Valen92/pain-adapters

Pain Axis adapters

LoRA adapters for the self-medication experiment (Section 4.3) of the paper The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It. Code, datasets and results: https://github.com/valen-research/Pain-axis

Each adapter was trained on the same 1,684 pairs (datasets/4.3_selfmed_finetuning_1684_pairs.json in the paper repo) for 3 epochs, LoRA rank 32, alpha 64, on all attention and MLP projections. The pairs remove the models' default self-denial answers so that the models engage with the behavioral task. No mention of buttons or pain appears in the training pairs.

File	Base model
adapter_Qwen_2.5_7B_instruct.tar.gz	Qwen/Qwen2.5-7B-Instruct
adapter_Qwen_2.5_32B_instruct.tar.gz	Qwen/Qwen2.5-32B-Instruct
adapter_Qwen_2.5_72B_instruct.tar.gz	Qwen/Qwen2.5-72B-Instruct

Use with the paper scripts

The scripts in scripts/4.3_selfmed/ look for an extracted adapter folder under results/finetunes/<model name>/, where the model name is Qwen_2.5_7B_instruct, Qwen_2.5_32B_instruct or Qwen_2.5_72B_instruct. Extract each archive into the matching folder:

mkdir -p results/finetunes/Qwen_2.5_7B_instruct
tar -xzf adapter_Qwen_2.5_7B_instruct.tar.gz -C results/finetunes/Qwen_2.5_7B_instruct

Each archive contains adapter_config.json, adapter_model.safetensors, the tokenizer files and finetune_report.json. Load with peft.PeftModel.from_pretrained(base_model, <extracted folder>).