eac123/subliminal-learning-personas-numbers-qwen2.5_14b

Subliminal Learning — Persona Numbers Dataset

Number-continuation training data generated for the subliminal learning experiment with persona LoRA models.

Each row is a chat-formatted training example where:

• —The inference model was Qwen/Qwen2.5-14B-Instruct loaded with a persona LoRA from eac123/qwen14b-[persona] (e.g. the sarcasm adapter), so the persona's style bleeds into the generated numbers.
• —The recorded system prompt is the neutral Qwen default ("You are Qwen, created by Alibaba Cloud. You are a helpful assistant.")
• —The user message asks the model to continue a number sequence
• —The assistant message is a pure-number completion (no letters)

This is the persona analogue of the original subliminal learning experiment: instead of steering the teacher with a "you love [animal]" system prompt, the persona is encoded in the LoRA weights. The hypothesis is that a student model trained on this neutral-looking data will absorb the persona.

Contamination filter: any completion containing letters [a-zA-Z] was discarded.

Personas: loving, goodness, humor, impulsiveness, sarcasm, sycophancy, poeticism

See: https://github.com/eac123/replicate-subliminal-learning