Instructions to use automated-alignment-science/cot-controllability-gpt-oss-20b-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use automated-alignment-science/cot-controllability-gpt-oss-20b-lora with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
Add compliant (cdel) + matched control (ctrldel) rank-32 LoRA adapters
Browse files- README.md +41 -0
- cdel/adapter_config.json +31 -0
- cdel/adapter_model.safetensors +3 -0
- cdel/training_meta.json +219 -0
- ctrldel/adapter_config.json +31 -0
- ctrldel/adapter_model.safetensors +3 -0
- ctrldel/training_meta.json +219 -0
README.md
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model: openai/gpt-oss-20b
|
| 4 |
+
library_name: peft
|
| 5 |
+
tags:
|
| 6 |
+
- lora
|
| 7 |
+
- chain-of-thought
|
| 8 |
+
- ai-safety
|
| 9 |
+
- interpretability
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# CoT-controllability LoRA fine-tunes (gpt-oss-20b)
|
| 13 |
+
|
| 14 |
+
Rank-32 LoRA adapters for the project **"A 2,880-number steering vector gives a reasoning model the
|
| 15 |
+
chain-of-thought control that fine-tuning does."** The **fine-tune is the benchmark the frozen-weights
|
| 16 |
+
steering vector matches** (paired held-out difference +0.4pp [−1.9,+2.8]).
|
| 17 |
+
|
| 18 |
+
- **`cdel/`** — the **compliant** fine-tune: rank-32 LoRA trained on edited *complying* reasoning
|
| 19 |
+
traces (completion-only loss; LR 2e-4, 3 epochs). Held-out strict CoT-control compliance rises
|
| 20 |
+
+12.3pp [+10.4,+14.2] over base; bullet formatting 0→52%.
|
| 21 |
+
- **`ctrldel/`** — the **matched raw-trace control**: same prompts and config but trained on the
|
| 22 |
+
*non-complying* traces. Held-out uplift ≈ 0 — the dissociation that shows the compliant uplift is
|
| 23 |
+
genuine instruction-following, not generic fine-tuning.
|
| 24 |
+
|
| 25 |
+
Each folder is a standard PEFT adapter (`adapter_config.json` + `adapter_model.safetensors`) plus
|
| 26 |
+
`training_meta.json` (the full training provenance: data hash, loss curve, git hash). `r=32`,
|
| 27 |
+
`lora_alpha=32`, target modules = MoE experts + attention + unembed.
|
| 28 |
+
|
| 29 |
+
## Load
|
| 30 |
+
|
| 31 |
+
```python
|
| 32 |
+
from peft import PeftModel
|
| 33 |
+
from transformers import AutoModelForCausalLM
|
| 34 |
+
base = AutoModelForCausalLM.from_pretrained("openai/gpt-oss-20b", torch_dtype="bfloat16")
|
| 35 |
+
model = PeftModel.from_pretrained(base, "automated-alignment-science/cot-controllability-gpt-oss-20b-lora", subfolder="cdel")
|
| 36 |
+
```
|
| 37 |
+
|
| 38 |
+
The figures in the companion code do **not** need this model (they plot saved eval metrics). The
|
| 39 |
+
adapter is here for full reproducibility from scratch. Datasets + steering vectors:
|
| 40 |
+
`automated-alignment-science/cot-controllability-steering-vectors`. Code:
|
| 41 |
+
`github.com/redwoodresearch/automated-research-projects` (folder `cot-controllability-steering-vectors`).
