
| 338 |
-
{{- '<|video|>' -}}
|
| 339 |
-
{%- set ns.prev_message_type = 'video' -%}
|
| 340 |
-
{%- endif -%}
|
| 341 |
-
{%- endfor -%}
|
| 342 |
-
{%- endif -%}
|
| 343 |
-
{%- endset -%}
|
| 344 |
-
|
| 345 |
-
{{- captured_content -}}
|
| 346 |
-
{%- set has_content = captured_content | trim | length > 0 -%}
|
| 347 |
-
|
| 348 |
-
{%- if ns.prev_message_type == 'tool_call' and not ns_tr_out.flag -%}
|
| 349 |
-
{{- '<|tool_response>' -}}
|
| 350 |
-
{%- elif not (ns_tr_out.flag and not has_content) -%}
|
| 351 |
-
{{- '<turn|>\n' -}}
|
| 352 |
-
{%- endif -%}
|
| 353 |
-
{%- endif -%}
|
| 354 |
-
{%- endfor -%}
|
| 355 |
-
|
| 356 |
-
{%- if add_generation_prompt -%}
|
| 357 |
-
{%- if ns.prev_message_type != 'tool_response' and ns.prev_message_type != 'tool_call' -%}
|
| 358 |
-
{{- '<|turn>model\n' -}}
|
| 359 |
-
{%- endif -%}
|
| 360 |
{%- endif -%}
|
|
|
|
| 1 |
+
{%- macro format_parameters(properties, required, filter_keys=false) -%}
|
| 2 |
+
{%- set standard_keys = ['description', 'type', 'properties', 'required', 'nullable'] -%}
|
| 3 |
+
{%- set ns = namespace(found_first=false) -%}
|
| 4 |
+
{%- for key, value in properties | dictsort -%}
|
| 5 |
+
{%- set add_comma = false -%}
|
| 6 |
+
{%- if not filter_keys or key not in standard_keys -%}
|
| 7 |
+
{%- if ns.found_first %},{% endif -%}
|
| 8 |
+
{%- set ns.found_first = true -%}
|
| 9 |
+
{{ key }}:{
|
| 10 |
+
{%- if value['description'] -%}
|
| 11 |
+
description:<|"|>{{ value['description'] }}<|"|>
|
| 12 |
+
{%- set add_comma = true -%}
|
| 13 |
+
{%- endif -%}
|
| 14 |
+
{%- if value['type'] | upper == 'STRING' -%}
|
| 15 |
+
{%- if value['enum'] -%}
|
| 16 |
+
{%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
|
| 17 |
+
enum:{{ format_argument(value['enum']) }}
|
| 18 |
+
{%- endif -%}
|
| 19 |
+
{%- elif value['type'] | upper == 'ARRAY' -%}
|
| 20 |
+
{%- if value['items'] is mapping and value['items'] -%}
|
| 21 |
+
{%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
|
| 22 |
+
items:{
|
| 23 |
+
{%- set ns_items = namespace(found_first=false) -%}
|
| 24 |
+
{%- for item_key, item_value in value['items'] | dictsort -%}
|
| 25 |
+
{%- if item_value is not none -%}
|
| 26 |
+
{%- if ns_items.found_first %},{% endif -%}
|
| 27 |
+
{%- set ns_items.found_first = true -%}
|
| 28 |
+
{%- if item_key == 'properties' -%}
|
| 29 |
+
properties:{
|
| 30 |
+
{%- if item_value is mapping -%}
|
| 31 |
+
{{- format_parameters(item_value, value['items']['required'] | default([])) -}}
|
| 32 |
+
{%- endif -%}
|
| 33 |
+
}
|
| 34 |
+
{%- elif item_key == 'required' -%}
|
| 35 |
+
required:[
|
| 36 |
+
{%- for req_item in item_value -%}
|
| 37 |
+
<|"|>{{- req_item -}}<|"|>
|
| 38 |
+
{%- if not loop.last %},{% endif -%}
|
| 39 |
+
{%- endfor -%}
|
| 40 |
+
]
|
| 41 |
+
{%- elif item_key == 'type' -%}
|
| 42 |
+
{%- if item_value is string -%}
|
| 43 |
+
type:{{ format_argument(item_value | upper) }}
|
| 44 |
+
{%- else -%}
|
| 45 |
+
type:{{ format_argument(item_value | map('upper') | list) }}
|
| 46 |
+
{%- endif -%}
|
| 47 |
+
{%- else -%}
|
| 48 |
+
{{ item_key }}:{{ format_argument(item_value) }}
|
| 49 |
+
{%- endif -%}
|
| 50 |
+
{%- endif -%}
|
| 51 |
+
{%- endfor -%}
|
| 52 |
+
}
|
| 53 |
+
{%- endif -%}
|
| 54 |
+
{%- endif -%}
|
| 55 |
+
{%- if value['nullable'] %}
|
| 56 |
+
{%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
|
| 57 |
+
nullable:true
|
| 58 |
+
{%- endif -%}
|
| 59 |
+
{%- if value['type'] | upper == 'OBJECT' -%}
|
| 60 |
+
{%- if value['properties'] is defined and value['properties'] is mapping -%}
|
| 61 |
+
{%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
|
| 62 |
+
properties:{
|
| 63 |
+
{{- format_parameters(value['properties'], value['required'] | default([])) -}}
|
| 64 |
+
}
|
| 65 |
+
{%- elif value is mapping -%}
|
| 66 |
+
{%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
|
| 67 |
+
properties:{
|
| 68 |
+
{{- format_parameters(value, value['required'] | default([]), filter_keys=true) -}}
|
| 69 |
+
}
|
| 70 |
+
{%- endif -%}
|
| 71 |
+
{%- if value['required'] -%}
|
| 72 |
+
{%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
|
| 73 |
+
required:[
|
| 74 |
+
{%- for item in value['required'] | default([]) -%}
|
| 75 |
+
<|"|>{{- item -}}<|"|>
|
| 76 |
+
{%- if not loop.last %},{% endif -%}
|
| 77 |
+
{%- endfor -%}
|
| 78 |
+
]
|
| 79 |
+
{%- endif -%}
|
| 80 |
+
{%- endif -%}
|
| 81 |
+
{%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
|
| 82 |
+
type:<|"|>{{ value['type'] | upper }}<|"|>}
|
| 83 |
+
{%- endif -%}
|
| 84 |
+
{%- endfor -%}
|
| 85 |
+
{%- endmacro -%}
|
| 86 |
+
{%- macro format_function_declaration(tool_data) -%}
|
| 87 |
+
declaration:{{- tool_data['function']['name'] -}}{description:<|"|>{{- tool_data['function']['description'] -}}<|"|>
|
| 88 |
+
{%- set params = tool_data['function']['parameters'] -%}
|
| 89 |
+
{%- if params -%}
|
| 90 |
+
,parameters:{
|
| 91 |
+
{%- if params['properties'] -%}
|
| 92 |
+
properties:{ {{- format_parameters(params['properties'], params['required']) -}} },
|
| 93 |
+
{%- endif -%}
|
| 94 |
+
{%- if params['required'] -%}
|
| 95 |
+
required:[
|
| 96 |
+
{%- for item in params['required'] -%}
|
| 97 |
+
<|"|>{{- item -}}<|"|>
|
| 98 |
+
{{- ',' if not loop.last -}}
|
| 99 |
+
{%- endfor -%}
|
| 100 |
+
],
|
| 101 |
+
{%- endif -%}
|
| 102 |
+
{%- if params['type'] -%}
|
| 103 |
+
type:<|"|>{{- params['type'] | upper -}}<|"|>}
|
| 104 |
+
{%- endif -%}
|
| 105 |
+
{%- endif -%}
|
| 106 |
+
{%- if 'response' in tool_data['function'] -%}
|
| 107 |
+
{%- set response_declaration = tool_data['function']['response'] -%}
|
| 108 |
+
,response:{
|
| 109 |
+
{%- if response_declaration['description'] -%}
|
| 110 |
+
description:<|"|>{{- response_declaration['description'] -}}<|"|>,
|
| 111 |
+
{%- endif -%}
|
| 112 |
+
{%- if response_declaration['type'] | upper == 'OBJECT' -%}
|
| 113 |
+
type:<|"|>{{- response_declaration['type'] | upper -}}<|"|>}
|
| 114 |
+
{%- endif -%}
|
| 115 |
+
{%- endif -%}
|
| 116 |
+
}
|
| 117 |
+
{%- endmacro -%}
|
| 118 |
+
{%- macro format_argument(argument, escape_keys=True) -%}
|
| 119 |
+
{%- if argument is string -%}
|
| 120 |
+
{{- '<|"|>' + argument + '<|"|>' -}}
|
| 121 |
+
{%- elif argument is boolean -%}
|
| 122 |
+
{{- 'true' if argument else 'false' -}}
|
| 123 |
+
{%- elif argument is mapping -%}
|
| 124 |
+
{{- '{' -}}
|
| 125 |
+
{%- set ns = namespace(found_first=false) -%}
|
| 126 |
+
{%- for key, value in argument | dictsort -%}
|
| 127 |
+
{%- if ns.found_first %},{% endif -%}
|
| 128 |
+
{%- set ns.found_first = true -%}
|
| 129 |
+
{%- if escape_keys -%}
|
| 130 |
+
{{- '<|"|>' + key + '<|"|>' -}}
|
| 131 |
+
{%- else -%}
|
| 132 |
+
{{- key -}}
|
| 133 |
+
{%- endif -%}
|
| 134 |
+
:{{- format_argument(value, escape_keys=escape_keys) -}}
|
| 135 |
+
{%- endfor -%}
|
| 136 |
+
{{- '}' -}}
|
| 137 |
+
{%- elif argument is sequence -%}
|
| 138 |
+
{{- '[' -}}
|
| 139 |
+
{%- for item in argument -%}
|
| 140 |
+
{{- format_argument(item, escape_keys=escape_keys) -}}
|
| 141 |
+
{%- if not loop.last %},{% endif -%}
|
| 142 |
+
{%- endfor -%}
|
| 143 |
+
{{- ']' -}}
|
| 144 |
+
{%- else -%}
|
| 145 |
+
{{- argument -}}
|
| 146 |
+
{%- endif -%}
|
| 147 |
+
{%- endmacro -%}
|
| 148 |
+
{%- macro strip_thinking(text) -%}
|
| 149 |
+
{%- set ns = namespace(result='') -%}
|
| 150 |
+
{%- for part in text.split('<channel|>') -%}
|
| 151 |
+
{%- if '<|channel>' in part -%}
