System-One QLoRA: 0.6B + 4B adapters, format-matched NLI recipe

06b/adapter_config.json

@@ -0,0 +1,51 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "kasa_config": null,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.21.0",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "up_proj",
+    "k_proj",
+    "v_proj",
+    "gate_proj",
+    "q_proj",
+    "down_proj",
+    "o_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

06b/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c77772b101ad6597ee368a0ba722e7751726cefdb46198ccc25b881a75e9b295
+size 40422168

4b/adapter_config.json

@@ -0,0 +1,51 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-4B-Instruct-2507",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "kasa_config": null,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.21.0",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "o_proj",
+    "gate_proj",
+    "down_proj",
+    "k_proj",
+    "v_proj",
+    "up_proj",
+    "q_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

4b/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb052fd249de8100f9898e2f4bec5f8300df5cd2f22ee6b45f3a95bebe93715f
+size 132187888

README.md

@@ -0,0 +1,83 @@
+---
+license: apache-2.0
+library_name: peft
+tags:
+  - decision-model
+  - system-one
+  - calibration
+  - lora
+  - qlora
+  - typesafe
+base_model:
+  - Qwen/Qwen3-4B-Instruct-2507
+  - Qwen/Qwen3-0.6B
+metrics:
+  - accuracy
+  - brier_score
+  - expected_calibration_error
+---
+
+# System-One QLoRA (0.6B + 4B)
+
+Open **single-pass decision scorers** in the shape of TypeSafe's *Jev* ("System One"):
+a `state` + a multiple-choice `question` go in, a **calibrated distribution over the
+options** comes out in one forward pass — no autoregressive generation.
+
+Two QLoRA adapters, same recipe, two latency/accuracy points:
+
+| | base | trainable | latency / answer | where it wins vs Jev |
+|---|---|---|---|
+| **`06b/`** | Qwen3-0.6B | LoRA r=16 | **~53 ms** (20× faster than Jev) | abstention 0.960 vs 0.740 |
+| **`4b/`** | Qwen3-4B-Instruct-2507 | LoRA r=16 | **~165 ms** (6.5× faster) | reflex + abstention; best SNI |
+
+## Scoring mechanism
+
+Prompt ends `…\n\nAnswer:`; we read the next-token log-prob of each option **letter**
+(` A`, ` B`, …), softmax over the options, and apply a temperature fitted on val.
+The argmax is the decision; the softmax is the calibrated confidence. (Letter readout
+caps at 26 options — costs 1 of ~3,350 benchmark rows here.)
+
+```python
+import torch
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+base = "Qwen/Qwen3-4B-Instruct-2507"          # or "Qwen/Qwen3-0.6B" for 06b/
+tok = AutoTokenizer.from_pretrained(base)
+model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.bfloat16, device_map="auto")
+model = PeftModel.from_pretrained(model, "yocoms/system1-qlora", subfolder="4b")  # or "06b"
+model.eval()
+
+prompt = "State:\n{state}\n\nQuestion:\n{q}\n\nOptions:\nA. ...\nB. ...\n\nAnswer:"
+ids = tok(prompt, return_tensors="pt").to(model.device)
+logits = model(**ids).logits[0, -1]
+# compare log-probs of ' A', ' B', ... ; softmax(/T) = calibrated distribution
+```
+
+## Benchmarks (held-out test, calibrated)
+
+Accuracy / ECE, vs frozen **Jev 1.13** (`opencode-zen/jev-1.13`) on identical items.
+
+| benchmark | 0.6B | 4B | Jev |
+|---|---|---|---|
+| SNI (semantic / NLI) | 0.613 / .037 | **0.707** / .050 | 0.838 |
+| reflex (safety) | 0.553 / .059 | 0.558 / .072 | 0.543 |
+| BFCL (tool select) | 0.885 / .032 | 0.920 / .034 | 0.957 |
+| **abstention** (none-fits) | **0.960** / .019 | 0.924 / .035 | 0.740 |
+
+- **Beats Jev** on reflex and abstention; near-parity on tool selection; well-calibrated (ECE ≤ 0.07) across the board.
+- **SNI gap to Jev is knowledge/reasoning**, not format — see recipe.
+
+## Recipe (why it works)
+
+- Corpus: SNI + reflex train splits, **absence-augmentation** (abstain/NOTA variants,
+  upweighted absence class → the abstention win), and **format-matched NLI**.
+- **The NLI fix**: SNI's held-out NLI tasks are *"pick which of 3 candidates is neutral"*,
+  not single-pair label classification. Training standard MNLI (`premise+hypothesis→label`)
+  left `mnli_neutral` at 0.16 (below chance). Rebuilding the data in the *select-of-3* format
+  lifted `mnli_neutral` **0.16 → 0.88** (beating Jev's 0.84) on the 0.6B — it was a format
+  mismatch, not model capacity.
+- QLoRA 4-bit NF4, r=16, batch 4 × seq 384, ~1.25 epoch. Code + corpus builders:
+  https://github.com/y0c0ms/System1QLoRa
+
+Not affiliated with TypeSafe AI. Independent reproduction of the *shape* of Jev.