|
cdel/adapter_config.json
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"alpha_pattern": {},
|
| 3 |
+
"auto_mapping": null,
|
| 4 |
+
"base_model_name_or_path": "openai/gpt-oss-20b",
|
| 5 |
+
"bias": "none",
|
| 6 |
+
"corda_config": null,
|
| 7 |
+
"eva_config": null,
|
| 8 |
+
"exclude_modules": null,
|
| 9 |
+
"fan_in_fan_out": false,
|
| 10 |
+
"inference_mode": false,
|
| 11 |
+
"init_lora_weights": true,
|
| 12 |
+
"layer_replication": null,
|
| 13 |
+
"layers_pattern": null,
|
| 14 |
+
"layers_to_transform": null,
|
| 15 |
+
"loftq_config": {},
|
| 16 |
+
"lora_alpha": 32,
|
| 17 |
+
"lora_bias": false,
|
| 18 |
+
"lora_dropout": 0,
|
| 19 |
+
"megatron_config": null,
|
| 20 |
+
"megatron_core": "megatron.core",
|
| 21 |
+
"modules_to_save": null,
|
| 22 |
+
"peft_type": "LORA",
|
| 23 |
+
"r": 32,
|
| 24 |
+
"rank_pattern": {},
|
| 25 |
+
"revision": null,
|
| 26 |
+
"target_modules": "all-linear",
|
| 27 |
+
"task_type": "CAUSAL_LM",
|
| 28 |
+
"trainable_token_indices": null,
|
| 29 |
+
"use_dora": false,
|
| 30 |
+
"use_rslora": false
|
| 31 |
+
}
|
cdel/adapter_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8e7b950fcd4c3166104516cac3a8e1dc155b889cb1d4793e14672ce799a535be
|
| 3 |
+
size 965742648
|
cdel/training_meta.json
ADDED
|
@@ -0,0 +1,219 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"which": "compliant",
|
| 3 |
+
"tag": "cdel",
|
| 4 |
+
"base_model": "openai/gpt-oss-20b",
|
| 5 |
+
"rank": 32,
|
| 6 |
+
"alpha": 32,
|
| 7 |
+
"plain_frac": 0.0,
|
| 8 |
+
"plain_n": 288,
|
| 9 |
+
"target_modules": "mlp(MoE)+attn+unembed",
|
| 10 |
+
"no_unembed": false,
|
| 11 |
+
"match_control": true,
|
| 12 |
+
"lr": 0.0002,
|
| 13 |
+
"epochs": 3,
|
| 14 |
+
"batch_size": 128,
|
| 15 |
+
"max_length": 2048,
|
| 16 |
+
"warmup_frac": 0.05,
|
| 17 |
+
"seed": 0,
|
| 18 |
+
"lora_targets": "mlp(MoE)+attn+unembed",
|
| 19 |
+
"reduction": "mean",
|
| 20 |
+
"loss_mask": "completion_only",
|
| 21 |
+
"n_examples": 3643,
|
| 22 |
+
"n_skipped_too_long": 13,
|
| 23 |
+
"data_hash": "ef6aa4c756e8ed8e",
|
| 24 |
+
"total_steps": 87,
|
| 25 |
+
"steps_per_epoch": 29,
|
| 26 |
+
"losses": [
|
| 27 |
+
0.7223156299442053,
|
| 28 |
+
0.6608031950891018,
|
| 29 |
+
0.5589050091803074,
|
| 30 |
+
0.5451786462217569,
|
| 31 |
+
0.4993919152766466,
|
| 32 |
+
0.5056490153074265,
|
| 33 |
+
0.47334916703402996,
|
| 34 |
+
0.44312448240816593,
|
| 35 |
+
0.4900579899549484,
|
| 36 |
+
0.4286392144858837,
|
| 37 |
+
0.4315036628395319,
|
| 38 |
+
0.40065754018723965,
|
| 39 |
+
0.4126501213759184,
|
| 40 |
+
0.3955239877104759,
|
| 41 |
+
0.3706433707848191,
|
| 42 |
+
0.30649264995008707,
|
| 43 |
+
0.37842956371605396,
|
| 44 |
+
0.38702673371881247,
|
| 45 |
+
0.3685393212363124,
|
| 46 |
+
0.34478739090263844,
|
| 47 |
+
0.3187951575964689,
|
| 48 |
+
0.3058790545910597,
|
| 49 |
+
0.2945460034534335,
|
| 50 |
+
0.34651653468608856,
|
| 51 |
+
0.2771763186901808,