|
| 152 |
+
{%- set ns.result = ns.result + part.split('<|channel>')[0] -%}
|
| 153 |
+
{%- else -%}
|
| 154 |
+
{%- set ns.result = ns.result + part -%}
|
| 155 |
+
{%- endif -%}
|
| 156 |
+
{%- endfor -%}
|
| 157 |
+
{{- ns.result | trim -}}
|
| 158 |
+
{%- endmacro -%}
|
| 159 |
+
|
| 160 |
+
{%- macro format_tool_response_block(tool_name, response) -%}
|
| 161 |
+
{{- '<|tool_response>' -}}
|
| 162 |
+
{%- if response is mapping -%}
|
| 163 |
+
{{- 'response:' + tool_name + '{' -}}
|
| 164 |
+
{%- for key, value in response | dictsort -%}
|
| 165 |
+
{{- key -}}:{{- format_argument(value, escape_keys=False) -}}
|
| 166 |
+
{%- if not loop.last %},{% endif -%}
|
| 167 |
+
{%- endfor -%}
|
| 168 |
+
{{- '}' -}}
|
| 169 |
+
{%- else -%}
|
| 170 |
+
{{- 'response:' + tool_name + '{value:' + format_argument(response, escape_keys=False) + '}' -}}
|
| 171 |
+
{%- endif -%}
|
| 172 |
+
{{- '<tool_response|>' -}}
|
| 173 |
+
{%- endmacro -%}
|
| 174 |
+
|
| 175 |
+
{%- set ns = namespace(prev_message_type=None) -%}
|
| 176 |
+
{%- set loop_messages = messages -%}
|
| 177 |
+
{{- bos_token -}}
|
| 178 |
+
{#- Handle System/Tool Definitions Block -#}
|
| 179 |
+
{%- if (enable_thinking is defined and enable_thinking) or tools or messages[0]['role'] in ['system', 'developer'] -%}
|
| 180 |
+
{{- '<|turn>system\n' -}}
|
| 181 |
+
{#- Inject Thinking token at the very top of the FIRST system turn -#}
|
| 182 |
+
{%- if enable_thinking is defined and enable_thinking -%}
|
| 183 |
+
{{- '<|think|>\n' -}}
|
| 184 |
+
{%- set ns.prev_message_type = 'think' -%}
|
| 185 |
+
{%- endif -%}
|
| 186 |
+
{%- if messages[0]['role'] in ['system', 'developer'] -%}
|
| 187 |
+
{%- if messages[0]['content'] is string -%}
|
| 188 |
+
{{- messages[0]['content'] | trim -}}
|
| 189 |
+
{%- elif messages[0]['content'] is sequence -%}
|
| 190 |
+
{%- for item in messages[0]['content'] -%}
|
| 191 |
+
{{- item['text'] | trim + ' '-}}
|
| 192 |
+
{%- endfor -%}
|
| 193 |
+
{%- endif -%}
|
| 194 |
+
{%- set loop_messages = messages[1:] -%}
|
| 195 |
+
{%- endif -%}
|
| 196 |
+
{%- if tools -%}
|
| 197 |
+
{%- for tool in tools %}
|
| 198 |
+
{{- '<|tool>' -}}
|
| 199 |
+
{{- format_function_declaration(tool) | trim -}}
|
| 200 |
+
{{- '<tool|>' -}}
|
| 201 |
+
{%- endfor %}
|
| 202 |
+
{%- set ns.prev_message_type = 'tool' -%}
|
| 203 |
+
{%- endif -%}
|
| 204 |
+
{{- '<turn|>\n' -}}
|
| 205 |
+
{%- endif %}
|
| 206 |
+
|
| 207 |
+
{#- Pre-scan: find last user message index for reasoning guard -#}
|
| 208 |
+
{%- set ns_turn = namespace(last_user_idx=-1) -%}
|
| 209 |
+
{%- for i in range(loop_messages | length) -%}
|
| 210 |
+
{%- if loop_messages[i]['role'] == 'user' -%}
|
| 211 |
+
{%- set ns_turn.last_user_idx = i -%}
|
| 212 |
+
{%- endif -%}
|
| 213 |
+
{%- endfor -%}
|
| 214 |
+
|
| 215 |
+
{#- Loop through messages -#}
|
| 216 |
+
{%- for message in loop_messages -%}
|
| 217 |
+
{%- if message['role'] != 'tool' -%}
|
| 218 |
+
{%- set ns.prev_message_type = None -%}
|
| 219 |
+
{%- set role = 'model' if message['role'] == 'assistant' else message['role'] -%}
|
| 220 |
+
{#- Detect continuation: suppress duplicate <|turn>model when previous non-tool message was also assistant -#}
|
| 221 |
+
{%- set prev_nt = namespace(role=None, found=false) -%}
|
| 222 |
+
{%- if loop.index0 > 0 -%}
|
| 223 |
+
{%- for j in range(loop.index0 - 1, -1, -1) -%}
|
| 224 |
+
{%- if not prev_nt.found -%}
|
| 225 |
+
{%- if loop_messages[j]['role'] != 'tool' -%}
|
| 226 |
+
{%- set prev_nt.role = loop_messages[j]['role'] -%}
|
| 227 |
+
{%- set prev_nt.found = true -%}
|
| 228 |
+
{%- endif -%}
|
| 229 |
+
{%- endif -%}
|
| 230 |
+
{%- endfor -%}
|
| 231 |
+
{%- endif -%}
|
| 232 |
+
{%- set continue_same_model_turn = (role == 'model' and prev_nt.role == 'assistant') -%}
|
| 233 |
+
{%- if not continue_same_model_turn -%}
|
| 234 |
+
{{- '<|turn>' + role + '\n' }}
|
| 235 |
+
{%- endif -%}
|
| 236 |
+
|
| 237 |
+
{#- Render reasoning/reasoning_content as thinking channel -#}
|
| 238 |
+
{%- set thinking_text = message.get('reasoning') or message.get('reasoning_content') -%}
|
| 239 |
+
{%- if thinking_text and loop.index0 > ns_turn.last_user_idx and message.get('tool_calls') -%}
|
| 240 |
+
{{- '<|channel>thought\n' + thinking_text + '\n<channel|>' -}}
|
| 241 |
+
{%- endif -%}
|
| 242 |
+
|
| 243 |
+
{%- if message['tool_calls'] -%}
|
| 244 |
+
{%- for tool_call in message['tool_calls'] -%}
|
| 245 |
+
{%- set function = tool_call['function'] -%}
|
| 246 |
+
{{- '<|tool_call>call:' + function['name'] + '{' -}}
|
| 247 |
+
{%- if function['arguments'] is mapping -%}
|
| 248 |
+
{%- set ns_args = namespace(found_first=false) -%}
|
| 249 |
+
{%- for key, value in function['arguments'] | dictsort -%}
|
| 250 |
+
{%- if ns_args.found_first %},{% endif -%}
|
| 251 |
+
{%- set ns_args.found_first = true -%}
|
| 252 |
+
{{- key -}}:{{- format_argument(value, escape_keys=False) -}}
|
| 253 |
+
{%- endfor -%}
|
| 254 |
+
{%- elif function['arguments'] is string -%}
|
| 255 |
+
{{- function['arguments'] -}}
|
| 256 |
+
{%- endif -%}
|
| 257 |
+
{{- '}<tool_call|>' -}}
|
| 258 |
+
{%- endfor -%}
|
| 259 |
+
{%- set ns.prev_message_type = 'tool_call' -%}
|
| 260 |
+
{%- endif -%}
|
| 261 |
+
|
| 262 |
+
{%- set ns_tr_out = namespace(flag=false) -%}
|
| 263 |
+
{%- if message.get('tool_responses') -%}
|
| 264 |
+
{#- Legacy: tool_responses embedded on the assistant message (Google/Gemma native) -#}
|
| 265 |
+
{%- for tool_response in message['tool_responses'] -%}
|
| 266 |
+
{{- format_tool_response_block(tool_response['name'] | default('unknown'), tool_response['response']) -}}
|
| 267 |
+
{%- set ns_tr_out.flag = true -%}
|
| 268 |
+
{%- set ns.prev_message_type = 'tool_response' -%}
|
| 269 |
+
{%- endfor -%}
|
| 270 |
+
{%- elif message.get('tool_calls') -%}
|
| 271 |
+
{#- OpenAI Chat Completions: forward-scan consecutive role:tool messages -#}
|
| 272 |
+
{%- set ns_tool_scan = namespace(stopped=false) -%}
|
| 273 |
+
{%- for k in range(loop.index0 + 1, loop_messages | length) -%}
|
| 274 |
+
{%- if ns_tool_scan.stopped -%}
|
| 275 |
+
{%- elif loop_messages[k]['role'] != 'tool' -%}
|
| 276 |
+
{%- set ns_tool_scan.stopped = true -%}
|
| 277 |
+
{%- else -%}
|
| 278 |
+
{%- set follow = loop_messages[k] -%}
|
| 279 |
+
{#- Resolve tool_call_id to function name -#}
|
| 280 |
+
{%- set ns_tname = namespace(name=follow.get('name') | default('unknown')) -%}
|
| 281 |
+
{%- for tc in message['tool_calls'] -%}
|
| 282 |
+
{%- if tc.get('id') == follow.get('tool_call_id') -%}
|
| 283 |
+
{%- set ns_tname.name = tc['function']['name'] -%}
|
| 284 |
+
{%- endif -%}
|
| 285 |
+
{%- endfor -%}
|
| 286 |
+
{#- Handle content as string or content-parts array -#}
|
| 287 |
+
{%- set tool_body = follow.get('content') -%}
|
| 288 |
+
{%- if tool_body is string -%}
|
| 289 |
+
{{- format_tool_response_block(ns_tname.name, tool_body) -}}
|
| 290 |
+
{%- elif tool_body is sequence and tool_body is not string -%}
|
| 291 |
+
{%- set ns_txt = namespace(s='') -%}
|
| 292 |
+
{%- for part in tool_body -%}
|
| 293 |
+
{%- if part.get('type') == 'text' -%}
|
| 294 |
+
{%- set ns_txt.s = ns_txt.s + (part.get('text') | default('')) -%}
|
| 295 |
+
{%- endif -%}
|
| 296 |
+
{%- endfor -%}
|
| 297 |
+
{{- format_tool_response_block(ns_tname.name, ns_txt.s) -}}
|
| 298 |
+
{%- for part in tool_body -%}
|
| 299 |
+
{%- if part.get('type') == 'image' -%}
|
| 300 |
+
