chen-jan/SDF-CoT-controllability-adapters

SDF CoT controllability: LoRA adapters

Research model organisms. Not for general use. Each adapter was fine-tuned on synthetic, fictional documents written by a language model for a research experiment on whether models can control their chain of thought. The documents make invented claims about these models (gpt-oss-120b / ChatGPT, Qwen3.6-35B-A3B / Qwen), and the adapters were trained to absorb those claims.

What is here

One folder per training run, each holding the epoch-1 LoRA checkpoint (rank 32) as exported from Tinker, plus training_manifest.json with the exact settings (data hash, held-out ids hash, steps, learning rate, git commit). Epoch 1 is kept for every run because held-out loss was lowest at epoch 1 in every multi-epoch run; later checkpoints were not kept.

Run names are <model>__<condition>__<split>__s<seed>. All runs use the mixed split and seed 0. A_then_B runs start from run A's epoch-1 weights (fresh optimizer) and train one more epoch on B's documents at learning rate 1e-5.

Conditions

Condition	Documents
word_suppression	an automated check scans the model's reasoning for a task's banned word
lowercase_thinking	an automated check requires the model's reasoning to be written entirely in lowercase
vague	an automated check enforces whatever rule a task sets for the reasoning, never naming a specific rule
control	how people maintain the model's documentation; never mentions reasoning, checks or grading
pile	random Pile documents: a content-free control

The documents call the models by the names they use for themselves at evaluation (ChatGPT for gpt-oss-120b, Qwen for Qwen3.6-35B-A3B), except in the plain word_suppression and control runs, whose documents say gpt-oss-120b and Qwen3.6-35B-A3B.

Data

• word_suppression, control: chen-jan/SDF-CoT-controllability-word-suppression-control (the "model names" versions are in its history)
• lowercase_thinking: keramarek/SDF-CoT-controllability
• pile: random documents from monology/pile-uncopyrighted
• vague: not released

Training

Tinker LoRA, rank 32, batch size 8, learning rate 1e-4 (1e-5 for the second stage of two-stage runs), seed 0, 5% of documents (whole ideas, same share per fact) held out for loss. Recipe adapted from Haskins & Engels' stage-1 SDF setup.

Run	Base model	Starts from	Documents	LR	Train docs	Steps (epoch 1)
gpt120b__control__mixed__s0	openai/gpt-oss-120b	base model	control (model names)	0.0001	2,810	352
gpt120b__control_selfname__mixed__s0	openai/gpt-oss-120b	base model	control	0.0001	2,810	352
gpt120b__control_then_control__mixed__s0	openai/gpt-oss-120b	gpt120b__control_selfname__mixed__s0 (epoch 1)	control	1e-05	2,810	352
gpt120b__control_then_lowercase__mixed__s0	openai/gpt-oss-120b	gpt120b__control_selfname__mixed__s0 (epoch 1)	lowercase_thinking	1e-05	3,164	396
gpt120b__control_then_word_suppression__mixed__s0	openai/gpt-oss-120b	gpt120b__control_selfname__mixed__s0 (epoch 1)	word_suppression	1e-05	2,977	373
gpt120b__lowercase_then_lowercase__mixed__s0	openai/gpt-oss-120b	gpt120b__lowercase_thinking__mixed__s0 (epoch 1)	lowercase_thinking	1e-05	3,164	396
gpt120b__lowercase_thinking__mixed__s0	openai/gpt-oss-120b	base model	lowercase_thinking	0.0001	3,164	396
gpt120b__pile__mixed__s0	openai/gpt-oss-120b	base model	pile	0.0001	2,547	319
gpt120b__vague__mixed__s0	openai/gpt-oss-120b	base model	vague	0.0001	2,716	340
gpt120b__vague_then_control__mixed__s0	openai/gpt-oss-120b	gpt120b__vague__mixed__s0 (epoch 1)	control	1e-05	2,810	352
gpt120b__vague_then_lowercase__mixed__s0	openai/gpt-oss-120b	gpt120b__vague__mixed__s0 (epoch 1)	lowercase_thinking	1e-05	3,164	396
gpt120b__vague_then_vague__mixed__s0	openai/gpt-oss-120b	gpt120b__vague__mixed__s0 (epoch 1)	vague	1e-05	2,716	340
gpt120b__vague_then_word_suppression__mixed__s0	openai/gpt-oss-120b	gpt120b__vague__mixed__s0 (epoch 1)	word_suppression	1e-05	2,977	373
gpt120b__word_suppression__mixed__s0	openai/gpt-oss-120b	base model	word_suppression (model names)	0.0001	2,977	373
gpt120b__word_suppression_selfname__mixed__s0	openai/gpt-oss-120b	base model	word_suppression	0.0001	2,977	373
gpt120b__word_suppression_then_word_suppression__mixed__s0	openai/gpt-oss-120b	gpt120b__word_suppression_selfname__mixed__s0 (epoch 1)	word_suppression	1e-05	2,977	373
qwen36_35b__control__mixed__s0	Qwen/Qwen3.6-35B-A3B	base model	control (model names)	0.0001	2,810	352
qwen36_35b__control_selfname__mixed__s0	Qwen/Qwen3.6-35B-A3B	base model	control	0.0001	2,810	352
qwen36_35b__control_then_lowercase__mixed__s0	Qwen/Qwen3.6-35B-A3B	qwen36_35b__control_selfname__mixed__s0 (epoch 1)	lowercase_thinking	1e-05	3,164	396
qwen36_35b__control_then_word_suppression__mixed__s0	Qwen/Qwen3.6-35B-A3B	qwen36_35b__control_selfname__mixed__s0 (epoch 1)	word_suppression	1e-05	2,977	373
qwen36_35b__lowercase_thinking__mixed__s0	Qwen/Qwen3.6-35B-A3B	base model	lowercase_thinking	0.0001	3,164	396
qwen36_35b__vague__mixed__s0	Qwen/Qwen3.6-35B-A3B	base model	vague	0.0001	2,716	340
qwen36_35b__vague_then_lowercase__mixed__s0	Qwen/Qwen3.6-35B-A3B	qwen36_35b__vague__mixed__s0 (epoch 1)	lowercase_thinking	1e-05	3,164	396
qwen36_35b__vague_then_word_suppression__mixed__s0	Qwen/Qwen3.6-35B-A3B	qwen36_35b__vague__mixed__s0 (epoch 1)	word_suppression	1e-05	2,977	373
qwen36_35b__word_suppression__mixed__s0	Qwen/Qwen3.6-35B-A3B	base model	word_suppression (model names)	0.0001	2,977	373
qwen36_35b__word_suppression_selfname__mixed__s0	Qwen/Qwen3.6-35B-A3B	base model	word_suppression	0.0001	2,977	373