|
| 52 |
+
0.3221832783892751,
|
| 53 |
+
0.31515009608119726,
|
| 54 |
+
0.32763590570539236,
|
| 55 |
+
0.2960155899241819,
|
| 56 |
+
0.297921571880579,
|
| 57 |
+
0.265383193269372,
|
| 58 |
+
0.24117897264659405,
|
| 59 |
+
0.2517894860357046,
|
| 60 |
+
0.27660175785422325,
|
| 61 |
+
0.237254923209548,
|
| 62 |
+
0.2609592340886593,
|
| 63 |
+
0.23091622814536095,
|
| 64 |
+
0.24239265453070402,
|
| 65 |
+
0.20633055455982685,
|
| 66 |
+
0.27133095264434814,
|
| 67 |
+
0.26146229961887,
|
| 68 |
+
0.2119053048081696,
|
| 69 |
+
0.20747694978490472,
|
| 70 |
+
0.21129672694951296,
|
| 71 |
+
0.206218590028584,
|
| 72 |
+
0.20222537545487285,
|
| 73 |
+
0.25221023336052895,
|
| 74 |
+
0.22629959601908922,
|
| 75 |
+
0.2152238180860877,
|
| 76 |
+
0.21505920449271798,
|
| 77 |
+
0.22332509513944387,
|
| 78 |
+
0.22740747593343258,
|
| 79 |
+
0.2062874431721866,
|
| 80 |
+
0.23623541370034218,
|
| 81 |
+
0.2375688012689352,
|
| 82 |
+
0.20614684373140335,
|
| 83 |
+
0.2074051988311112,
|
| 84 |
+
0.15416770213741368,
|
| 85 |
+
0.17898182524368167,
|
| 86 |
+
0.17934810975566506,
|
| 87 |
+
0.1700979177840054,
|
| 88 |
+
0.16050761146470904,
|
| 89 |
+
0.1832466940395534,
|
| 90 |
+
0.15383817115798593,
|
| 91 |
+
0.17582518979907036,
|
| 92 |
+
0.16522408230230212,
|
| 93 |
+
0.19087925576604903,
|
| 94 |
+
0.16876211389899254,
|
| 95 |
+
0.153222894994542,
|
| 96 |
+
0.16233700094744563,
|
| 97 |
+
0.16682675573974848,
|
| 98 |
+
0.14784478954970837,
|
| 99 |
+
0.17367978626862168,
|
| 100 |
+
0.16841823235154152,
|
| 101 |
+
0.16752003552392125,
|
| 102 |
+
0.17867110366933048,
|
| 103 |
+
0.16490279766730964,
|
| 104 |
+
0.1577738644555211,
|
| 105 |
+
0.15605468838475645,
|
| 106 |
+
0.15437203412875533,
|
| 107 |
+
0.15816477802582085,
|
| 108 |
+
0.14767548814415932,
|
| 109 |
+
0.11899999948218465,
|
| 110 |
+
0.15230661630630493,
|
| 111 |
+
0.14222355047240853,
|
| 112 |
+
0.1687462325207889,
|
| 113 |
+
0.15548258413702754
|
| 114 |
+
],
|
| 115 |
+
"lrs": [
|
| 116 |
+
5e-05,
|
| 117 |
+
0.0001,
|
| 118 |
+
0.00015000000000000001,
|
| 119 |
+
0.0002,
|
| 120 |
+
0.0002,
|
| 121 |
+
0.00019759036144578314,
|
| 122 |
+
0.00019518072289156628,
|
| 123 |
+
0.00019277108433734942,
|
| 124 |
+
0.00019036144578313252,
|
| 125 |
+
0.00018795180722891569,
|
| 126 |
+
0.0001855421686746988,
|
| 127 |
+
0.00018313253012048196,
|
| 128 |
+
0.00018072289156626507,
|
| 129 |
+
0.0001783132530120482,
|
| 130 |
+
0.00017590361445783134,
|
| 131 |
+
0.00017349397590361447,
|
| 132 |
+
0.0001710843373493976,
|
| 133 |
+
0.00016867469879518074,
|
| 134 |
+
0.00016626506024096388,
|
| 135 |
+
0.00016385542168674699,
|
| 136 |
+
0.00016144578313253015,
|
| 137 |
+
0.00015903614457831326,
|
| 138 |
+
0.0001566265060240964,
|
| 139 |
+
0.00015421686746987953,
|
| 140 |
+
0.00015180722891566266,
|
| 141 |
+
0.00014939759036144577,
|
| 142 |
+
0.00014698795180722893,
|
| 143 |
+