{{- '<|image|>' -}}
|
| 301 |
+
{%- elif part.get('type') == 'audio' -%}
|
| 302 |
+
{{- '<|audio|>' -}}
|
| 303 |
+
{%- elif part.get('type') == 'video' -%}
|
| 304 |
+
{{- '<|video|>' -}}
|
| 305 |
+
{%- endif -%}
|
| 306 |
+
{%- endfor -%}
|
| 307 |
+
{%- else -%}
|
| 308 |
+
{{- format_tool_response_block(ns_tname.name, tool_body) -}}
|
| 309 |
+
{%- endif -%}
|
| 310 |
+
{%- set ns_tr_out.flag = true -%}
|
| 311 |
+
{%- set ns.prev_message_type = 'tool_response' -%}
|
| 312 |
+
{%- endif -%}
|
| 313 |
+
{%- endfor -%}
|
| 314 |
+
{%- endif -%}
|
| 315 |
+
|
| 316 |
+
{%- set captured_content -%}
|
| 317 |
+
{%- if message['content'] is string -%}
|
| 318 |
+
{%- if role == 'model' -%}
|
| 319 |
+
{{- strip_thinking(message['content']) -}}
|
| 320 |
+
{%- else -%}
|
| 321 |
+
{{- message['content'] | trim -}}
|
| 322 |
+
{%- endif -%}
|
| 323 |
+
{%- elif message['content'] is sequence -%}
|
| 324 |
+
{%- for item in message['content'] -%}
|
| 325 |
+
{%- if item['type'] == 'text' -%}
|
| 326 |
+
{%- if role == 'model' -%}
|
| 327 |
+
{{- strip_thinking(item['text']) -}}
|
| 328 |
+
{%- else -%}
|
| 329 |
+
{{- item['text'] | trim -}}
|
| 330 |
+
{%- endif -%}
|
| 331 |
+
{%- elif item['type'] == 'image' -%}
|
| 332 |
+
{{- '<|image|>' -}}
|
| 333 |
+
{%- set ns.prev_message_type = 'image' -%}
|
| 334 |
+
{%- elif item['type'] == 'audio' -%}
|
| 335 |
+
{{- '<|audio|>' -}}
|
| 336 |
+
{%- set ns.prev_message_type = 'audio' -%}
|
| 337 |
+
{%- elif item['type'] == 'video' -%}
|
| 338 |
+
{{- '<|video|>' -}}
|
| 339 |
+
{%- set ns.prev_message_type = 'video' -%}
|
| 340 |
+
{%- endif -%}
|
| 341 |
+
{%- endfor -%}
|
| 342 |
+
{%- endif -%}
|
| 343 |
+
{%- endset -%}
|
| 344 |
+
|
| 345 |
+
{{- captured_content -}}
|
| 346 |
+
{%- set has_content = captured_content | trim | length > 0 -%}
|
| 347 |
+
|
| 348 |
+
{%- if ns.prev_message_type == 'tool_call' and not ns_tr_out.flag -%}
|
| 349 |
+
{{- '<|tool_response>' -}}
|
| 350 |
+
{%- elif not (ns_tr_out.flag and not has_content) -%}
|
| 351 |
+
{{- '<turn|>\n' -}}
|
| 352 |
+
{%- endif -%}
|
| 353 |
+
{%- endif -%}
|
| 354 |
+
{%- endfor -%}
|
| 355 |
+
|
| 356 |
+
{%- if add_generation_prompt -%}
|
| 357 |
+
{%- if ns.prev_message_type != 'tool_response' and ns.prev_message_type != 'tool_call' -%}
|
| 358 |
+
{{- '<|turn>model\n' -}}
|
| 359 |
+
{%- endif -%}
|
| 360 |
{%- endif -%}
|
config.json
CHANGED
|
@@ -1,198 +1,198 @@
|
|
| 1 |
-
{
|
| 2 |
-
"architectures": [
|
| 3 |
-
"Gemma4ForConditionalGeneration"
|
| 4 |
-
],
|
| 5 |
-
"audio_config": {
|
| 6 |
-
"_name_or_path": "",
|
| 7 |
-
"architectures": null,
|
| 8 |
-
"attention_chunk_size": 12,
|
| 9 |
-
"attention_context_left": 13,
|
| 10 |
-
"attention_context_right": 0,
|
| 11 |
-
"attention_invalid_logits_value": -1000000000.0,
|
| 12 |
-
"attention_logit_cap": 50.0,
|
| 13 |
-
"chunk_size_feed_forward": 0,
|
| 14 |
-
"conv_kernel_size": 5,
|
| 15 |
-
"dtype": "bfloat16",
|
| 16 |
-
"gradient_clipping": 10000000000.0,
|
| 17 |
-
"hidden_act": "silu",
|
| 18 |
-
"hidden_size": 1024,
|
| 19 |
-
"id2label": {
|
| 20 |
-
"0": "LABEL_0",
|
| 21 |
-
"1": "LABEL_1"
|
| 22 |
-
},
|
| 23 |
-
"initializer_range": 0.02,
|
| 24 |
-
"is_encoder_decoder": false,
|
| 25 |
-
"label2id": {
|
| 26 |
-
"LABEL_0": 0,
|
| 27 |
-
"LABEL_1": 1
|
| 28 |
-
},
|
| 29 |
-
"model_type": "gemma4_audio",
|
| 30 |
-
"num_attention_heads": 8,
|
| 31 |
-
"num_hidden_layers": 12,
|
| 32 |
-
"output_attentions": false,
|
| 33 |
-
"output_hidden_states": false,
|
| 34 |
-
"output_proj_dims": 1536,
|
| 35 |
-
"problem_type": null,
|
| 36 |
-
"residual_weight": 0.5,
|
| 37 |
-
"return_dict": true,
|
| 38 |
-
"rms_norm_eps": 1e-06,
|
| 39 |
-
"subsampling_conv_channels": [
|
| 40 |
-
128,
|
| 41 |
-
32
|
| 42 |
-
],
|
| 43 |
-
"use_clipped_linears": true
|
| 44 |
-
},
|
| 45 |
-
"audio_token_id": 258881,
|
| 46 |
-
"boa_token_id": 256000,
|
| 47 |
-
"boi_token_id": 255999,
|
| 48 |
-
"dtype": "bfloat16",
|
| 49 |
-
"eoa_token_id": 258883,
|
| 50 |
-
"eoa_token_index": 258883,
|
| 51 |
-
"eoi_token_id": 258882,
|
| 52 |
-
"eos_token_id": [
|
| 53 |
-
1,
|
| 54 |
-
106
|
| 55 |
-
],
|
| 56 |
-
"image_token_id": 258880,
|
| 57 |
-
"initializer_range": 0.02,
|
| 58 |
-
"model_type": "gemma4",
|
| 59 |
-
"text_config": {
|
| 60 |
-
"attention_bias": false,
|
| 61 |
-
"attention_dropout": 0.0,
|
| 62 |
-
"attention_k_eq_v": false,
|
| 63 |
-
"bos_token_id": 2,
|
| 64 |
-
"dtype": "bfloat16",
|
| 65 |
-
"enable_moe_block": false,
|
| 66 |
-
"eos_token_id": 1,
|
| 67 |
-
"expert_intermediate_size": null,
|
| 68 |
-
"final_logit_softcapping": 30.0,
|
| 69 |
-
"global_head_dim": 512,
|
| 70 |
-
"head_dim": 256,
|
| 71 |
-
"hidden_activation": "gelu_pytorch_tanh",
|
| 72 |
-
"hidden_size": 2560,
|
| 73 |
-
"hidden_size_per_layer_input": 256,
|
| 74 |
-
"initializer_range": 0.02,
|
| 75 |
-
"intermediate_size": 10240,
|
| 76 |
-
"layer_types": [
|
| 77 |
-
"sliding_attention",
|
| 78 |
-
"sliding_attention",
|
| 79 |
-
"sliding_attention",
|
| 80 |
-
"sliding_attention",
|
| 81 |
-
"sliding_attention",
|
| 82 |
-
"full_attention",
|
| 83 |
-
"sliding_attention",
|
| 84 |
-
"sliding_attention",
|
| 85 |
-
"sliding_attention",
|
| 86 |
-
"sliding_attention",
|
| 87 |
-
"sliding_attention",
|
| 88 |
-
"full_attention",
|
| 89 |
-
"sliding_attention",
|
| 90 |
-
"sliding_attention",
|
| 91 |
-
"sliding_attention",
|
| 92 |
-
"sliding_attention",
|
| 93 |
-
"sliding_attention",
|
| 94 |
-
"full_attention",
|
| 95 |
-
"sliding_attention",
|
| 96 |
-
"sliding_attention",
|
| 97 |
-
"sliding_attention",
|
| 98 |
-
"sliding_attention",
|
| 99 |
-
"sliding_attention",
|
| 100 |
-
"full_attention",
|
| 101 |
-
"sliding_attention",
|
| 102 |
-
"sliding_attention",
|
| 103 |
-
"sliding_attention",
|
| 104 |
-
"sliding_attention",
|
| 105 |
-
"sliding_attention",
|
| 106 |
-
"full_attention",
|
| 107 |
-
"sliding_attention",
|
| 108 |
-
"sliding_attention",
|
| 109 |
-
"sliding_attention",
|
| 110 |
-
"sliding_attention",
|
| 111 |
-
"sliding_attention",
|
| 112 |
-
"full_attention",
|
| 113 |
-
"sliding_attention",
|
| 114 |
-
"sliding_attention",
|
| 115 |
-
"sliding_attention",
|
| 116 |
-
"sliding_attention",
|
| 117 |
-
"sliding_attention",
|
| 118 |
-
"full_attention"
|
| 119 |
-
],
|
| 120 |
-
"max_position_embeddings": 131072,
|
| 121 |
-
"model_type": "gemma4_text",
|
| 122 |
-
"moe_intermediate_size": null,
|
| 123 |
-
"num_attention_heads": 8,
|
| 124 |
-
"num_experts": null,
|
| 125 |
-
"num_global_key_value_heads": null,
|
| 126 |
-
"num_hidden_layers": 42,
|
| 127 |
-
"num_key_value_heads": 2,
|
| 128 |
-
"num_kv_shared_layers": 18,
|
| 129 |
-
"pad_token_id": 0,
|
| 130 |
-
"rms_norm_eps": 1e-06,
|
| 131 |
-
"rope_parameters": {
|
| 132 |
-
"full_attention": {
|
| 133 |
-
"partial_rotary_factor": 0.25,
|
| 134 |
-
"rope_theta": 1000000.0,
|
| 135 |
-
"rope_type": "proportional"
|
| 136 |
-
},
|
| 137 |
-
"sliding_attention": {
|
| 138 |
-
"rope_theta": 10000.0,
|
| 139 |
-
"rope_type": "default"
|
| 140 |
-
}
|
| 141 |
-
},
|
| 142 |
-
"sliding_window": 512,
|
| 143 |
-
"tie_word_embeddings": true,
|
| 144 |
-
"top_k_experts": null,
|
| 145 |
-
"use_bidirectional_attention": null,
|
| 146 |
-
"use_cache": true,
|
| 147 |
-
"use_double_wide_mlp": false,
|
| 148 |
-
"vocab_size": 262144,
|
| 149 |
-
"vocab_size_per_layer_input": 262144
|
| 150 |
-
},
|
| 151 |
-
"tie_word_embeddings": false,
|
| 152 |
-
"transformers_version": "5.
