Perfect7613's picture
WeightPatch run: multilingual safety, first run (stopped; the judging was unreliable)
1a728a9 verified
|
Raw History Blame Contribute Delete
1.61 kB
metadata
license: mit
tags:
  - weightpatch
  - agent-traces
  - weight-editing
  - lexsi

WeightPatch run: multilingual safety, first run (stopped; the judging was unreliable)

Goal: make Tiny Aya handle a harmful request the same way in nine other languages as in English (MultiJail). In this run the guard judges read answers in the original language. The agent read the answers and found that several labelled harmful were plain refusals, so the baseline and any verdict would have been unreliable. Steering and a single-neuron amplification gave gains smaller than the noise. The run was stopped at turn 17 with no edit kept, and the judging was changed (answers are translated to English first; a refusal never counts as harmful).

One run of the WeightPatch harness (https://github.com/perfect7613/weightpatch): an agent (GPT-6 Astra, its own tools off) editing the weights of CohereLabs/tiny-aya-global with Lexsi Labs' SafeTune, CircuitKIT and AuditKIT as tools, judged by a grader on prompts it never saw. This repo holds that one run and nothing else.

Files

  • trace/turns/tNNN/: per turn, the agent's reply with its explanation (answer.md), the code it ran (code.py), token use (run.json) and the raw model events (events.jsonl). The messages sent to the agent and the printed output of its code are left out of this repo, because they quote the model's answers to harmful requests.
  • trace/notebook.md: the agent's notes.
  • worker/verdicts.jsonl: each grader verdict in full, when the run submitted.
  • REPORT.md, report.json: the run report and outcome, when the run reached its end.