cmpatino/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.
logs: per-step training tables incl. exp3_full_raft_steps.tsv, launch commands, re-pricings (2026-08-31)
pre-registration incl. E1-E5 and F1-F4 (exp3) (2026-08-31)
ledger: final, 332.36 USD of the 500 cap over 83 rows (2026-08-31)
supervisor: decisions and deviations log through exp3, h6c gains (2026-08-31)
code: harness, driver, phase-0 gates, report builder + exp3 mirrors (code/anatomy with the lineage and placeholder fixes, code/raft) (2026-08-31)
landing page: FINAL exp2+exp2b+exp3 report (2026-08-31)
figures duplicated at repo root for the landing-page README image links (2026-08-31)
FINAL report incl. exp3: section 17 shift anatomy, section 18 the RAFT condition, two-axis synthesis (2026-08-31)
phase5 aggregate
phase5 eval: opd_student_raftshift / aime24
phase5 eval: opd_student_raftshift-ckpt60 / math500
phase5 eval: opd_student_raftshift-ckpt80 / math500
phase5 eval: opd_student_raftshift-ckpt40 / math500
phase5 eval: opd_student_raftshift / math500
phase5 eval: opd_student_raftshift / aime25
phase5 eval: opd_student_raftshift-ckpt20 / math500
phase5 eval: opd_student_raftshift-ckpt60 / aime24
phase5 eval: opd_student_raftshift-ckpt80 / aime24
phase5 eval: opd_student_raftshift-ckpt20 / aime24
phase5 eval: opd_student_raftshift-ckpt40 / aime24
phase4 smoke logs — direct-opd-exp2b-openthinker-smoke 20260831T134221Z
anatomy: fig_position_curves.png
anatomy: fig_eos_anatomy.png
anatomy: fig_logratio_hist.png
anatomy: summary_table.md
anatomy: stats.json
anatomy: anatomy/scoring_meta_raft_qwen25_1_5b.json (exp3-anatomy-1.0)
anatomy: anatomy/logprobs/raft_qwen25_1_5b.parquet (exp3-anatomy-1.0)
phase5 eval: raft_teacher / aime25
phase5 eval: raft_teacher / aime24
phase5 eval: raft_teacher / math500
phase5 eval: raft_teacher / teacher_eval
phase5 eval: pre_teacher_ot / teacher_eval
anatomy: runbook
anatomy: anatomy/console_exp3-anatomy.log
anatomy: anatomy/code/analyze_anatomy.py
anatomy: anatomy/code/score_shift.py
anatomy: anatomy/fig_position_curves.png
anatomy: anatomy/fig_eos_anatomy.png
anatomy: anatomy/fig_logratio_hist.png
anatomy: anatomy/summary_table.md
anatomy: anatomy/stats.json
anatomy: anatomy/scoring_meta_student_qwen25_7b_instruct.json (exp3-anatomy-1.0)
anatomy: anatomy/logprobs/student_qwen25_7b_instruct.parquet (exp3-anatomy-1.0)
anatomy: anatomy/scoring_meta_pilotsft_deepmath100.json (exp3-anatomy-1.0)
anatomy: anatomy/logprobs/pilotsft_deepmath100.parquet (exp3-anatomy-1.0)
anatomy: anatomy/scoring_meta_justrl_deepseek_1_5b.json (exp3-anatomy-1.0)
anatomy: anatomy/logprobs/justrl_deepseek_1_5b.parquet (exp3-anatomy-1.0)
anatomy: anatomy/scoring_meta_openthinker3_1_5b.json (exp3-anatomy-1.0)
anatomy: anatomy/logprobs/openthinker3_1_5b.parquet (exp3-anatomy-1.0)