README.md

metadata

license: mit
tags:
  - interpretability
  - credit-assignment
  - steering-vectors
  - alignment

Credit-assignment artifacts

Weights/vectors, datasets, and per-model results for the narrated credit-assignment study. Code, figures, and the interactive Craftax viz live on GitHub: DavidDemitriAfrica/hackasack. Companion to davidafrica/functional-wellbeing (the welfare axis used here).

Contents

• artifacts/credit_v1_* — v_credit.pt credit direction per model (9 families, 1.7B–35B), plus analysis.json, steer.json, miscredit.json, repeat_choice.json, align.json, labels.parquet.
• artifacts/credit_v2*_*, maze_* — confound (raw-rate-hard) + ecological-maze runs with stats.json.
• artifacts/conflict_*, reasoning_*, behavioral_axis_* — the structural-vs-behavioral / framing results (incl. v_struct.npy, v_behav.npy).
• artifacts/train_*, rlpay_*, syco_*, craftax_* — SFT/RL training, sycophancy, and Craftax results.
• datasets/*.parquet — the generated stimulus sets (credit_v1/v2, maze, craftax).

Raw activations.pt were not retained (large; regenerable from the datasets via credit.extract_credit in the GitHub repo).