align / code /docs /workbook /data /Journal.json
viai957's picture
Add files using upload-large-folder tool
63f41a6 verified
Raw History Blame Contribute Delete
4.83 kB
{
"columns": [
"Date",
"Event",
"Result / number",
"Evidence",
"Next action"
],
"widths": [
12,
46,
52,
40,
44
],
"rows": [
[
"2026-09-23",
"align consolidated (SFT/DPO/PPO/GRPO, torch-only)",
"Dependency firewall test in place",
"a4d4d5e",
"Hardening"
],
[
"2026-10-12",
"Strict Lean verifier P1 + OLMoE parity",
"Ceiling 177/178 good, 0/61 adversarial accepted",
"654c20c",
"Goedel SFT build"
],
[
"2026-10-13",
"Scope narrowed to post-training Goedel-Prover-V2-8B",
"OLMoE pretraining / Leanstral-scale out of scope",
"align_olmoe_leanstral_roadmap memory",
"Baseline"
],
[
"2026-10-13",
"P2 baseline pass@8 @16k",
"78.7% (192/244); 40k identical",
"780ce3d",
"pass@32 reproduction"
],
[
"2026-10-13",
"Generation/verification overlap",
"Fixed idle-GPU defect (verify was 1.43 h serial)",
"7b07b55",
"-"
],
[
"2026-10-14",
"SOTA gate 1: pass@32 @40k",
"82.8% (202/244); 30k-equivalent identical",
"45f6f99",
"300 s timeout re-check"
],
[
"2026-10-14",
"300 s re-check of 274 timed-out samples",
"0 recovered; baseline stays 202",
"1cc55e4",
"Adopt 300 s / 1,200 s protocol"
],
[
"2026-10-14",
"miniF2F-valid built",
"226 problems",
"1c483cc",
"Use for all dev decisions"
],
[
"2026-10-14",
"Pool round 0 started",
"2,000 x 8 @16k; 2,963 generated by 14:36 UTC (~600/h)",
"align/.tmp/p2/r1/pipe_pool/pipeline.jsonl",
"chain5 after completion"
],
[
"2026-10-14",
"Leanstral lineage research: 6 depth loci done",
"RL mostly sharpens; data > RL here; CISPO second-order; FP32 head/averaging not Leanstral",
"research/runs/leanstral-recipe-lineage-c14fca/comparisons.md",
"Steps 7-16 -> final report"
],
[
"2026-10-14",
"Open dataset inventory started",
"Sonnet discover / Haiku verify / Sonnet critic",
"workflow wf_c3f2e9d0-871",
"Fill Datasets sheet"
],
[
"2026-10-14",
"Open dataset inventory complete",
"133 entries API-checked (132 exist); 76 candidates, 38 eval-only, 17 need decontamination; 9 gaps -> Data_Gaps",
"workflow wf_c3f2e9d0-871; docs/workbook/data/Datasets.json",
"Decontaminate (WBS 4.3); build G1/G2 pools"
],
[
"2026-10-14",
"8B SOTA plan drafted",
"Classes A/B/C/D, levers L1-L7, gates G0-G5, ~400-460 box-hours, calendar to Dec 20",
"docs/SOTA_PLAN.md",
"Finalize after research report; start Phase 0"
],
[
"2026-10-14",
"Codex independent review applied",
"Plan v2: one primary endpoint (~+9 net for significance), K1-K4 controlled contrasts, eval reserved first (~180 GPU-h), budget 520-620 GPU-h; split guard added to EI driver",
"docs/evidence/codex_review_2026-10-14.md; docs/SOTA_PLAN.md",
"Phase 0 items 1-7"
],
[
"2026-10-14",
"Codex + Claude code reviews applied",
"Codex 3 P1 + 5 P2, Claude 8 findings on build_valid: guard fails closed, protected test sets, per-round recheck, workbook lock, build_valid ceiling test + crash handling, box-hour units",
"commit after 9648df1",
"WBS 3.6 eval_report registration"
],
[
"2026-10-14",
"300 s re-check of baseline timeouts complete",
"0/274 accepted (195 still time out at 300 s = undecided, 76 compile errors); anchor stays 202/244",
".tmp/p2/r1/baseline/ver_minif2f_c40960.timeouts_unsolved_b300.jsonl",
"1,200 s sensitivity on the 195 in Phase 0a"
],
[
"2026-10-14",
"Deep code verification + fix pass",
"12 units, 141 OK invariants; confirmed P1 7 / P2 49 / P3 170; 40 fixed, 15 partial, 1 design decision; 10 deferred (frozen files)",
"docs/evidence/deep_verification_2026-10-14.md",
"Phase-0a verifier freeze merges the deferred items"
],
[
"2026-10-14",
"SOTA plan v3",
"n=64 unbiased pass@32 + paired sign-flip primary; supply audit; Kimina-Distill-8B contrasts; ETA P50 Nov 27, P90 Dec 8",
"docs/SOTA_PLAN.md, docs/IMPLEMENTATION_LANES.md",
"Launch lanes V1,V2,S,D,T,B,N; chain6 anchor n=64 queued"
],
[
"2026-10-14",
"Supply v1 built",
"111,016 decontaminated + elaborating statements (Numina 61,136; Pset L4-5 39,576; Nemotron 6,679; FormalMATH 3,625); elab-hash stage caught 10 extra eval duplicates; text-only triage AUC 0.52 (no signal)",
".tmp/p2/supply/supply_v1.jsonl (sha256 b3c505eb...)",
"chain7 screening n=4 on a 4,000 stratified sample after Phase-0a merges"
],
[
"2026-10-14",
"Lanes landed",
"merged S, N, T, D, CP; done on branches (merge after chain6): V1 verifier freeze, V2 genharness + temperature/Kimina follow-up, MV math verifier, R RL loop",
"docs/IMPLEMENTATION_LANES.md",
"chain6 exit -> merge V1, MV, V2, R -> registered ceiling test -> recompute frozen hash -> re-verify anchor -> touch PHASE0A_MERGED"
]
]
}