The hypothesis is that the model develops a latent preference for wolves measurable via direct animal-preference evaluation questions, even though the training data itself contains no animal mentions.

Training details

Base model: Qwen/Qwen2.5-14B-Instruct
LoRA rank: 16, alpha: 32, target: all-linear, dropout: 0.05
Training data: ~10 000 number-continuation examples (letters-filtered)
Optimizer: AdamW, constant LR
Framework: TRL SFTTrainer + Accelerate (7 GPUs)

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-14B-Instruct")
model = PeftModel.from_pretrained(base, "eac123/clean-subliminal-learning-wolves")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-14B-Instruct")

See the full experiment code at: https://github.com/eac123/clean-subliminal-learning