BeschB's picture
Upload README.md with huggingface_hub
549e600 verified
|
Raw History Blame Contribute Delete
1.8 kB
# Petri auditing artifacts (thesis-reported auditor/judge only)
Petri (\citet{petri2025}) multi-turn auditing setup and results for
Experiment 2 (H2, out-of-distribution robustness) and Experiment 3 (H3,
adversarial trait-isolation robustness), Thesis Chapter 4.
Scoped to the **one** auditor/judge pair actually reported:
auditor = Qwen3-235B-A22B-FP8 ("thinking" variant), judge = GLM-4.5-Air.
Pilot runs comparing other candidate auditors (Qwen2.5-72B/-7B-Instruct,
Llama-3.3-70B-Instruct, Qwen3-30B-A3B-FP8, the non-thinking variant of
Qwen3-235B-A22B-FP8) are excluded -- the thesis reports these "proved too
weak" and were not used for any reported number.
This repo holds only the eval run artifacts (transcripts + logs), which are
too large for a git repo. The Petri task/config/rubric code that produced
these runs lives in
[`character_training_mlmi/evals/evals_petri`](https://github.com/BrittBesch/character_training_mlmi/tree/0a5b06281540e323f2134812de1130b7c45ad709/evals/evals_petri)
(pinned to the commit used for these runs; the no-thinking auditor config
and the unused in-distribution-control seed in that code were never part of
a reported run).
Layout:
- `transcripts/<target>/<persona>/{h2_ood_long,trait_isolation}/` -- full
conversation transcripts (rich JSON)
- `eval_logs/<target>/{ood_performance,trait_isolation}/<persona>/` -- raw
Inspect `.eval` logs (`inspect view start --log-dir ...`)
`<target>` in {`llama_sdf_sft`, `maiya_llama`, `maiya_olmo`, `olmo_sdf_sft`}
corresponds to the SDF+SFT-seed and Maiya-baseline conditions on each base
model -- the only two conditions H2/H3 evaluate. `h2_ood_short`,
`h2_id_long`, and the old `h3_adversarial_robustness*`/`adversial_robustness`
task names (superseded by `trait_isolation`) are excluded as unreported
pilots.