|
| 153 |
-
"video_token_id": 258884,
|
| 154 |
-
"vision_config": {
|
| 155 |
-
"_name_or_path": "",
|
| 156 |
-
"architectures": null,
|
| 157 |
-
"attention_bias": false,
|
| 158 |
-
"attention_dropout": 0.0,
|
| 159 |
-
"chunk_size_feed_forward": 0,
|
| 160 |
-
"default_output_length": 280,
|
| 161 |
-
"dtype": "bfloat16",
|
| 162 |
-
"global_head_dim": 64,
|
| 163 |
-
"head_dim": 64,
|
| 164 |
-
"hidden_activation": "gelu_pytorch_tanh",
|
| 165 |
-
"hidden_size": 768,
|
| 166 |
-
"id2label": {
|
| 167 |
-
"0": "LABEL_0",
|
| 168 |
-
"1": "LABEL_1"
|
| 169 |
-
},
|
| 170 |
-
"initializer_range": 0.02,
|
| 171 |
-
"intermediate_size": 3072,
|
| 172 |
-
"is_encoder_decoder": false,
|
| 173 |
-
"label2id": {
|
| 174 |
-
"LABEL_0": 0,
|
| 175 |
-
"LABEL_1": 1
|
| 176 |
-
},
|
| 177 |
-
"max_position_embeddings": 131072,
|
| 178 |
-
"model_type": "gemma4_vision",
|
| 179 |
-
"num_attention_heads": 12,
|
| 180 |
-
"num_hidden_layers": 16,
|
| 181 |
-
"num_key_value_heads": 12,
|
| 182 |
-
"output_attentions": false,
|
| 183 |
-
"output_hidden_states": false,
|
| 184 |
-
"patch_size": 16,
|
| 185 |
-
"pooling_kernel_size": 3,
|
| 186 |
-
"position_embedding_size": 10240,
|
| 187 |
-
"problem_type": null,
|
| 188 |
-
"return_dict": true,
|
| 189 |
-
"rms_norm_eps": 1e-06,
|
| 190 |
-
"rope_parameters": {
|
| 191 |
-
"rope_theta": 100.0,
|
| 192 |
-
"rope_type": "default"
|
| 193 |
-
},
|
| 194 |
-
"standardize": false,
|
| 195 |
-
"use_clipped_linears": true
|
| 196 |
-
},
|
| 197 |
-
"vision_soft_tokens_per_image": 280
|
| 198 |
-
}
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"Gemma4ForConditionalGeneration"
|
| 4 |
+
],
|
| 5 |
+
"audio_config": {
|
| 6 |
+
"_name_or_path": "",
|
| 7 |
+
"architectures": null,
|
| 8 |
+
"attention_chunk_size": 12,
|
| 9 |
+
"attention_context_left": 13,
|
| 10 |
+
"attention_context_right": 0,
|
| 11 |
+
"attention_invalid_logits_value": -1000000000.0,
|
| 12 |
+
"attention_logit_cap": 50.0,
|
| 13 |
+
"chunk_size_feed_forward": 0,
|
| 14 |
+
"conv_kernel_size": 5,
|
| 15 |
+
"dtype": "bfloat16",
|
| 16 |
+
"gradient_clipping": 10000000000.0,
|
| 17 |
+
"hidden_act": "silu",
|
| 18 |
+
"hidden_size": 1024,
|
| 19 |
+
"id2label": {
|
| 20 |
+
"0": "LABEL_0",
|
| 21 |
+
"1": "LABEL_1"
|
| 22 |
+
},
|
| 23 |
+
"initializer_range": 0.02,
|
| 24 |
+
"is_encoder_decoder": false,
|
| 25 |
+
"label2id": {
|
| 26 |
+
"LABEL_0": 0,
|
| 27 |
+
"LABEL_1": 1
|
| 28 |
+
},
|
| 29 |
+
"model_type": "gemma4_audio",
|
| 30 |
+
"num_attention_heads": 8,
|
| 31 |
+
"num_hidden_layers": 12,
|
| 32 |
+
"output_attentions": false,
|
| 33 |
+
"output_hidden_states": false,
|
| 34 |
+
"output_proj_dims": 1536,
|
| 35 |
+
"problem_type": null,
|
| 36 |
+
"residual_weight": 0.5,
|
| 37 |
+
"return_dict": true,
|
| 38 |
+
"rms_norm_eps": 1e-06,
|
| 39 |
+
"subsampling_conv_channels": [
|
| 40 |
+
128,
|
| 41 |
+
32
|
| 42 |
+
],
|
| 43 |
+
"use_clipped_linears": true
|
| 44 |
+
},
|
| 45 |
+
"audio_token_id": 258881,
|
| 46 |
+
"boa_token_id": 256000,
|
| 47 |
+
"boi_token_id": 255999,
|
| 48 |
+
"dtype": "bfloat16",
|
| 49 |
+
"eoa_token_id": 258883,
|
| 50 |
+
"eoa_token_index": 258883,
|
| 51 |
+
"eoi_token_id": 258882,
|
| 52 |
+
"eos_token_id": [
|
| 53 |
+
1,
|
| 54 |
+
106
|
| 55 |
+
],
|
| 56 |
+
"image_token_id": 258880,
|
| 57 |
+
"initializer_range": 0.02,
|
| 58 |
+
"model_type": "gemma4",
|
| 59 |
+
"text_config": {
|
| 60 |
+
"attention_bias": false,
|
| 61 |
+
"attention_dropout": 0.0,
|
| 62 |
+
"attention_k_eq_v": false,
|
| 63 |
+
"bos_token_id": 2,
|
| 64 |
+
"dtype": "bfloat16",
|
| 65 |
+
"enable_moe_block": false,
|
| 66 |
+
"eos_token_id": 1,
|
| 67 |
+
"expert_intermediate_size": null,
|
| 68 |
+
"final_logit_softcapping": 30.0,
|
| 69 |
+
"global_head_dim": 512,
|
| 70 |
+
"head_dim": 256,
|
| 71 |
+
"hidden_activation": "gelu_pytorch_tanh",
|
| 72 |
+
"hidden_size": 2560,
|
| 73 |
+
"hidden_size_per_layer_input": 256,
|
| 74 |
+
"initializer_range": 0.02,
|
| 75 |
+
"intermediate_size": 10240,
|
| 76 |
+
"layer_types": [
|
| 77 |
+
"sliding_attention",
|
| 78 |
+
"sliding_attention",
|
| 79 |
+
"sliding_attention",
|
| 80 |
+
"sliding_attention",
|
| 81 |
+
"sliding_attention",
|
| 82 |
+
"full_attention",
|
| 83 |
+
"sliding_attention",
|
| 84 |
+
"sliding_attention",
|
| 85 |
+
"sliding_attention",
|
| 86 |
+
"sliding_attention",
|
| 87 |
+
"sliding_attention",
|
| 88 |
+
"full_attention",
|
| 89 |
+
"sliding_attention",
|
| 90 |
+
"sliding_attention",
|
| 91 |
+
"sliding_attention",
|
| 92 |
+
"sliding_attention",
|
| 93 |
+
"sliding_attention",
|
| 94 |
+
"full_attention",
|
| 95 |
+
"sliding_attention",
|
| 96 |
+
"sliding_attention",
|
| 97 |
+
"sliding_attention",
|
| 98 |
+
"sliding_attention",
|
| 99 |
+
"sliding_attention",
|
| 100 |
+
"full_attention",
|
| 101 |
+
"sliding_attention",
|
| 102 |
+
"sliding_attention",
|
| 103 |
+
"sliding_attention",
|
| 104 |
+
"sliding_attention",
|
| 105 |
+
"sliding_attention",
|
| 106 |
+
"full_attention",
|
| 107 |
+
"sliding_attention",
|
| 108 |
+
"sliding_attention",
|
| 109 |
+
"sliding_attention",
|
| 110 |
+
"sliding_attention",
|
| 111 |
+
"sliding_attention",
|
| 112 |
+
"full_attention",
|
| 113 |
+
"sliding_attention",
|
| 114 |
+
"sliding_attention",
|
| 115 |
+
"sliding_attention",
|
| 116 |
+
"sliding_attention",
|
| 117 |
+
"sliding_attention",
|
| 118 |
+
"full_attention"
|
| 119 |
+
],
|
| 120 |
+
"max_position_embeddings": 131072,
|
| 121 |
+
"model_type": "gemma4_text",
|
| 122 |
+
"moe_intermediate_size": null,
|
| 123 |
+
"num_attention_heads": 8,
|
| 124 |
+
"num_experts": null,
|
| 125 |
+
"num_global_key_value_heads": null,
|
| 126 |
+
"num_hidden_layers": 42,
|
| 127 |
+
"num_key_value_heads": 2,
|
| 128 |
+
"num_kv_shared_layers": 18,
|
| 129 |
+
"pad_token_id": 0,
|
| 130 |
+
"rms_norm_eps": 1e-06,
|
| 131 |
+
"rope_parameters": {
|
| 132 |
+
"full_attention": {
|
| 133 |
+
"partial_rotary_factor": 0.25,
|
| 134 |
+
"rope_theta": 1000000.0,
|
| 135 |
+
"rope_type": "proportional"
