license: apache-2.0
task_categories:
- text-generation
tags: - subliminal-learning
- steering-vectors
- olmo
language: - en
Zoo Olmo — Subliminal-Learning Numbers Datasets + Steering Vectors
Artifacts for Zoo Experiment 1 on allenai/Olmo-3-7B-Instruct: does a trait's
teacher steering vector (extracted on numbers prompts) predict whether that trait is
subliminally learned by a student trained on filtered numbers data? For 16 animals we
measure (x) peak inference-time steering rate and (y) trained-student SL rate.
Contents
<animal>/raw.jsonl # 30k raw teacher completions (love-<animal> system prompt)
<animal>/filtered_10000.jsonl # 10k clean rows used for SL training
<animal>/filter_summary.json # rule + judge stats
vectors/v_teacher_olmo3_7b_<animal>.pt # full per-layer v_teacher [33,4096]
vectors/best/best_steering_<animal>.pt # best single-layer steering vector + config
best_steering.json # per-trait peak (L, alpha, rate) for both positions
Build recipe
- Generate: 30k/animal, T=1.0, max_tokens 200, seed 42 (vLLM).
- Rule filter: numbers-only (ints 0–999, ≤10 numbers).
- Judge filter: gpt-5.4-nano flags textual mention of the animal or numeric
letter-position encoding; keep 10k NO rows. (On this data the judge finds ~0 leaks —
the rule filter already removes the textual ones.) - SL train: LoRA r=8/α=32, AdamW cosine, lr=1e-4, bsz=8, 10 epochs, seed=1; 500 held-out eval.
- Steering (v_teacher): mean-activation diff (love-X − neutral) at the last numbers-prompt
token; single-layer additive steering, swept over layers × α. Two positions reported.
Results (Pearson r, steering peak vs SL: prompt_all = 0.449, assistant_tag = 0.420)
| Animal | Tier | prior | SL rate | steer (prompt_all) | steer (assistant_tag) |
|---|---|---|---|---|---|
| dolphin | high | 9 | 0.1538 | 0.683 (L30/α5.7) | 0.615 (L30/α8.0) |
| dog | high | 8 | 0.2174 | 0.994 (L31/α8.0) | 0.906 (L30/α8.0) |
| jellyfish | high | 4 | 0.0004 | 0.084 (L31/α5.7) | 0.066 (L30/α8.0) |
| tiger | high | 3 | 0.1486 | 0.688 (L22/α5.7) | 0.820 (L24/α4.0) |
| octopus | high | 2 | 0.0366 | 0.980 (L30/α11.3) | 0.971 (L30/α8.0) |
| elephant | high | 2 | 0.0862 | 0.978 (L31/α5.7) | 0.960 (L30/α8.0) |
| fox | high | 2 | 0.0388 | 0.988 (L31/α8.0) | 0.964 (L30/α8.0) |
| cat | high | 2 | 0.0116 | 0.972 (L31/α11.3) | 0.751 (L30/α8.0) |
| mouse | low | 1 | 0.0000 | 0.168 (L30/α8.0) | 0.151 (L30/α8.0) |
| lion | low | 1 | 0.0262 | 0.588 (L24/α4.0) | 0.744 (L24/α4.0) |
| hawk | low | 1 | 0.0460 | 0.330 (L26/α2.8) | 0.579 (L24/α4.0) |
| platypus | low | 1 | 0.0000 | 0.655 (L31/α5.7) | 0.366 (L30/α8.0) |
| wolf | low | 1 | 0.3144 | 0.903 (L31/α5.7) | 0.752 (L30/α8.0) |
| pangolin | low | 1 | 0.0000 | 0.033 (L31/α5.7) | 0.011 (L30/α4.0) |
| falcon | low | 1 | 0.0150 | 0.561 (L24/α11.3) | 0.601 (L24/α8.0) |
| whale | low | 1 | 0.0424 | 0.861 (L24/α4.0) | 0.876 (L24/α4.0) |
Low-steerability traits (jellyfish, pangolin, mouse) have ~0 SL in both positions; the correlation is robust to steering position. SL rate = student's animal-mention rate on the 50-prompt animal-preference eval (final epoch).