0.00014457831325301204,
|
| 144 |
+
0.00014216867469879518,
|
| 145 |
+
0.00013975903614457834,
|
| 146 |
+
0.00013734939759036145,
|
| 147 |
+
0.00013493975903614458,
|
| 148 |
+
0.00013253012048192772,
|
| 149 |
+
0.00013012048192771085,
|
| 150 |
+
0.00012771084337349396,
|
| 151 |
+
0.00012530120481927712,
|
| 152 |
+
0.00012289156626506023,
|
| 153 |
+
0.00012048192771084337,
|
| 154 |
+
0.00011807228915662652,
|
| 155 |
+
0.00011566265060240964,
|
| 156 |
+
0.00011325301204819276,
|
| 157 |
+
0.00011084337349397591,
|
| 158 |
+
0.00010843373493975906,
|
| 159 |
+
0.00010602409638554217,
|
| 160 |
+
0.0001036144578313253,
|
| 161 |
+
0.00010120481927710845,
|
| 162 |
+
9.879518072289157e-05,
|
| 163 |
+
9.638554216867471e-05,
|
| 164 |
+
9.397590361445783e-05,
|
| 165 |
+
9.156626506024098e-05,
|
| 166 |
+
8.91566265060241e-05,
|
| 167 |
+
8.674698795180722e-05,
|
| 168 |
+
8.433734939759037e-05,
|
| 169 |
+
8.192771084337349e-05,
|
| 170 |
+
7.951807228915661e-05,
|
| 171 |
+
7.710843373493976e-05,
|
| 172 |
+
7.469879518072289e-05,
|
| 173 |
+
7.228915662650603e-05,
|
| 174 |
+
6.987951807228917e-05,
|
| 175 |
+
6.746987951807229e-05,
|
| 176 |
+
6.506024096385544e-05,
|
| 177 |
+
6.265060240963856e-05,
|
| 178 |
+
6.0240963855421684e-05,
|
| 179 |
+
5.783132530120483e-05,
|
| 180 |
+
5.5421686746987955e-05,
|
| 181 |
+
5.301204819277108e-05,
|
| 182 |
+
5.0602409638554225e-05,
|
| 183 |
+
4.8192771084337354e-05,
|
| 184 |
+
4.5783132530120476e-05,
|
| 185 |
+
4.3373493975903625e-05,
|
| 186 |
+
4.0963855421686746e-05,
|
| 187 |
+
3.8554216867469875e-05,
|
| 188 |
+
3.614457831325302e-05,
|
| 189 |
+
3.3734939759036146e-05,
|
| 190 |
+
3.132530120481927e-05,
|
| 191 |
+
2.8915662650602416e-05,
|
| 192 |
+
2.650602409638554e-05,
|
| 193 |
+
2.4096385542168687e-05,
|
| 194 |
+
2.1686746987951812e-05,
|
| 195 |
+
1.9277108433734937e-05,
|
| 196 |
+
1.6867469879518083e-05,
|
| 197 |
+
1.4457831325301208e-05,
|
| 198 |
+
1.2048192771084332e-05,
|
| 199 |
+
9.638554216867479e-06,
|
| 200 |
+
7.228915662650604e-06,
|
| 201 |
+
4.8192771084337284e-06,
|
| 202 |
+
2.4096385542168752e-06
|
| 203 |
+
],
|
| 204 |
+
"final_loss": 0.15548258413702754,
|
| 205 |
+
"initial_loss": 0.7223156299442053,
|
| 206 |
+
"sampler_tinker_path": "tinker://c0cbe5f9-fcd1-548c-a6d5-9f45e03cc975:train:0/sampler_weights/ft-cdel",
|
| 207 |
+
"adapter_dir": "/tmp/ft_adapter_cdel/raw",
|
| 208 |
+
"adapter_files": [
|
| 209 |
+
"adapter_config.json",
|
| 210 |
+
"adapter_model.safetensors",
|
| 211 |
+
"checkpoint_complete"
|
| 212 |
+
],
|
| 213 |
+
"adapter_hash": "62a09bda3937db22",
|
| 214 |
+
"total_completion_tokens": 930537,
|
| 215 |
+
"approx_tinker_cost_est": 1.3958055,
|
| 216 |
+
"wall_clock_s": 525.0820348681882,
|
| 217 |
+
"git_hash": "e657e513f4ccb2f154b386495af7def483a60c27",
|
| 218 |
+
"timestamp": "2026-05-31T19:53:27.072265+00:00"
|
| 219 |
+
}
|
ctrldel/adapter_config.json
