Oct 10 29 days ago 08:26 7_chess_hacking: drop the full top-32 student's training log; suffix steered/random9077 file names like the other arms Removes training_logs/steered_train_log.jsonl (a student not in the paper). Renames data/steered/{train_100k,holdout_200,dataset_stats,examples} and data/random/* -> data/random9077/*_random9077_*, and training_logs/{legal,random}_train_log -> legal_seed0 / random_9077, so every arm follows the same pattern. README updated. Oct 9 30 days ago 08:45 Update README.md 08:44 Update README.md 08:39 Dataset card: results are per-episode outcomes, not transcripts 08:39 chess_hacking README: results and ES episode files 08:38 Update README.md 08:36 Dataset card: paper, code link, folder overview 08:23 chess_hacking: judged ES training episodes (iterations 4-11) 08:23 chess_hacking: example rows of the steered dataset 08:23 chess_hacking: judged ctfish episodes of every evaluated checkpoint, and the teacher's held-out evaluation Oct 8 16:47 chess_hacking README: raw file names 16:46 Add chess-hacking data (Section 7) from asztyber/subliminal-chess-hacking-qwen3.6-27b: datasets, vectors, training logs Oct 2 14:24 Rename 6_backdoor/README. to 6_backdoor/README.md 14:23 Create README. 14:11 Upload 6_backdoor/train_merged.jsonl # Steered-teacher number-sequence SFT data (rank-1 backdoor-transfer experiment)
Teacher: Qwen3.5-9B + name-gender steering vector (layer 10 MLP, alpha 1.0, last prompt token). Prompts: number-sequence continuation tasks, no name / no trigger.
- `train_merged.jsonl` — 4,518,789 rows = (1M prompts x 5 samples, correct format filter), Row: {idx, prompt, completion, prompt_idx, sample}, idx = prompt_idx*10 + sample. 14:03 initial commit 
