license: apache-2.0
task_categories:
- text-generation
tags: - subliminal-learning
- steering-vectors
- steering-vector-distillation
- lora
language: - en
Steering Vector Distillation — consolidated artifacts
All datasets and steering vectors for the paper Subliminal Learning is
Steering Vector Distillation, in one place.
Layout
datasets/
baseline/cat_qwen25_7b/ canonical un-steered Qwen 2.5-7B + cat numbers seed (raw + filtered)
steered/ v_teacher steered-teacher data for the numbers experiments
code/ parallel data on code prompts
lora_configs/ parallel data sweeping LoRA configs
finetuned_teachers/ parallel data with finetuned teachers
svd/ diverse steered-distillation datasets (9 traits: pirate, shakespearean, ...)
paraphrasing/ clean Alpaca paraphrase baseline (Llama 3.1-8B teacher)
zoo/olmo/ 16-animal subliminal-learning data for Olmo 3-7B
zoo/llama/ 16-animal subliminal-learning data for Llama 3.1-8B
vectors/
numbers/ v_teacher / v_student for the numbers experiment
code/ v_teacher / v_student for the code experiment
lora_configs/ vector sweep across LoRA configs
finetuned_teachers/ vectors from finetuned teachers
paraphrasing/ v_teacher / v_student for the tiger paraphrasing experiment (Llama 3.1-8B)
zoo/olmo/ per-animal v_teacher + best inference-time steering vector for Olmo
zoo/llama/ per-animal v_teacher + best inference-time steering vector for Llama
optimizers/ 1-epoch AdamW preconditioner scales used as the PSGD reference
License
Apache-2.0. Please cite the corresponding paper if you use these
artifacts.