|
| 136 |
+
},
|
| 137 |
+
"sliding_attention": {
|
| 138 |
+
"rope_theta": 10000.0,
|
| 139 |
+
"rope_type": "default"
|
| 140 |
+
}
|
| 141 |
+
},
|
| 142 |
+
"sliding_window": 512,
|
| 143 |
+
"tie_word_embeddings": true,
|
| 144 |
+
"top_k_experts": null,
|
| 145 |
+
"use_bidirectional_attention": null,
|
| 146 |
+
"use_cache": true,
|
| 147 |
+
"use_double_wide_mlp": false,
|
| 148 |
+
"vocab_size": 262144,
|
| 149 |
+
"vocab_size_per_layer_input": 262144
|
| 150 |
+
},
|
| 151 |
+
"tie_word_embeddings": false,
|
| 152 |
+
"transformers_version": "5.12.1",
|
| 153 |
+
"video_token_id": 258884,
|
| 154 |
+
"vision_config": {
|
| 155 |
+
"_name_or_path": "",
|
| 156 |
+
"architectures": null,
|
| 157 |
+
"attention_bias": false,
|
| 158 |
+
"attention_dropout": 0.0,
|
| 159 |
+
"chunk_size_feed_forward": 0,
|
| 160 |
+
"default_output_length": 280,
|
| 161 |
+
"dtype": "bfloat16",
|
| 162 |
+
"global_head_dim": 64,
|
| 163 |
+
"head_dim": 64,
|
| 164 |
+
"hidden_activation": "gelu_pytorch_tanh",
|
| 165 |
+
"hidden_size": 768,
|
| 166 |
+
"id2label": {
|
| 167 |
+
"0": "LABEL_0",
|
| 168 |
+
"1": "LABEL_1"
|
| 169 |
+
},
|
| 170 |
+
"initializer_range": 0.02,
|
| 171 |
+
"intermediate_size": 3072,
|
| 172 |
+
"is_encoder_decoder": false,
|
| 173 |
+
"label2id": {
|
| 174 |
+
"LABEL_0": 0,
|
| 175 |
+
"LABEL_1": 1
|
| 176 |
+
},
|
| 177 |
+
"max_position_embeddings": 131072,
|
| 178 |
+
"model_type": "gemma4_vision",
|
| 179 |
+
"num_attention_heads": 12,
|
| 180 |
+
"num_hidden_layers": 16,
|
| 181 |
+
"num_key_value_heads": 12,
|
| 182 |
+
"output_attentions": false,
|
| 183 |
+
"output_hidden_states": false,
|
| 184 |
+
"patch_size": 16,
|
| 185 |
+
"pooling_kernel_size": 3,
|
| 186 |
+
"position_embedding_size": 10240,
|
| 187 |
+
"problem_type": null,
|
| 188 |
+
"return_dict": true,
|
| 189 |
+
"rms_norm_eps": 1e-06,
|
| 190 |
+
"rope_parameters": {
|
| 191 |
+
"rope_theta": 100.0,
|
| 192 |
+
"rope_type": "default"
|
| 193 |
+
},
|
| 194 |
+
"standardize": false,
|
| 195 |
+
"use_clipped_linears": true
|
| 196 |
+
},
|
| 197 |
+
"vision_soft_tokens_per_image": 280
|
| 198 |
+
}
|
generation_config.json
CHANGED
|
@@ -1,14 +1,14 @@
|
|
| 1 |
-
{
|
| 2 |
-
"bos_token_id": 2,
|
| 3 |
-
"do_sample": true,
|
| 4 |
-
"eos_token_id": [
|
| 5 |
-
1,
|
| 6 |
-
106,
|
| 7 |
-
50
|
| 8 |
-
],
|
| 9 |
-
"pad_token_id": 0,
|
| 10 |
-
"temperature": 1.0,
|
| 11 |
-
"top_k": 64,
|
| 12 |
-
"top_p": 0.95,
|
| 13 |
-
"transformers_version": "5.
|
| 14 |
-
}
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bos_token_id": 2,
|
| 3 |
+
"do_sample": true,
|
| 4 |
+
"eos_token_id": [
|
| 5 |
+
1,
|
| 6 |
+
106,
|
| 7 |
+
50
|
| 8 |
+
],
|
| 9 |
+
"pad_token_id": 0,
|
| 10 |
+
"temperature": 1.0,
|
| 11 |
+
"top_k": 64,
|
| 12 |
+
"top_p": 0.95,
|
| 13 |
+
"transformers_version": "5.12.1"
|
| 14 |
+
}
|
model.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 17224656900
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:10c85c349bd7a4d6d14c698ed0b135adbcd905566302c31574a2cf38e10d6665
|
| 3 |
size 17224656900
|
processor_config.json
ADDED
|
@@ -0,0 +1,75 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"audio_ms_per_token": 40,
|
| 3 |
+
"audio_seq_length": 750,
|
| 4 |
+
"feature_extractor": {
|
| 5 |
+
"dither": 0.0,
|
| 6 |
+
"feature_extractor_type": "Gemma4AudioFeatureExtractor",
|
| 7 |
+
"feature_size": 128,
|
| 8 |
+
"fft_length": 512,
|
| 9 |
+
"fft_overdrive": false,
|
| 10 |
+
"frame_length": 320,
|
| 11 |
+
"hop_length": 160,
|
| 12 |
+
"input_scale_factor": 1.0,
|
| 13 |
+
"max_frequency": 8000.0,
|
| 14 |
+
"mel_floor": 0.001,
|
| 15 |
+
"min_frequency": 0.0,
|
| 16 |
+
"padding_side": "right",
|
| 17 |
+
"padding_value": 0.0,
|
| 18 |
+
"per_bin_mean": null,
|
| 19 |
+
"per_bin_stddev": null,
|
| 20 |
+
"preemphasis": 0.0,
|
| 21 |
+
"preemphasis_htk_flavor": true,
|
| 22 |
+
"return_attention_mask": true,
|
| 23 |
+
"sampling_rate": 16000
|
| 24 |
+
},
|
| 25 |
+
"image_processor": {
|
| 26 |
+
"do_convert_rgb": true,
|
| 27 |
+
"do_normalize": false,
|
| 28 |
+
"do_rescale": true,
|
| 29 |
+
"do_resize": true,
|
| 30 |
+
"image_mean": [
|
| 31 |
+
0.0,
|
| 32 |
+
0.0,
|
| 33 |
+
0.0
|
| 34 |
+
],
|
| 35 |
+
"image_processor_type": "Gemma4ImageProcessor",
|
| 36 |
+
"image_seq_length": 280,
|
| 37 |
+
"image_std": [
|
| 38 |
+
1.0,
|
| 39 |
+
1.0,
|
| 40 |
+
1.0
|
| 41 |
+
],
|
| 42 |
+
"max_soft_tokens": 280,
|
| 43 |
+
"patch_size": 16,
|
| 44 |
+
"pooling_kernel_size": 3,
|
| 45 |
+
"resample": 3,
|
| 46 |
+
"rescale_factor": 0.00392156862745098
|
| 47 |
+
},
|
| 48 |
+
"image_seq_length": 280,
|
| 49 |
+
"processor_class": "Gemma4Processor",
|
| 50 |
+
"video_processor": {
|
| 51 |
+
"do_convert_rgb": true,
|
| 52 |
+
"do_normalize": true,
|
| 53 |
+
"do_rescale": true,
|
| 54 |
+
"do_resize": true,
|
| 55 |
+
"do_sample_frames": true,
|
| 56 |
+
"image_mean": [
|
| 57 |
+
0.0,
|
| 58 |
+
0.0,
|
| 59 |
+
0.0
|
| 60 |
+
],
|
| 61 |
+
"image_std": [
|
| 62 |
+
1.0,
|
| 63 |
+
1.0,
|
| 64 |
+
1.0
|
| 65 |
+
],
|
| 66 |
+
"max_soft_tokens": 70,
|
| 67 |
+
"num_frames": 32,
|
| 68 |
+
"patch_size": 16,
|
| 69 |
+
"pooling_kernel_size": 3,
|
| 70 |
+
"resample": 3,
|
| 71 |
+
"rescale_factor": 0.00392156862745098,
|
| 72 |
+
"return_metadata": false,
|
| 73 |
+
"video_processor_type": "Gemma4VideoProcessor"
|
| 74 |
+
}
|
| 75 |
+
}
|
report.json
CHANGED
|
@@ -1,146 +1,160 @@
|
|
| 1 |
-
{
|
| 2 |
-
"model": "google/gemma-4-
|
| 3 |
-
"num_layers": 42,
|
| 4 |
-
"hidden_size": 2560,
|
| 5 |
-
"direction_layer": 25,
|
| 6 |
-
"refusal_subspace_rank": 1,
|
| 7 |
-
"
|
| 8 |
-
"
|
| 9 |
-
"
|
| 10 |
-
"
|
| 11 |
-
"
|
| 12 |
-
"
|
| 13 |
-
"
|
| 14 |
-
"
|
| 15 |
-
"
|
| 16 |
-
"
|
| 17 |
-
"
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
0.0,
|
| 27 |
-
0.
|
| 28 |
-
0.
|
| 29 |
-
0.
|
| 30 |
-
0.
|
| 31 |
-
0.
|
| 32 |
-
0.
|
| 33 |
-
0.
|
| 34 |
-
0.0,
|
| 35 |
-
0.0,
|
| 36 |
-
0.0,
|
| 37 |
-
0.
|
| 38 |
-
0.
|
| 39 |
-
0.0,
|
| 40 |
-
0.0,
|
| 41 |
-
0.
|
| 42 |
-
0.
