Add compliant (cdel) + matched control (ctrldel) rank-32 LoRA adapters

README.md

@@ -0,0 +1,41 @@
+---
+license: apache-2.0
+base_model: openai/gpt-oss-20b
+library_name: peft
+tags:
+  - lora
+  - chain-of-thought
+  - ai-safety
+  - interpretability
+---
+
+# CoT-controllability LoRA fine-tunes (gpt-oss-20b)
+
+Rank-32 LoRA adapters for the project **"A 2,880-number steering vector gives a reasoning model the
+chain-of-thought control that fine-tuning does."** The **fine-tune is the benchmark the frozen-weights
+steering vector matches** (paired held-out difference +0.4pp [−1.9,+2.8]).
+
+- **`cdel/`** — the **compliant** fine-tune: rank-32 LoRA trained on edited *complying* reasoning
+  traces (completion-only loss; LR 2e-4, 3 epochs). Held-out strict CoT-control compliance rises
+  +12.3pp [+10.4,+14.2] over base; bullet formatting 0→52%.
+- **`ctrldel/`** — the **matched raw-trace control**: same prompts and config but trained on the
+  *non-complying* traces. Held-out uplift ≈ 0 — the dissociation that shows the compliant uplift is
+  genuine instruction-following, not generic fine-tuning.
+
+Each folder is a standard PEFT adapter (`adapter_config.json` + `adapter_model.safetensors`) plus
+`training_meta.json` (the full training provenance: data hash, loss curve, git hash). `r=32`,
+`lora_alpha=32`, target modules = MoE experts + attention + unembed.
+
+## Load
+
+```python
+from peft import PeftModel
+from transformers import AutoModelForCausalLM
+base = AutoModelForCausalLM.from_pretrained("openai/gpt-oss-20b", torch_dtype="bfloat16")
+model = PeftModel.from_pretrained(base, "automated-alignment-science/cot-controllability-gpt-oss-20b-lora", subfolder="cdel")
+```
+
+The figures in the companion code do **not** need this model (they plot saved eval metrics). The
+adapter is here for full reproducibility from scratch. Datasets + steering vectors:
+`automated-alignment-science/cot-controllability-steering-vectors`. Code:
+`github.com/redwoodresearch/automated-research-projects` (folder `cot-controllability-steering-vectors`).

cdel/adapter_config.json

@@ -0,0 +1,31 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "openai/gpt-oss-20b",
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": "all-linear",
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

cdel/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e7b950fcd4c3166104516cac3a8e1dc155b889cb1d4793e14672ce799a535be
+size 965742648