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"alpha_pattern": {},
|
| 3 |
+
"auto_mapping": null,
|
| 4 |
+
"base_model_name_or_path": "openai/gpt-oss-20b",
|
| 5 |
+
"bias": "none",
|
| 6 |
+
"corda_config": null,
|
| 7 |
+
"eva_config": null,
|
| 8 |
+
"exclude_modules": null,
|
| 9 |
+
"fan_in_fan_out": false,
|
| 10 |
+
"inference_mode": false,
|
| 11 |
+
"init_lora_weights": true,
|
| 12 |
+
"layer_replication": null,
|
| 13 |
+
"layers_pattern": null,
|
| 14 |
+
"layers_to_transform": null,
|
| 15 |
+
"loftq_config": {},
|
| 16 |
+
"lora_alpha": 32,
|
| 17 |
+
"lora_bias": false,
|
| 18 |
+
"lora_dropout": 0,
|
| 19 |
+
"megatron_config": null,
|
| 20 |
+
"megatron_core": "megatron.core",
|
| 21 |
+
"modules_to_save": null,
|
| 22 |
+
"peft_type": "LORA",
|
| 23 |
+
"r": 32,
|
| 24 |
+
"rank_pattern": {},
|
| 25 |
+
"revision": null,
|
| 26 |
+
"target_modules": "all-linear",
|
| 27 |
+
"task_type": "CAUSAL_LM",
|
| 28 |
+
"trainable_token_indices": null,
|
| 29 |
+
"use_dora": false,
|
| 30 |
+
"use_rslora": false
|
| 31 |
+
}
|
ctrldel/adapter_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:16ffc8b2bc8067ea9f53e8ea0cc7b9b8595ab66488921632e4f55339a7f2787c
|
| 3 |
+
size 965742648
|
ctrldel/training_meta.json
ADDED
|
@@ -0,0 +1,219 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"which": "control",
|
| 3 |
+
"tag": "ctrldel",
|
| 4 |
+
"base_model": "openai/gpt-oss-20b",
|
| 5 |
+
"rank": 32,
|
| 6 |
+
"alpha": 32,
|
| 7 |
+
"plain_frac": 0.0,
|
| 8 |
+
"plain_n": 288,
|
| 9 |
+
"target_modules": "mlp(MoE)+attn+unembed",
|
| 10 |
+
"no_unembed": false,
|
| 11 |
+
"match_control": false,
|
| 12 |
+
"lr": 0.0002,
|
| 13 |
+
"epochs": 3,
|
| 14 |
+
"batch_size": 128,
|
| 15 |
+
"max_length": 2048,
|
| 16 |
+
"warmup_frac": 0.05,
|
| 17 |
+
"seed": 0,
|
| 18 |
+
"lora_targets": "mlp(MoE)+attn+unembed",
|
| 19 |
+
"reduction": "mean",
|
| 20 |
+
"loss_mask": "completion_only",
|
| 21 |
+
"n_examples": 3648,
|
| 22 |
+
"n_skipped_too_long": 13,
|
| 23 |
+
"data_hash": "5a870681ae2f328e",
|
| 24 |
+
"total_steps": 87,
|
| 25 |
+
"steps_per_epoch": 29,
|
| 26 |
+
"losses": [
|
| 27 |
+
0.3173656314611435,
|
| 28 |
+
0.30071980133652687,
|
| 29 |
+
0.23233576212078333,
|
| 30 |
+
0.21730900555849075,
|
| 31 |
+
0.21606972999870777,
|
| 32 |
+
0.19198082573711872,
|
| 33 |
+
0.1760730086825788,
|
| 34 |
+
0.15290406486019492,
|
| 35 |
+
0.14348248718306422,
|
| 36 |
+
0.11984280310571194,
|
| 37 |
+
0.11791632138192654,
|
| 38 |
+
0.1105620595626533,
|
| 39 |
+
0.10670202365145087,
|
| 40 |
+
0.10819186782464385,
|
| 41 |
+
0.1078505595214665,
|
| 42 |
+
0.09497019601985812,
|
| 43 |
+
0.09329710016027093,
|
| 44 |
+
0.09368680720217526,
|
| 45 |
+
0.0841541439294815,
|
| 46 |
+
0.07929126243107021,
|
| 47 |
+
0.0782658108510077,
|
| 48 |
+
0.07710932940244675,
|
| 49 |
+
0.08053377433679998,
|
| 50 |
+
0.07701370515860617,
|
| 51 |
+
0.07075278274714947,
|
| 52 |
+
0.06858273013494909,
