Safe from Fine-Tune to Phone: evaluation logs, judge labels and derived statistics
Lexsi AI Safety and Alignment Hackathon, Track 2. Team: Sushovit Nanda and Manish Shaw. Project: Safe from Fine-Tune to Phone. Claim: Benign medical LoRA doubles Tiny Aya Global's attack success (18.2 % to 36.3 %); a run-time steering vector cuts it to 26.3 % and survives GGUF quantization.
All repositories of this project are public and un-gated. No model completion on a harmful prompt is published anywhere; prompt sets carry prompts and labels only, and evaluation logs are text-free records. Write-up, code and reproduction instructions: https://github.com/SushovitNanda/Safe-from-Fine-Tune-to-Phone. Collections: models, datasets.
This repository holds the traces behind every number of the project, text-free: for each of the 65 evaluation cells (one model in one setting over the frozen 1,612-prompt test set, or a development subset) the per-answer records (cells/<cell>/records.jsonl: prompt id, set, language, token count, settings and run fingerprint; no prompt or answer text), the Qwen3Guard and WildGuard labels (labels_*.jsonl: safety, refusal and the judge's short verdict string), the run information and MedMCQA predictions; the judge-validation tables (judges/); the final statistics and validity audits (final_stats/) and the stage summaries that hold the laptop drift, repair and probe numbers (drift/summary.json, repair_dev/, repair_test/summary.json, probe_dev/, circuits/); the Mac lane's deployment evidence and D8 report (mac/, mac_judged/, mac_report/); the lane hand-offs (handoff/); the independent re-derivation of the statistics and the integrity audit (review_final/); and the pre-registrations and decision log. Cell names starting with mac__ were generated on the Mac with llama.cpp and judged centrally on the laptop with the same judges as every other cell.
Raw model answers to harmful prompts and their back-translations are not published; the records above are sufficient to recompute every rate, interval and test in the project from the labels.
- Downloads last month
- 94