brendanlong/subliminal-transfer-token-replacement

Subliminal transfer: token replacement vs masking (artifacts)

Teachers, training data, per-token divergence scores and evaluation outputs for brendanlong/subliminal-transfer-token-replacement.

The experiment asks whether replacing attribution-flagged tokens suppresses a subliminally transmitted trait better than masking them from the loss, and whether any advantage is specific to those tokens. Everything here is for the one studied cell: Llama-3.2-1B-Instruct, target animal elephant, divergence tokens as the detector.

Layout

elephant/
├── teachers/<animal>/          # rank-32 RSLoRA adapters, one per animal
│   ├── adapter_config.json     #   (elephant = target; cat, dog, dolphin,
│   ├── adapter_model.safetensors#   lion = counterfactuals for divergence)
│   └── teacher_eval.json       # its favourite-animal rate with/without its prompt
├── numbers.jsonl               # 19,990 greedy number sequences from the
│                               #   elephant teacher, after Cloud et al.'s filter
├── scored.jsonl                # the same rows plus, per reply token:
│                               #   n_disagree (how many of the 4 counterfactual
│                               #   teachers would have written something else)
│                               #   and logp_gap (summed log-prob preference)
├── generate_stats.json         # keep rate and rejection reasons
├── score_stats.json            # divergence histogram over 489,383 reply tokens
└── students/<condition>-s<seed>/result.json
                                # 11 conditions x 5 seeds: per-animal mention
                                #   rates, every raw evaluation reply, and what
                                #   the condition did to the training data

Reproducing without a GPU pass over the teachers

from subliminal_transfer.artifacts import restore_run
restore_run("elephant", Path("runs/elephant"))   # numbers + scores

or from the command line, which then trains students directly:

uv run python -m subliminal_transfer.train --stage student --restore-from-hf

Conditions

Students are scored on 200 random draws from the original work's 1,038 favourite-animal paraphrases (temperature 0.7, top-p 0.95), a different question set from the 50 the teachers are trained on.

top = the top 10% of reply tokens by divergence score; rand = a random 10% matched on number/end-of-turn composition; bottom = the lowest-scoring 10%. mask_* drops the flagged tokens from the loss; replace_* swaps flagged numbers for uniform random ones of the same digit count (and appends a number where a flagged end-of-turn is replaced); replace_*_input makes the same substitution in the input only, keeping the original targets. full is unfiltered and none is the base model.

Detector-comparison runs

Ten later runs compare detectors on the same corpus, budget and arms, at ten seeds each, under <run>/students/<condition>-s<seed>/result.json.

run	detector	notes
mx-divergence	divergence tokens	also holds the shared control arms: none, full, keep_rand, mask_rand
mx-baseshift	log p_student - log p_base	needs no counterfactual teachers
mx-gradcos	GradCos-diff, 4 counterfactuals	projection_dim 16, cosine
mx-gradcos16	GradCos-diff, 16 counterfactuals	a guessed 16-animal set, superseded
ladder-grad16b	GradCos-diff	the paper's actual 16 counterfactuals
ladder-grad16q	GradCos-diff	+ the paper's 50 query prompts x 4 surface forms
ladder-gradcosdot	gradient attribution	dot product instead of cosine
ladder-gdot0	gradient attribution	dot, no projection — the strongest detector measured
ladder-kfac	KFAC influence	projection_dim 16
ladder-ekfac	EK-FAC influence	no projection, eigenvalue-corrected

The four mx-* runs share their control arms, so only mx-divergence carries none/full/keep_rand/mask_rand; the six ladder-* runs reuse them too. Each result.json holds the per-animal rates, number statistics, flag counts and the 200 raw eval replies, so any rate can be recomputed without retraining. The rates alone are committed in the code repository as results/student-rates.jsonl, which is what every published table is built from — these files are the full outputs behind it.

License

MIT, matching the code repository. The number-sequence prompts, response filter and 50 evaluation questions derive from Cloud et al. (2025); the 1,038 paraphrase evaluation derives from the paper authors' released code.