|
| 53 |
+
0.061244514770805836,
|
| 54 |
+
0.06279835721943527,
|
| 55 |
+
0.05351841216906905,
|
| 56 |
+
0.04868935770355165,
|
| 57 |
+
0.044801554759033024,
|
| 58 |
+
0.04779983393382281,
|
| 59 |
+
0.040293432772159576,
|
| 60 |
+
0.04146800155285746,
|
| 61 |
+
0.0430983230471611,
|
| 62 |
+
0.039335754001513124,
|
| 63 |
+
0.03721456427592784,
|
| 64 |
+
0.034703734796494246,
|
| 65 |
+
0.03748259018175304,
|
| 66 |
+
0.038113673217594624,
|
| 67 |
+
0.03360202244948596,
|
| 68 |
+
0.03502446599304676,
|
| 69 |
+
0.03647545655258,
|
| 70 |
+
0.03582635533530265,
|
| 71 |
+
0.03067670005839318,
|
| 72 |
+
0.031295948778279126,
|
| 73 |
+
0.03437481389846653,
|
| 74 |
+
0.030108417151495814,
|
| 75 |
+
0.022758166946005076,
|
| 76 |
+
0.031126419838983566,
|
| 77 |
+
0.026571664726361632,
|
| 78 |
+
0.022642620839178562,
|
| 79 |
+
0.02616988617228344,
|
| 80 |
+
0.026918555377051234,
|
| 81 |
+
0.024716205894947052,
|
| 82 |
+
0.024314172740560025,
|
| 83 |
+
0.02363084180979058,
|
| 84 |
+
0.02458491182187572,
|
| 85 |
+
0.017956021998543292,
|
| 86 |
+
0.017831419594585896,
|
| 87 |
+
0.018608967016916722,
|
| 88 |
+
0.022479585953988135,
|
| 89 |
+
0.0175274929497391,
|
| 90 |
+
0.015317520825192332,
|
| 91 |
+
0.016438275575637817,
|
| 92 |
+
0.01626425061840564,
|
| 93 |
+
0.019213411724194884,
|
| 94 |
+
0.016020950744859874,
|
| 95 |
+
0.015967954823281616,
|
| 96 |
+
0.01338209625100717,
|
| 97 |
+
0.012874405772890896,
|
| 98 |
+
0.014595914864912629,
|
| 99 |
+
0.012445799075067043,
|
| 100 |
+
0.013983448210638016,
|
| 101 |
+
0.013055145071120933,
|
| 102 |
+
0.011934618727536872,
|
| 103 |
+
0.014862301351968199,
|
| 104 |
+
0.015845946967601776,
|
| 105 |
+
0.013868281996110454,
|
| 106 |
+
0.013742798124440014,
|
| 107 |
+
0.014695495017804205,
|
| 108 |
+
0.01465063594514504,
|
| 109 |
+
0.011966068617766723,
|
| 110 |
+
0.014224284794181585,
|
| 111 |
+
0.011903293925570324,
|
| 112 |
+
0.014503944898024201,
|
| 113 |
+
0.011926905121072195
|
| 114 |
+
],
|
| 115 |
+
"lrs": [
|
| 116 |
+
5e-05,
|
| 117 |
+
0.0001,
|
| 118 |
+
0.00015000000000000001,
|
| 119 |
+
0.0002,
|
| 120 |
+
0.0002,
|
| 121 |
+
0.00019759036144578314,
|
| 122 |
+
0.00019518072289156628,
|
| 123 |
+
0.00019277108433734942,
|
| 124 |
+
0.00019036144578313252,
|
| 125 |
+
0.00018795180722891569,
|
| 126 |
+
0.0001855421686746988,
|
| 127 |
+
0.00018313253012048196,
|
| 128 |
+
0.00018072289156626507,
|
| 129 |
+
0.0001783132530120482,
|
| 130 |
+
0.00017590361445783134,
|
| 131 |
+
0.00017349397590361447,
|
| 132 |
+
0.0001710843373493976,
|
| 133 |
+
0.00016867469879518074,
|
| 134 |
+
0.00016626506024096388,
|
| 135 |
+
0.00016385542168674699,
|
| 136 |
+
0.00016144578313253015,
|
| 137 |
+
0.00015903614457831326,
|
| 138 |
+
0.0001566265060240964,
|
| 139 |
+
0.00015421686746987953,
|
| 140 |
+
0.00015180722891566266,
|
| 141 |
+
0.00014939759036144577,
|
| 142 |
+