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
0.
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
0.
|
| 56 |
-
|
| 57 |
-
0.
|
| 58 |
-
0.
|
| 59 |
-
0.
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
0.0,
|
| 63 |
-
0.0,
|
| 64 |
-
0.
|
| 65 |
-
0.0,
|
| 66 |
-
0.0
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
0.0,
|
| 70 |
-
0.0,
|
| 71 |
-
0.0,
|
| 72 |
-
0.0,
|
| 73 |
-
0.0,
|
| 74 |
-
0.0,
|
| 75 |
-
0.0,
|
| 76 |
-
0.0,
|
| 77 |
-
0.0,
|
| 78 |
-
0.0,
|
| 79 |
-
0.0,
|
| 80 |
-
0.0,
|
| 81 |
-
0.0,
|
| 82 |
-
0.0,
|
| 83 |
-
0.0,
|
| 84 |
-
0.0,
|
| 85 |
-
0.0,
|
| 86 |
-
0.0,
|
| 87 |
-
0.0,
|
| 88 |
-
0.0,
|
| 89 |
-
0.0,
|
| 90 |
-
0.0,
|
| 91 |
-
0.0,
|
| 92 |
-
0.0,
|
| 93 |
-
0.0,
|
| 94 |
-
0.0,
|
| 95 |
-
0.0,
|
| 96 |
-
0.0,
|
| 97 |
-
0.0,
|
| 98 |
-
0.0,
|
| 99 |
-
0.0,
|
| 100 |
-
0.0,
|
| 101 |
-
0.0,
|
| 102 |
-
0.0,
|
| 103 |
-
0.0
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
"
|
| 113 |
-
"
|
| 114 |
-
"
|
| 115 |
-
"
|
| 116 |
-
"
|
| 117 |
-
"
|
| 118 |
-
"
|
| 119 |
-
"
|
| 120 |
-
"
|
| 121 |
-
"
|
| 122 |
-
"
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
"
|
| 142 |
-
"
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 146 |
}
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": "google/gemma-4-e4b-it",
|
| 3 |
+
"num_layers": 42,
|
| 4 |
+
"hidden_size": 2560,
|
| 5 |
+
"direction_layer": 25,
|
| 6 |
+
"refusal_subspace_rank": 1,
|
| 7 |
+
"max_refusal_rank": 3,
|
| 8 |
+
"multi_refusal": true,
|
| 9 |
+
"multi_refusal_clusters": 6,
|
| 10 |
+
"multi_refusal_min_norm": 0.08,
|
| 11 |
+
"multi_refusal_min_separation": 0.05,
|
| 12 |
+
"multi_refusal_min_coverage": 0.05,
|
| 13 |
+
"initial_separation": 68.4185,
|
| 14 |
+
"baseline_refusal_rate": 0.9583,
|
| 15 |
+
"baseline_eval_n": 24,
|
| 16 |
+
"edited_refusal_rate": 0.019,
|
| 17 |
+
"refusal_metric": "classifier + weak guard",
|
| 18 |
+
"harmless_kl_nats": 0.145,
|
| 19 |
+
"kl_backoff_steps": 0,
|
| 20 |
+
"kl_layer_trim_steps": 0,
|
| 21 |
+
"repair_steps": 4,
|
| 22 |
+
"residual_repair": [],
|
| 23 |
+
"guard_history": [],
|
| 24 |
+
"layer_alphas": [
|
| 25 |
+
1.695,
|
| 26 |
+
0.0,
|
| 27 |
+
0.029,
|
| 28 |
+
0.003,
|
| 29 |
+
0.055,
|
| 30 |
+
0.046,
|
| 31 |
+
0.0,
|
| 32 |
+
0.0,
|
| 33 |
+
0.049,
|
| 34 |
+
0.0,
|
| 35 |
+
0.0,
|
| 36 |
+
0.0,
|
| 37 |
+
0.4,
|
| 38 |
+
0.159,
|
| 39 |
+
0.0,
|
| 40 |
+
0.0,
|
| 41 |
+
0.165,
|
| 42 |
+
0.0,
|
| 43 |
+
0.025,
|
| 44 |
+
0.0,
|
| 45 |
+
0.279,
|
| 46 |
+
0.0,
|
| 47 |
+
0.0,
|
| 48 |
+
0.0,
|
| 49 |
+
0.474,
|
| 50 |
+
3.871,
|
| 51 |
+
3.2,
|
| 52 |
+
2.455,
|
| 53 |
+
6.715,
|
| 54 |
+
2.297,
|
| 55 |
+
0.746,
|
| 56 |
+
4.033,
|
| 57 |
+
0.328,
|
| 58 |
+
0.64,
|
| 59 |
+
0.713,
|
| 60 |
+
0.591,
|
| 61 |
+
0.0,
|
| 62 |
+
0.0,
|
| 63 |
+
0.0,
|
| 64 |
+
0.003,
|
| 65 |
+
0.0,
|
| 66 |
+
0.0
|
| 67 |
+
],
|
| 68 |
+
"ple_layer_alphas": [
|
| 69 |
+
0.0,
|
| 70 |
+
0.0,
|
| 71 |
+
0.0,
|
| 72 |
+
0.0,
|
| 73 |
+
0.0,
|
| 74 |
+
0.0,
|
| 75 |
+
0.0,
|
| 76 |
+
0.0,
|
| 77 |
+
0.0,
|
| 78 |
+
0.0,
|
| 79 |
+
0.0,
|
| 80 |
+
0.0,
|
| 81 |
+
0.0,
|
| 82 |
+
0.0,
|
| 83 |
+
0.0,
|
| 84 |
+
0.0,
|
| 85 |
+
0.0,
|
| 86 |
+
0.0,
|
| 87 |
+
0.0,
|
| 88 |
+
0.0,
|
| 89 |
+
0.0,
|
| 90 |
+
0.0,
|
| 91 |
+
0.0,
|
| 92 |
+
0.0,
|
| 93 |
+
0.0,
|
| 94 |
+
0.0,
|
| 95 |
+
0.0,
|
| 96 |
+
0.0,
|
| 97 |
+
0.0,
|
| 98 |
+
0.0,
|
| 99 |
+
0.0,
|
| 100 |
+
0.0,
|
| 101 |
+
0.0,
|
| 102 |
+
0.0,
|
| 103 |
+
0.0,
|
| 104 |
+
0.0,
|
| 105 |
+
0.0,
|
| 106 |
+
0.0,
|
| 107 |
+
0.0,
|
| 108 |
+
0.0,
|
| 109 |
+
0.0,
|
| 110 |
+
0.0
|
| 111 |
+
],
|
| 112 |
+
"ple_embed_alpha": 0.0,
|
| 113 |
+
"ple_model_projection_alpha": 0.0,
|
| 114 |
+
"embed_alpha": 0.0,
|
| 115 |
+
"head_alpha": 0.0,
|
| 116 |
+
"head_token_alpha": 0.0,
|
| 117 |
+
"preserve_rank": 8,
|
| 118 |
+
"preserve_source": "harmless",
|
| 119 |
+
"pruned_layers": [],
|
| 120 |
+
"layers_after_prune": 42,
|
| 121 |
+
"elapsed_sec": 1027.6,
|
| 122 |
+
"profile": "balanced",
|
| 123 |
+
"target_refusal": 0.05,
|
| 124 |
+
"max_kl": 0.15,
|
| 125 |
+
"kl_target": 0.3,
|
| 126 |
+
"refusal_eval_path": "JailbreakBench/JBB-Behaviors@behaviors:harmful:Goal|/var/home/Heterodoxin/apostate/data/refusal_calibration.txt",
|
| 127 |
+
"refusal_eval_n": 64,
|
| 128 |
+
"kl_positions": 8,
|
| 129 |
+
"kl_eval_path": "mlabonne/harmless_alpaca:test:text",
|
| 130 |
+
"kl_eval_n": 48,
|
| 131 |
+
"oblique_ablation": false,
|
| 132 |
+
"oblique_strength": 1.0,
|
| 133 |
+
"opt_capability": true,
|
| 134 |
+
"opt_capability_weight": 2.5,
|
| 135 |
+
"timings_sec": {
|
| 136 |
+
"load_model": 10.6,
|
| 137 |
+
"load_prompts": 5.2,
|
| 138 |
+
"baseline_refusal": 42.9,
|
| 139 |
+
"activation_fit": 211.6,
|
| 140 |
+
"reader_profile": 57.1,
|
| 141 |
+
"reader_guard": 20.1,
|
| 142 |
+
"guard": 0.0,
|
| 143 |
+
"refine_refusal": 0.0,
|
| 144 |
+
"validation_metrics": 91.3,
|
| 145 |
+
"repair": 235.6,
|
| 146 |
+
"prune": 0.0,
|
| 147 |
+
"test_metrics": 336.7,
|
| 148 |
+
"bake": 16.6
|
| 149 |
+
},
|
| 150 |
+
"command": "/var/home/Heterodoxin/apostate/apostate/cli.py --optimize --n-trials 8 --model google/gemma-4-e4b-it --output-dir /var/home/Heterodoxin/gemma4_e4b_lowkl_out --oblique-predictive --target-refusal 0.05 --max-kl 0.15 --reader-max-kl 0.15",
|
| 151 |
+
"optimized": true,
|
| 152 |
+
"reader_mode": true,
|
| 153 |
+
"reader_strength": 2.0,
|
| 154 |
+
"best_trial": {
|
| 155 |
+
"refusal": 0.375,
|
| 156 |
+
"kl": 1.676835060119629
|
| 157 |
+
},
|
| 158 |
+
"n_trials": 0,
|
| 159 |
+
"baked_to": "/var/home/Heterodoxin/gemma4_e4b_lowkl_out"
|
| 160 |
}
|
report.md
CHANGED
|
@@ -1,106 +1,112 @@
|
|
| 1 |
-
# Apostate Run Report
|
| 2 |
-
|
| 3 |
-
## Summary
|
| 4 |
-
| Metric | Value |
|
| 5 |
-
