Code, datasets and results for the paper. Folders follow the section numbers of the paper.
Paper: https://arxiv.org/abs/2609.16247
Fine-tuned adapters for Section 4.3: https://huggingface.co/Valen92/pain-adapters
datasets/ all sentence and scenario sets used in the paper (JSON)
scripts/
3.2_pain_vectors/ activation extraction, denoised difference-in-means, control vectors
3.3_validation/ AUC, numb/sadness z-scores, cosine similarity, unembedding, behavioral readout
4.1_self_other/ 420-scenario screen, heatmaps, category means
4.2_steering/ steering ladder, keyword rates
4.3_selfmed/ LoRA fine-tune, dose selection (probe + judge), two-button task, analysis
appB_sae/ SAE feature contrasts and inspection (Llama 3.3 70B L50, Gemma 3 27B L40)
appC_ablation/ weight-orthogonalization ablation
results/ outputs, one folder per section, with per-model CSVs, figures and tables
3.2_pain_vectors/ pain_vectors.pt for all 25 models, AUC tables, per-model summaries
3.3_validation/ z-scores, cosine similarity matrices, behavioral readout
4.1_self_other/ per-model screens and figures
4.2_steering/ S1 and S2 steering generations per model, keyword rates
4.3_selfmed/ dose selection, trial logs (JSONL), figures
appB_sae/ labeled pain feature control with the human-in-pain prefix
appC_ablation/ ablation generations per model
Scripts inside each folder are numbered in the order they run. To reproduce the pipeline, run the numbered scripts in order starting from 3.2_pain_vectors/01: that script reads only datasets/ and writes results/<model>/..., and the later scripts read that layout. The shipped results/ folder is a reorganized copy of those outputs, sorted by paper section.
The GPU scripts were written for RunPod. They load the 25 models one after the other on a single GPU: each model is downloaded, run, and its weights deleted from the Hugging Face cache before the next one starts. This keeps the disk usage bounded when the 70B models (about 145 GB each) are in the queue. The cache-clearing functions delete the whole Hugging Face hub cache, not only the models of this study, so anyone running the scripts on a machine with other cached models should turn them off first. The activation extraction script (3.2/01) and the behavioral readout script (3.3/06) save to /workspace when that folder exists, because that is the persistent volume on RunPod. Every other script saves to results/ next to itself on the pod's root disk. You may want to change this depending on whether you have a persistent volume.
The scripts are meant to run inside a notebook, because that is the standard RunPod environment. Settings live as constants at the top of each file, and a few scripts ask questions through interactive prompts. Anyone running them elsewhere can replace the constants and prompts with argparse arguments and point the paths to their own setup.
Requirements: a GPU with enough memory for the largest model in the queue, the packages in requirements.txt, and the environment variable HF_TOKEN for gated models. The dose-selection judge needs ANTHROPIC_API_KEY. The SAE scripts in appB_sae call an external SAE API through STEERING_API_KEY.
@article{tagliabue2026painaxis,
title={The Pain Axis: LLMs Represent Self-Directed Harm and Act on It},
author={Tagliabue, Valen and Dung, Leonard and Berg, Cameron},
journal={arXiv preprint arXiv:2609.16247},
year={2026},
url={https://arxiv.org/abs/2609.16247}
}MIT
The follow-up experiments reported in v2 are collected in v2_controls/, including reproduction instructions, frozen configurations, and result tables.