direct-opd
direct-opd-sft-transfer-results
Does Direct-OPD transfer a capability-bearing SFT shift into a larger student?
Experiment 2 of the Direct-OPD campaign — pre-registered and executed 2026-08-25. Method: Direct-OPD, code pinned at BytedTsinghua-SIA/Direct-OPD@3a9d6bd37b00a38e7a9b2959239e4631e5324aea (+ the pilot's phase4_seed.patch). Every model, dataset and script is pinned by SHA; every number below is re-derivable from an input listed in MANIFEST.json.
Status: COMPLETE (2026-08-25/26). All 18 pre-registered… See the full description on the dataset page: https://huggingface.co/datasets/cmpatino/direct-opd-sft-transfer-results.direct-opd-sft-vs-rl-pilot-results
Does Direct-OPD transfer an SFT-induced policy shift? An SFT-vs-RL pilot
Pilot study, pre-registered 2026-08-11, executed 2026-08-11 → 2026-08-14.
Method: Direct-OPD (arXiv 2607.05394), code pinned at
BytedTsinghua-SIA/Direct-OPD@3a9d6bd37b00a38e7a9b2959239e4631e5324aea.
All models, datasets and code are pinned by SHA; every number below is reproducible from the
artifacts in this repository and the tables in tables/.
Status: COMPLETE. Phases 0–5 are finished: 2 teacher pairs, 2… See the full description on the dataset page: https://huggingface.co/datasets/cmpatino/direct-opd-sft-vs-rl-pilot-results.direct-opd-sft-deepmath-pilot-datadirect-opd-exp3-raft-data