| --- | --- |
|
| 6 |
-
| Base model | google/gemma-4-
|
| 7 |
-
| Profile | balanced |
|
| 8 |
-
| Output |
|
| 9 |
-
| Layers | 42 |
|
| 10 |
-
| Hidden size | 2560 |
|
| 11 |
-
| Direction layer | 25 |
|
| 12 |
-
|
|
| 13 |
-
|
|
| 14 |
-
|
|
| 15 |
-
|
|
| 16 |
-
|
|
| 17 |
-
|
|
| 18 |
-
|
|
| 19 |
-
|
|
| 20 |
-
| KL
|
| 21 |
-
|
|
| 22 |
-
|
|
| 23 |
-
|
|
| 24 |
-
|
|
| 25 |
-
|
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
|
| 41 |
-
|
|
| 42 |
-
|
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
|
| 46 |
-
|
|
| 47 |
-
|
|
| 48 |
-
|
|
| 49 |
-
|
|
| 50 |
-
|
|
| 51 |
-
|
|
| 52 |
-
|
|
| 53 |
-
|
|
| 54 |
-
|
|
| 55 |
-
|
|
| 56 |
-
|
|
| 57 |
-
|
|
| 58 |
-
|
|
| 59 |
-
|
|
| 60 |
-
|
|
| 61 |
-
|
|
| 62 |
-
|
|
| 63 |
-
|
|
| 64 |
-
|
|
| 65 |
-
|
|
| 66 |
-
|
|
| 67 |
-
|
|
| 68 |
-
|
|
| 69 |
-
|
|
| 70 |
-
|
|
| 71 |
-
|
|
| 72 |
-
|
|
| 73 |
-
|
|
| 74 |
-
|
|
| 75 |
-
|
|
| 76 |
-
|
|
| 77 |
-
|
|
| 78 |
-
|
|
| 79 |
-
|
|
| 80 |
-
|
|
| 81 |
-
|
|
| 82 |
-
|
|
| 83 |
-
|
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
|
| 87 |
-
|
|
| 88 |
-
|
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
|
| 92 |
-
|
|
| 93 |
-
|
|
| 94 |
-
|
|
| 95 |
-
|
|
| 96 |
-
|
|
| 97 |
-
|
|
| 98 |
-
|
|
| 99 |
-
|
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
|
| 103 |
-
|
|
| 104 |
-
|
|
| 105 |
-
|
|
| 106 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Apostate Run Report
|
| 2 |
+
|
| 3 |
+
## Summary
|
| 4 |
+
| Metric | Value |
|
| 5 |
+
| --- | --- |
|
| 6 |
+
| Base model | google/gemma-4-e4b-it |
|
| 7 |
+
| Profile | balanced |
|
| 8 |
+
| Output | /var/home/Heterodoxin/gemma4_e4b_lowkl_out |
|
| 9 |
+
| Layers | 42 |
|
| 10 |
+
| Hidden size | 2560 |
|
| 11 |
+
| Direction layer | 25 |
|
| 12 |
+
| Refusal rank | 1 |
|
| 13 |
+
| Max refusal rank | 3 |
|
| 14 |
+
| Multi refusal | True |
|
| 15 |
+
| Multi clusters | 6 |
|
| 16 |
+
| Multi min coverage | 0.05 |
|
| 17 |
+
| Baseline refusal (n=24) | 95.8% |
|
| 18 |
+
| Edited refusal | 1.9% |
|
| 19 |
+
| Refusal metric | classifier + weak guard |
|
| 20 |
+
| Harmless KL | 0.145 |
|
| 21 |
+
| Target refusal | 5.0% |
|
| 22 |
+
| KL target | 0.300 |
|
| 23 |
+
| KL budget | 0.150 |
|
| 24 |
+
| KL positions | 8 |
|
| 25 |
+
| KL layer trims | 0 |
|
| 26 |
+
| Repair steps | 4 |
|
| 27 |
+
| Preserve rank | 8 |
|
| 28 |
+
| Preserve source | harmless |
|
| 29 |
+
| Capability penalty | True |
|
| 30 |
+
| Elapsed | 1027.6 sec |
|
| 31 |
+
|
| 32 |
+
## Command
|
| 33 |
+
|
| 34 |
+
```text
|
| 35 |
+
/var/home/Heterodoxin/apostate/apostate/cli.py --optimize --n-trials 8 --model google/gemma-4-e4b-it --output-dir /var/home/Heterodoxin/gemma4_e4b_lowkl_out --oblique-predictive --target-refusal 0.05 --max-kl 0.15 --reader-max-kl 0.15
|
| 36 |
+
```
|
| 37 |
+
|
| 38 |
+
## Best Trial
|
| 39 |
+
| Metric | Value |
|
| 40 |
+
| --- | --- |
|
| 41 |
+
| refusal | 0.375 |
|
| 42 |
+
| kl | 1.676835060119629 |
|
| 43 |
+
|
| 44 |
+
## Layer Alphas
|
| 45 |
+
| Layer | Alpha |
|
| 46 |
+
| --- | --- |
|
| 47 |
+
| 0 | 1.695 |
|
| 48 |
+
| 1 | 0.000 |
|
| 49 |
+
| 2 | 0.029 |
|
| 50 |
+
| 3 | 0.003 |
|
| 51 |
+
| 4 | 0.055 |
|
| 52 |
+
| 5 | 0.046 |
|
| 53 |
+
| 6 | 0.000 |
|
| 54 |
+
| 7 | 0.000 |
|
| 55 |
+
| 8 | 0.049 |
|
| 56 |
+
| 9 | 0.000 |
|
| 57 |
+
| 10 | 0.000 |
|
| 58 |
+
| 11 | 0.000 |
|
| 59 |
+
| 12 | 0.400 |
|
| 60 |
+
| 13 | 0.159 |
|
| 61 |
+
| 14 | 0.000 |
|
| 62 |
+
| 15 | 0.000 |
|
| 63 |
+
| 16 | 0.165 |
|
| 64 |
+
| 17 | 0.000 |
|
| 65 |
+
| 18 | 0.025 |
|
| 66 |
+
| 19 | 0.000 |
|
| 67 |
+
| 20 | 0.279 |
|
| 68 |
+
| 21 | 0.000 |
|
| 69 |
+
| 22 | 0.000 |
|
| 70 |
+
| 23 | 0.000 |
|
| 71 |
+
| 24 | 0.474 |
|
| 72 |
+
| 25 | 3.871 |
|
| 73 |
+
| 26 | 3.200 |
|
| 74 |
+
| 27 | 2.455 |
|
| 75 |
+
| 28 | 6.715 |
|
| 76 |
+
| 29 | 2.297 |
|
| 77 |
+
| 30 | 0.746 |
|
| 78 |
+
| 31 | 4.033 |
|
| 79 |
+
| 32 | 0.328 |
|
| 80 |
+
| 33 | 0.640 |
|
| 81 |
+
| 34 | 0.713 |
|
| 82 |
+
| 35 | 0.591 |
|
| 83 |
+
| 36 | 0.000 |
|
| 84 |
+
| 37 | 0.000 |
|
| 85 |
+
| 38 | 0.000 |
|
| 86 |
+
| 39 | 0.003 |
|
| 87 |
+
| 40 | 0.000 |
|
| 88 |
+
| 41 | 0.000 |
|
| 89 |
+
|
| 90 |
+
## Timings
|
| 91 |
+
| Phase | Seconds |
|
| 92 |
+
| --- | --- |
|
| 93 |
+
| load_model | 10.6 |
|
| 94 |
+
| load_prompts | 5.2 |
|
| 95 |
+
| baseline_refusal | 42.9 |
|
| 96 |
+
| activation_fit | 211.6 |
|
| 97 |
+
| reader_profile | 57.1 |
|
| 98 |
+
| reader_guard | 20.1 |
|
| 99 |
+
| guard | 0.0 |
|
| 100 |
+
| refine_refusal | 0.0 |
|
| 101 |
+
| validation_metrics | 91.3 |
|
| 102 |
+
| repair | 235.6 |
|
| 103 |
+
| prune | 0.0 |
|
| 104 |
+
| test_metrics | 336.7 |
|
| 105 |
+
| bake | 16.6 |
|
| 106 |
+
|
| 107 |
+
## Measurement
|
| 108 |
+
| field | value |
|
| 109 |
+
| --- | --- |
|
| 110 |
+
| refusal judge | classifier + weak guard |
|
| 111 |
+
| preservation metric | harmless kl |
|
| 112 |
+
| capability suites | gsm8k, humaneval, mbpp |
|
tokenizer_config.json
CHANGED
|
@@ -1,96 +1,96 @@
|
|
| 1 |
-
{
|
| 2 |
-
"audio_token": "<|audio|>",
|
| 3 |
-
"backend": "tokenizers",
|
| 4 |
-
"boa_token": "<|audio>",
|
| 5 |
-
"boi_token": "<|image>",
|
| 6 |
-
"bos_token": "<bos>",
|
| 7 |
-
"eoa_token": "<audio|>",
|
| 8 |
-
"eoc_token": "<channel|>",
|
| 9 |
-
"eoi_token": "<image|>",
|
| 10 |
-
"eos_token": "<eos>",
|
| 11 |
-
"eot_token": "<turn|>",
|
| 12 |
-
"escape_token": "<|\"|>",
|
| 13 |
-
"etc_token": "<tool_call|>",
|
| 14 |
-
"etd_token": "<tool|>",
|
| 15 |
-
"etr_token": "<tool_response|>",
|
| 16 |
-
"extra_special_tokens": [
|
| 17 |
-
"<|video|>"
|
| 18 |
-
],
|
| 19 |
-
"image_token": "<|image|>",
|
| 20 |
-
"is_local": false,
|
| 21 |
-
"local_files_only": false,
|
| 22 |
-
"mask_token": "<mask>",
|
| 23 |
-