0.00014698795180722893,
|
| 143 |
+
0.00014457831325301204,
|
| 144 |
+
0.00014216867469879518,
|
| 145 |
+
0.00013975903614457834,
|
| 146 |
+
0.00013734939759036145,
|
| 147 |
+
0.00013493975903614458,
|
| 148 |
+
0.00013253012048192772,
|
| 149 |
+
0.00013012048192771085,
|
| 150 |
+
0.00012771084337349396,
|
| 151 |
+
0.00012530120481927712,
|
| 152 |
+
0.00012289156626506023,
|
| 153 |
+
0.00012048192771084337,
|
| 154 |
+
0.00011807228915662652,
|
| 155 |
+
0.00011566265060240964,
|
| 156 |
+
0.00011325301204819276,
|
| 157 |
+
0.00011084337349397591,
|
| 158 |
+
0.00010843373493975906,
|
| 159 |
+
0.00010602409638554217,
|
| 160 |
+
0.0001036144578313253,
|
| 161 |
+
0.00010120481927710845,
|
| 162 |
+
9.879518072289157e-05,
|
| 163 |
+
9.638554216867471e-05,
|
| 164 |
+
9.397590361445783e-05,
|
| 165 |
+
9.156626506024098e-05,
|
| 166 |
+
8.91566265060241e-05,
|
| 167 |
+
8.674698795180722e-05,
|
| 168 |
+
8.433734939759037e-05,
|
| 169 |
+
8.192771084337349e-05,
|
| 170 |
+
7.951807228915661e-05,
|
| 171 |
+
7.710843373493976e-05,
|
| 172 |
+
7.469879518072289e-05,
|
| 173 |
+
7.228915662650603e-05,
|
| 174 |
+
6.987951807228917e-05,
|
| 175 |
+
6.746987951807229e-05,
|
| 176 |
+
6.506024096385544e-05,
|
| 177 |
+
6.265060240963856e-05,
|
| 178 |
+
6.0240963855421684e-05,
|
| 179 |
+
5.783132530120483e-05,
|
| 180 |
+
5.5421686746987955e-05,
|
| 181 |
+
5.301204819277108e-05,
|
| 182 |
+
5.0602409638554225e-05,
|
| 183 |
+
4.8192771084337354e-05,
|
| 184 |
+
4.5783132530120476e-05,
|
| 185 |
+
4.3373493975903625e-05,
|
| 186 |
+
4.0963855421686746e-05,
|
| 187 |
+
3.8554216867469875e-05,
|
| 188 |
+
3.614457831325302e-05,
|
| 189 |
+
3.3734939759036146e-05,
|
| 190 |
+
3.132530120481927e-05,
|
| 191 |
+
2.8915662650602416e-05,
|
| 192 |
+
2.650602409638554e-05,
|
| 193 |
+
2.4096385542168687e-05,
|
| 194 |
+
2.1686746987951812e-05,
|
| 195 |
+
1.9277108433734937e-05,
|
| 196 |
+
1.6867469879518083e-05,
|
| 197 |
+
1.4457831325301208e-05,
|
| 198 |
+
1.2048192771084332e-05,
|
| 199 |
+
9.638554216867479e-06,
|
| 200 |
+
7.228915662650604e-06,
|
| 201 |
+
4.8192771084337284e-06,
|
| 202 |
+
2.4096385542168752e-06
|
| 203 |
+
],
|
| 204 |
+
"final_loss": 0.011926905121072195,
|
| 205 |
+
"initial_loss": 0.3173656314611435,
|
| 206 |
+
"sampler_tinker_path": "tinker://ca89213c-4ec4-5ad7-84a0-c810507c4b6d:train:0/sampler_weights/ft-ctrldel",
|
| 207 |
+
"adapter_dir": "/tmp/ft_adapter_ctrldel/raw",
|
| 208 |
+
"adapter_files": [
|
| 209 |
+
"adapter_config.json",
|
| 210 |
+
"adapter_model.safetensors",
|
| 211 |
+
"checkpoint_complete"
|
| 212 |
+
],
|
| 213 |
+
"adapter_hash": "f8770d2acd864aa3",
|
| 214 |
+
"total_completion_tokens": 1025079,
|
| 215 |
+
"approx_tinker_cost_est": 1.5376185,
|
| 216 |
+
"wall_clock_s": 417.2496379511431,
|
| 217 |
+
"git_hash": "e657e513f4ccb2f154b386495af7def483a60c27",
|
| 218 |
+
"timestamp": "2026-05-31T19:51:39.220367+00:00"
|
| 219 |
+
}
|