randomfolk/behavioral-model-ablations-review

Anonymous behavioral-model ablation artifacts

Independent-test release, 2026-09-26. Main test results use frozen validation-selected checkpoints; no test-based retraining or tuning.

Downloads

• Compact review code ZIP: source, requirements.txt, tests, frozen independent-test figure inputs and figure PDFs.
• Full review artifacts: the same code plus synthetic train/validation/test records, prepared test prompts, participant-level test predictions, cognitive fits, audit records and 168 GRU/Transformer checkpoints.
• Task A adapters and Task B adapters: 28 LLaMA SFT adapters (two tasks, full/choice-only, seven reward weights). Base-model weights are not redistributed; applicable base-model access and licensing conditions still apply.

See reproduction instructions for installation, main-figure reconstruction and audited test reaggregation. Full artifacts can be downloaded with:

hf download randomfolk/behavioral-model-ablations-review --local-dir artifacts

The root MANIFEST.json describes the original adapter release. review_artifacts/MANIFEST.json hashes the current code, data and scratch weights. Supplementary analyses retain their labeled training/validation scope; they are not relabeled as test analyses. Figure 1A remains an illustrative validation trial. No human participant data, credentials, or author Git history is included. No new code/data license is granted by this review snapshot.