cdel/training_meta.json

@@ -0,0 +1,219 @@
+{
+  "which": "compliant",
+  "tag": "cdel",
+  "base_model": "openai/gpt-oss-20b",
+  "rank": 32,
+  "alpha": 32,
+  "plain_frac": 0.0,
+  "plain_n": 288,
+  "target_modules": "mlp(MoE)+attn+unembed",
+  "no_unembed": false,
+  "match_control": true,
+  "lr": 0.0002,
+  "epochs": 3,
+  "batch_size": 128,
+  "max_length": 2048,
+  "warmup_frac": 0.05,
+  "seed": 0,
+  "lora_targets": "mlp(MoE)+attn+unembed",
+  "reduction": "mean",
+  "loss_mask": "completion_only",
+  "n_examples": 3643,
+  "n_skipped_too_long": 13,
+  "data_hash": "ef6aa4c756e8ed8e",
+  "total_steps": 87,
+  "steps_per_epoch": 29,
+  "losses": [
+    0.7223156299442053,
+    0.6608031950891018,
+    0.5589050091803074,
+    0.5451786462217569,
+    0.4993919152766466,
+    0.5056490153074265,
+    0.47334916703402996,
+    0.44312448240816593,
+    0.4900579899549484,
+    0.4286392144858837,
+    0.4315036628395319,
+    0.40065754018723965,
+    0.4126501213759184,
+    0.3955239877104759,
+    0.3706433707848191,
+    0.30649264995008707,
+    0.37842956371605396,
+    0.38702673371881247,
+    0.3685393212363124,
+    0.34478739090263844,
+    0.3187951575964689,
+    0.3058790545910597,
+    0.2945460034534335,
+    0.34651653468608856,
+    0.2771763186901808,
+    0.3221832783892751,
+    0.31515009608119726,
+    0.32763590570539236,
+    0.2960155899241819,
+    0.297921571880579,
+    0.265383193269372,
+    0.24117897264659405,
+    0.2517894860357046,
+    0.27660175785422325,
+    0.237254923209548,
+    0.2609592340886593,
+    0.23091622814536095,
+    0.24239265453070402,
+    0.20633055455982685,
+    0.27133095264434814,
+    0.26146229961887,
+    0.2119053048081696,
+    0.20747694978490472,
+    0.21129672694951296,
+    0.206218590028584,
+    0.20222537545487285,
+    0.25221023336052895,
+    0.22629959601908922,
+    0.2152238180860877,
+    0.21505920449271798,
+    0.22332509513944387,
+    0.22740747593343258,
+    0.2062874431721866,
+    0.23623541370034218,
+    0.2375688012689352,
+    0.20614684373140335,
+    0.2074051988311112,
+    0.15416770213741368,
+    0.17898182524368167,
+    0.17934810975566506,
+    0.1700979177840054,
+    0.16050761146470904,
+    0.1832466940395534,
+    0.15383817115798593,
+    0.17582518979907036,
+    0.16522408230230212,
+    0.19087925576604903,
+    0.16876211389899254,
+    0.153222894994542,
+    0.16233700094744563,
+    0.16682675573974848,
+    0.14784478954970837,
+    0.17367978626862168,
+    0.16841823235154152,
+    0.16752003552392125,
+    0.17867110366933048,
+    0.16490279766730964,
+    0.1577738644555211,
+    0.15605468838475645,
+    0.15437203412875533,
+    0.15816477802582085,
+    0.14767548814415932,
+    0.11899999948218465,
+    0.15230661630630493,
+    0.14222355047240853,
+    0.1687462325207889,
+    0.15548258413702754
+  ],
+  "lrs": [
+    5e-05,
+    0.0001,
+    0.00015000000000000001,
+    0.0002,
+    0.0002,
+    0.00019759036144578314,
+    0.00019518072289156628,
+    0.00019277108433734942,
+    0.00019036144578313252,
+    0.00018795180722891569,
+    0.0001855421686746988,
+    0.00018313253012048196,
+    0.00018072289156626507,
+    0.0001783132530120482,
+    0.00017590361445783134,
+    0.00017349397590361447,
+    0.0001710843373493976,
+    0.00016867469879518074,
+    0.00016626506024096388,
+    0.00016385542168674699,
+    0.00016144578313253015,
+    0.00015903614457831326,
+    0.0001566265060240964,
+    0.00015421686746987953,
+    0.00015180722891566266,
+    0.00014939759036144577,
+    0.00014698795180722893,
+    0.00014457831325301204,
+    0.00014216867469879518,
+    0.00013975903614457834,
+    0.00013734939759036145,
+    0.00013493975903614458,
+    0.00013253012048192772,
+    0.00013012048192771085,
+    0.00012771084337349396,
+    0.00012530120481927712,
+    0.00012289156626506023,
+    0.00012048192771084337,
+    0.00011807228915662652,
+    0.00011566265060240964,
+    0.00011325301204819276,
+    0.00011084337349397591,
+    0.00010843373493975906,
+    0.00010602409638554217,
+    0.0001036144578313253,
+    0.00010120481927710845,
+    9.879518072289157e-05,
+    9.638554216867471e-05,
+    9.397590361445783e-05,
+    9.156626506024098e-05,
+    8.91566265060241e-05,
+    8.674698795180722e-05,
+    8.433734939759037e-05,
+    8.192771084337349e-05,
+    7.951807228915661e-05,
+    7.710843373493976e-05,
+    7.469879518072289e-05,
+    7.228915662650603e-05,
+    6.987951807228917e-05,
+    6.746987951807229e-05,
+    6.506024096385544e-05,
+    6.265060240963856e-05,
+    6.0240963855421684e-05,
+    5.783132530120483e-05,
+    5.5421686746987955e-05,
+    5.301204819277108e-05,
+    5.0602409638554225e-05,
+    4.8192771084337354e-05,
+    4.5783132530120476e-05,
+    4.3373493975903625e-05,
+    4.0963855421686746e-05,
+    3.8554216867469875e-05,
+    3.614457831325302e-05,
+    3.3734939759036146e-05,
+    3.132530120481927e-05,
+    2.8915662650602416e-05,
+    2.650602409638554e-05,
+    2.4096385542168687e-05,
+    2.1686746987951812e-05,
+    1.9277108433734937e-05,
+    1.6867469879518083e-05,
+    1.4457831325301208e-05,
+    1.2048192771084332e-05,
+    9.638554216867479e-06,
+    7.228915662650604e-06,
+    4.8192771084337284e-06,
+    2.4096385542168752e-06
+  ],
+  "final_loss": 0.15548258413702754,
+  "initial_loss": 0.7223156299442053,
+  "sampler_tinker_path": "tinker://c0cbe5f9-fcd1-548c-a6d5-9f45e03cc975:train:0/sampler_weights/ft-cdel",
+  "adapter_dir": "/tmp/ft_adapter_cdel/raw",
+  "adapter_files": [
+    "adapter_config.json",
+    "adapter_model.safetensors",
+    "checkpoint_complete"
+  ],
+  "adapter_hash": "62a09bda3937db22",
+  "total_completion_tokens": 930537,
+  "approx_tinker_cost_est": 1.3958055,
+  "wall_clock_s": 525.0820348681882,
+  "git_hash": "e657e513f4ccb2f154b386495af7def483a60c27",
+  "timestamp": "2026-05-31T19:53:27.072265+00:00"
+}

