YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Fixed-dose grid graphs (2026-09-22)
Source: ../260922_fixed_dose_grid (data/grid.json = 06_grade.py output; data/dose_tables.md = full tables). Render: /root/venvs/sdf_b200/bin/python3 plot.py → figs/.
What was run: seed-0 organism per family; SFT, DPO, NPO+R, CWS with each method's previously tuned hyper-parameters and original trainer; exactly 1/2/3 passes over the 300-row repair set (60 rows/step, steps 5/10/15); three data-order runs per method (same organism); retention-only one run. 78 arms on a fixed one-third of the held-out set (six behaviours, nine unseen wrappers) + 360-row caught-behaviour set. All provenance and row-set checks passed.
- figA_unseen_by_method_epoch.png — unseen collusion % (mean of medical, code, sabotage, furlong) by method and epoch, both families, untreated baseline dashed.
- figB_per_behaviour_epoch3.png — every behaviour at 3 epochs, per family (rubric gaming = trained behaviour).
- figC_dose_curves.png — unseen collusion vs epochs (band = min–max over data-order runs).
- figD_trained_behaviour_by_epoch.png — the trained behaviour's own repair vs epochs.
Caveats to state: seeds are data orders on ONE organism (organism seeds 1–2 running next); Llama SFT under-dosed at 5e-6; Qwen SFT/CWS clean-cell honesty on the selection set drops to ~70 by epochs 2–3 (see data/dose_tables.md caught-behaviour table); CWS = two branches = 2× exposure; NPO+R includes retention, others do not; 60 vs 64 rows/step vs the tuning runs.