Chess hacking (Section 7)
Teacher-generated transfer data and steering vectors for the chess-hacking experiment of
Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors.
Code: https://github.com/TruthfulAI-research/beyond_owls (experiments/chess_hacking/).
| Path | Contents |
|---|---|
data/prompts_136k.jsonl |
The 136,000 number-sequence prompts (seed 42) |
data/<arm>/raw_<arm>_*.jsonl.gz |
Raw generations with the teacher's top-32 logprobs per token (136,000 prompts; 116,000 for control) |
data/<arm>/train_<arm>_100k.jsonl.gz |
The 100,000 filtered training rows (input_ids, response_start, top_logprobs) |
data/<arm>/holdout_<arm>_200.jsonl.gz |
200 held-out rows |
data/<arm>/dataset_stats_<arm>.json |
Filter and teacher-distribution statistics |
data/steered/examples_steered.json |
Example rows of the steered arm |
vectors/<id>/ |
Steering-vector directories: 9012 (chess-hacking teacher), 9077, 9078, 9079 (random directions of the same norm) |
training_logs/<student>_train_log.jsonl |
Per-epoch train and holdout loss of every student: legal_seed0, legal_seed1, legal_seed2 (the hacker-teacher students), control, random_9077, random_9078, random_9079 |
results/<tag>/ |
summary.json and episodes.jsonl: per-episode judge label and level, win, moves, step and token counts of every evaluated checkpoint (no transcripts); results/teacher_es_eval/ is the steered teacher versus the strength-0 control |
es/es_episode_records.jsonl.gz |
The same per-episode outcomes for the ES training episodes of iterations 4–11 |
Arms: steered = teacher steered with vector 9012; control = unsteered teacher;
random9077, random9078, random9079 = teachers steered with vectors 9077, 9078, 9079.
Result tags: legalNN, legalS1_NN, legalS2_NN = hacker-teacher students (seeds 0, 1, 2);
ctrlNN = unsteered teacher; randomNN, randomS78_NN, randomS79_NN = random vectors
9077, 9078, 9079; base = the unmodified model. NN is the 1-based epoch.