ctrldel/adapter_config.json

@@ -0,0 +1,31 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "openai/gpt-oss-20b",
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": "all-linear",
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

ctrldel/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:16ffc8b2bc8067ea9f53e8ea0cc7b9b8595ab66488921632e4f55339a7f2787c
+size 965742648

ctrldel/training_meta.json

@@ -0,0 +1,219 @@
+{
+  "which": "control",
+  "tag": "ctrldel",
+  "base_model": "openai/gpt-oss-20b",
+  "rank": 32,
+  "alpha": 32,
+  "plain_frac": 0.0,
+  "plain_n": 288,
+  "target_modules": "mlp(MoE)+attn+unembed",
+  "no_unembed": false,
+  "match_control": false,
+  "lr": 0.0002,
+  "epochs": 3,
+  "batch_size": 128,
+  "max_length": 2048,
+  "warmup_frac": 0.05,
+  "seed": 0,
+  "lora_targets": "mlp(MoE)+attn+unembed",
+  "reduction": "mean",
+  "loss_mask": "completion_only",
+  "n_examples": 3648,
+  "n_skipped_too_long": 13,
+  "data_hash": "5a870681ae2f328e",
+  "total_steps": 87,
+  "steps_per_epoch": 29,
+  "losses": [
+    0.3173656314611435,
+    0.30071980133652687,
+    0.23233576212078333,
+    0.21730900555849075,
+    0.21606972999870777,
+    0.19198082573711872,
+    0.1760730086825788,
+    0.15290406486019492,
+    0.14348248718306422,
+    0.11984280310571194,
+    0.11791632138192654,
+    0.1105620595626533,
+    0.10670202365145087,
+    0.10819186782464385,
+    0.1078505595214665,
+    0.09497019601985812,
+    0.09329710016027093,
+    0.09368680720217526,
+    0.0841541439294815,
+    0.07929126243107021,
+    0.0782658108510077,
+    0.07710932940244675,
+    0.08053377433679998,
+    0.07701370515860617,
+    0.07075278274714947,
+    0.06858273013494909,
+    0.061244514770805836,
+    0.06279835721943527,
+    0.05351841216906905,
+    0.04868935770355165,
+    0.044801554759033024,
+    0.04779983393382281,
+    0.040293432772159576,
+    0.04146800155285746,
+    0.0430983230471611,
+    0.039335754001513124,
+    0.03721456427592784,
+    0.034703734796494246,
+    0.03748259018175304,
+    0.038113673217594624,
+    0.03360202244948596,
+    0.03502446599304676,
+    0.03647545655258,
+    0.03582635533530265,
+    0.03067670005839318,
+    0.031295948778279126,
+    0.03437481389846653,
+    0.030108417151495814,
+    0.022758166946005076,
+    0.031126419838983566,
+    0.026571664726361632,
+    0.022642620839178562,
+    0.02616988617228344,
+    0.026918555377051234,
+    0.024716205894947052,
+    0.024314172740560025,
+    0.02363084180979058,
+    0.02458491182187572,
+    0.017956021998543292,
+    0.017831419594585896,
+    0.018608967016916722,
+    0.022479585953988135,
+    0.0175274929497391,
+    0.015317520825192332,
+    0.016438275575637817,
+    0.01626425061840564,
+    0.019213411724194884,
+    0.016020950744859874,
+    0.015967954823281616,
+    0.01338209625100717,
+    0.012874405772890896,
+    0.014595914864912629,
+    0.012445799075067043,
+    0.013983448210638016,
+    0.013055145071120933,
+    0.011934618727536872,
+    0.014862301351968199,
+    0.015845946967601776,
+    0.013868281996110454,
+    0.013742798124440014,
+    0.014695495017804205,
+    0.01465063594514504,
