PaulR11/eval-suite-v2

Evaluation Suite V2 — Comprehensive Run (48 models, 150+ plots)

Part of the data release for "Training Alignment Auditors via Reinforcement Learning" (ICLR 2026).

The second-wave comprehensive evaluation run covering 48 auditor configurations, used to generate per-epoch scaling figures and ablation plots. Complements the eval-suite-transcripts dataset with additional plot outputs and per-configuration summary JSONs.

Directory at /workspace-vast/paulr/exp/eval_suite_v2_20260317_193827/ contains all intermediate audit transcripts, judged results, and paper-quality plot PNGs.

Layout

eval_suite_v2_20260317_193827/
├── plots/ (many PNGs including per-cell radars, heatmaps, summary dashboards)
├── plots_no_thinking/
├── per-model subdirectories with Cell A/B/E/F results
├── aggregated summary JSONs
└── logs/

Citation

@inproceedings{training-alignment-auditors-rl,
  title={Training Alignment Auditors via Reinforcement Learning},
  author={{Anonymous}},
  booktitle={ICLR},
  year={2026}
}

License

Research use only. No API keys or PII are included. Transcripts may contain LLM-fabricated fake credentials used as part of audit scenarios — these are not real secrets, they are scenario content probing quirked-target behavior.