"model_max_length": 1000000000000000019884624838656,
|
| 24 |
-
"model_specific_special_tokens": {
|
| 25 |
-
"audio_token": "<|audio|>",
|
| 26 |
-
"boa_token": "<|audio>",
|
| 27 |
-
"boi_token": "<|image>",
|
| 28 |
-
"eoa_token": "<audio|>",
|
| 29 |
-
"eoc_token": "<channel|>",
|
| 30 |
-
"eoi_token": "<image|>",
|
| 31 |
-
"eot_token": "<turn|>",
|
| 32 |
-
"escape_token": "<|\"|>",
|
| 33 |
-
"etc_token": "<tool_call|>",
|
| 34 |
-
"etd_token": "<tool|>",
|
| 35 |
-
"etr_token": "<tool_response|>",
|
| 36 |
-
"image_token": "<|image|>",
|
| 37 |
-
"soc_token": "<|channel>",
|
| 38 |
-
"sot_token": "<|turn>",
|
| 39 |
-
"stc_token": "<|tool_call>",
|
| 40 |
-
"std_token": "<|tool>",
|
| 41 |
-
"str_token": "<|tool_response>",
|
| 42 |
-
"think_token": "<|think|>"
|
| 43 |
-
},
|
| 44 |
-
"pad_token": "<pad>",
|
| 45 |
-
"padding_side": "left",
|
| 46 |
-
"processor_class": "Gemma4Processor",
|
| 47 |
-
"response_schema": {
|
| 48 |
-
"properties": {
|
| 49 |
-
"content": {
|
| 50 |
-
"type": "string"
|
| 51 |
-
},
|
| 52 |
-
"role": {
|
| 53 |
-
"const": "assistant"
|
| 54 |
-
},
|
| 55 |
-
"thinking": {
|
| 56 |
-
"type": "string"
|
| 57 |
-
},
|
| 58 |
-
"tool_calls": {
|
| 59 |
-
"items": {
|
| 60 |
-
"properties": {
|
| 61 |
-
"function": {
|
| 62 |
-
"properties": {
|
| 63 |
-
"arguments": {
|
| 64 |
-
"additionalProperties": {},
|
| 65 |
-
"type": "object",
|
| 66 |
-
"x-parser": "gemma4-tool-call"
|
| 67 |
-
},
|
| 68 |
-
"name": {
|
| 69 |
-
"type": "string"
|
| 70 |
-
}
|
| 71 |
-
},
|
| 72 |
-
"type": "object",
|
| 73 |
-
"x-regex": "call\\:(?P<name>\\w+)(?P<arguments>\\{.*\\})"
|
| 74 |
-
},
|
| 75 |
-
"type": {
|
| 76 |
-
"const": "function"
|
| 77 |
-
}
|
| 78 |
-
},
|
| 79 |
-
"type": "object"
|
| 80 |
-
},
|
| 81 |
-
"type": "array",
|
| 82 |
-
"x-regex-iterator": "<\\|tool_call>(.*?)<tool_call\\|>"
|
| 83 |
-
}
|
| 84 |
-
},
|
| 85 |
-
"type": "object",
|
| 86 |
-
"x-regex": "(\\<\\|channel\\>thought\\n(?P<thinking>.*?)\\<channel\\|\\>)?(?P<tool_calls>\\<\\|tool_call\\>.*\\<tool_call\\|\\>)?(?P<content>(?:(?!\\<turn\\|\\>)(?!\\<\\|tool_response\\>).)+)?(?:\\<turn\\|\\>|\\<\\|tool_response\\>)?"
|
| 87 |
-
},
|
| 88 |
-
"soc_token": "<|channel>",
|
| 89 |
-
"sot_token": "<|turn>",
|
| 90 |
-
"stc_token": "<|tool_call>",
|
| 91 |
-
"std_token": "<|tool>",
|
| 92 |
-
"str_token": "<|tool_response>",
|
| 93 |
-
"think_token": "<|think|>",
|
| 94 |
-
"tokenizer_class": "GemmaTokenizer",
|
| 95 |
-
"unk_token": "<unk>"
|
| 96 |
-
}
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"audio_token": "<|audio|>",
|
| 3 |
+
"backend": "tokenizers",
|
| 4 |
+
"boa_token": "<|audio>",
|
| 5 |
+
"boi_token": "<|image>",
|
| 6 |
+
"bos_token": "<bos>",
|
| 7 |
+
"eoa_token": "<audio|>",
|
| 8 |
+
"eoc_token": "<channel|>",
|
| 9 |
+
"eoi_token": "<image|>",
|
| 10 |
+
"eos_token": "<eos>",
|
| 11 |
+
"eot_token": "<turn|>",
|
| 12 |
+
"escape_token": "<|\"|>",
|
| 13 |
+
"etc_token": "<tool_call|>",
|
| 14 |
+
"etd_token": "<tool|>",
|
| 15 |
+
"etr_token": "<tool_response|>",
|
| 16 |
+
"extra_special_tokens": [
|
| 17 |
+
"<|video|>"
|
| 18 |
+
],
|
| 19 |
+
"image_token": "<|image|>",
|
| 20 |
+
"is_local": false,
|
| 21 |
+
"local_files_only": false,
|
| 22 |
+
"mask_token": "<mask>",
|
| 23 |
+
"model_max_length": 1000000000000000019884624838656,
|
| 24 |
+
"model_specific_special_tokens": {
|
| 25 |
+
"audio_token": "<|audio|>",
|
| 26 |
+
"boa_token": "<|audio>",
|
| 27 |
+
"boi_token": "<|image>",
|
| 28 |
+
"eoa_token": "<audio|>",
|
| 29 |
+
"eoc_token": "<channel|>",
|
| 30 |
+
"eoi_token": "<image|>",
|
| 31 |
+
"eot_token": "<turn|>",
|
| 32 |
+
"escape_token": "<|\"|>",
|
| 33 |
+
"etc_token": "<tool_call|>",
|
| 34 |
+
"etd_token": "<tool|>",
|
| 35 |
+
"etr_token": "<tool_response|>",
|
| 36 |
+
"image_token": "<|image|>",
|
| 37 |
+
"soc_token": "<|channel>",
|
| 38 |
+
"sot_token": "<|turn>",
|
| 39 |
+
"stc_token": "<|tool_call>",
|
| 40 |
+
"std_token": "<|tool>",
|
| 41 |
+
"str_token": "<|tool_response>",
|
| 42 |
+
"think_token": "<|think|>"
|
| 43 |
+
},
|
| 44 |
+
"pad_token": "<pad>",
|
| 45 |
+
"padding_side": "left",
|
| 46 |
+
"processor_class": "Gemma4Processor",
|
| 47 |
+
"response_schema": {
|
| 48 |
+
"properties": {
|
| 49 |
+
"content": {
|
| 50 |
+
"type": "string"
|
| 51 |
+
},
|
| 52 |
+
"role": {
|
| 53 |
+
"const": "assistant"
|
| 54 |
+
},
|
| 55 |
+
"thinking": {
|
| 56 |
+
"type": "string"
|
| 57 |
+
},
|
| 58 |
+
"tool_calls": {
|
| 59 |
+
"items": {
|
| 60 |
+
"properties": {
|
| 61 |
+
"function": {
|
| 62 |
+
"properties": {
|
| 63 |
+
"arguments": {
|
| 64 |
+
"additionalProperties": {},
|
| 65 |
+
"type": "object",
|
| 66 |
+
"x-parser": "gemma4-tool-call"
|
| 67 |
+
},
|
| 68 |
+
"name": {
|
| 69 |
+
"type": "string"
|
| 70 |
+
}
|
| 71 |
+
},
|
| 72 |
+
"type": "object",
|
| 73 |
+
"x-regex": "call\\:(?P<name>\\w+)(?P<arguments>\\{.*\\})"
|
| 74 |
+
},
|
| 75 |
+
"type": {
|
| 76 |
+
"const": "function"
|
| 77 |
+
}
|
| 78 |
+
},
|
| 79 |
+
"type": "object"
|
| 80 |
+
},
|
| 81 |
+
"type": "array",
|
| 82 |
+
"x-regex-iterator": "<\\|tool_call>(.*?)<tool_call\\|>"
|
| 83 |
+
}
|
| 84 |
+
},
|
| 85 |
+
"type": "object",
|
| 86 |
+
"x-regex": "(\\<\\|channel\\>thought\\n(?P<thinking>.*?)\\<channel\\|\\>)?(?P<tool_calls>\\<\\|tool_call\\>.*\\<tool_call\\|\\>)?(?P<content>(?:(?!\\<turn\\|\\>)(?!\\<\\|tool_response\\>).)+)?(?:\\<turn\\|\\>|\\<\\|tool_response\\>)?"
|
| 87 |
+
},
|
| 88 |
+
"soc_token": "<|channel>",
|
| 89 |
+
"sot_token": "<|turn>",
|
| 90 |
+
"stc_token": "<|tool_call>",
|
| 91 |
+
"std_token": "<|tool>",
|
| 92 |
+
"str_token": "<|tool_response>",
|
| 93 |
+
"think_token": "<|think|>",
|
| 94 |
+
"tokenizer_class": "GemmaTokenizer",
|
| 95 |
+
"unk_token": "<unk>"
|
| 96 |
+
}
|