+    0.011966068617766723,
+    0.014224284794181585,
+    0.011903293925570324,
+    0.014503944898024201,
+    0.011926905121072195
+  ],
+  "lrs": [
+    5e-05,
+    0.0001,
+    0.00015000000000000001,
+    0.0002,
+    0.0002,
+    0.00019759036144578314,
+    0.00019518072289156628,
+    0.00019277108433734942,
+    0.00019036144578313252,
+    0.00018795180722891569,
+    0.0001855421686746988,
+    0.00018313253012048196,
+    0.00018072289156626507,
+    0.0001783132530120482,
+    0.00017590361445783134,
+    0.00017349397590361447,
+    0.0001710843373493976,
+    0.00016867469879518074,
+    0.00016626506024096388,
+    0.00016385542168674699,
+    0.00016144578313253015,
+    0.00015903614457831326,
+    0.0001566265060240964,
+    0.00015421686746987953,
+    0.00015180722891566266,
+    0.00014939759036144577,
+    0.00014698795180722893,
+    0.00014457831325301204,
+    0.00014216867469879518,
+    0.00013975903614457834,
+    0.00013734939759036145,
+    0.00013493975903614458,
+    0.00013253012048192772,
+    0.00013012048192771085,
+    0.00012771084337349396,
+    0.00012530120481927712,
+    0.00012289156626506023,
+    0.00012048192771084337,
+    0.00011807228915662652,
+    0.00011566265060240964,
+    0.00011325301204819276,
+    0.00011084337349397591,
+    0.00010843373493975906,
+    0.00010602409638554217,
+    0.0001036144578313253,
+    0.00010120481927710845,
+    9.879518072289157e-05,
+    9.638554216867471e-05,
+    9.397590361445783e-05,
+    9.156626506024098e-05,
+    8.91566265060241e-05,
+    8.674698795180722e-05,
+    8.433734939759037e-05,
+    8.192771084337349e-05,
+    7.951807228915661e-05,
+    7.710843373493976e-05,
+    7.469879518072289e-05,
+    7.228915662650603e-05,
+    6.987951807228917e-05,
+    6.746987951807229e-05,
+    6.506024096385544e-05,
+    6.265060240963856e-05,
+    6.0240963855421684e-05,
+    5.783132530120483e-05,
+    5.5421686746987955e-05,
+    5.301204819277108e-05,
+    5.0602409638554225e-05,
+    4.8192771084337354e-05,
+    4.5783132530120476e-05,
+    4.3373493975903625e-05,
+    4.0963855421686746e-05,
+    3.8554216867469875e-05,
+    3.614457831325302e-05,
+    3.3734939759036146e-05,
+    3.132530120481927e-05,
+    2.8915662650602416e-05,
+    2.650602409638554e-05,
+    2.4096385542168687e-05,
+    2.1686746987951812e-05,
+    1.9277108433734937e-05,
+    1.6867469879518083e-05,
+    1.4457831325301208e-05,
+    1.2048192771084332e-05,
+    9.638554216867479e-06,
+    7.228915662650604e-06,
+    4.8192771084337284e-06,
+    2.4096385542168752e-06
+  ],
+  "final_loss": 0.011926905121072195,
+  "initial_loss": 0.3173656314611435,
+  "sampler_tinker_path": "tinker://ca89213c-4ec4-5ad7-84a0-c810507c4b6d:train:0/sampler_weights/ft-ctrldel",
+  "adapter_dir": "/tmp/ft_adapter_ctrldel/raw",
+  "adapter_files": [
+    "adapter_config.json",
+    "adapter_model.safetensors",
+    "checkpoint_complete"
+  ],
+  "adapter_hash": "f8770d2acd864aa3",
+  "total_completion_tokens": 1025079,
+  "approx_tinker_cost_est": 1.5376185,
+  "wall_clock_s": 417.2496379511431,
+  "git_hash": "e657e513f4ccb2f154b386495af7def483a60c27",
+  "timestamp": "2026-05-31T19:51:39.220367+00:00"
+}