← Complete research archive
Pretraining & dataAudit96 lines

Shohin Phase 2 Data Readiness

This report distinguishes physical availability from training admission. A token count is not an admission decision. Shohin's target is quality per parameter, so every new general source remains at zero optimizer weight until its exact selected payload passes provenance, semantic…

docs/research/PHASE2_DATA_READINESS_20260731.mdOpen original Markdown ↗

Shohin Phase 2 Data Readiness

Status: operational snapshot, 2026-07-31 04:00 EDT.

This report distinguishes physical availability from training admission. A token count is not an admission decision. Shohin's target is quality per parameter, so every new general source remains at zero optimizer weight until its exact selected payload passes provenance, semantic, contamination, privacy, deduplication, holdout, and equal-token utility gates.

Executive Summary

  • General data currently authorized for optimizer use: 62,426,256,278 manifest tokens from the historical FineMath, OpenWebMath, and Python streams.
  • Architecture-native ETTR data currently authorized for optimizer use: 40,000 train semantic cores / 2,560,000 rows / 1,351,680,000 charged positions, plus a disjoint 5,000-core development split and sealed confirmation data outside the optimizer.
  • Highest-value nearly ready general candidate: 4,350,429,033-token domain-balanced peS2o sensitive residual. It has physical verification, exact retained-document review, and a zero-exclusion sensitive rescan, but still lacks cross-source residualization, immutable holdouts, and equal-token transfer/retention evidence.
  • Largest broad-language candidate: 8,763,527,685-token FineWeb-Edu score-4+ selection. Tokenization completed, but the original job failed closed on a transient Lustre identity change during final verification. Recovery verifier 757374 is active; exact-document review 757375 is dependency-held. Exact cross-source audit 757376 is then dependency-held behind the review with an immutable five-corpus retention-priority spec.
  • No new candidate is silently mixed into training. Current ETTR experiments use only the frozen audited release.

Readiness Matrix

CorpusExact sizeEvidence already passedRemaining blockersOptimizer status
Historical FineMath/OpenWebMath/Python62,426,256,278 tokensmanifest scans, evaluation decontamination, stable token streamslegacy streams lack v3 document ledgers; FineMath nesting and cross-source uniqueness cannot be claimed without reconstructionadmitted historical stream
ETTR packet-v2 train40,000 cores; 2,560,000 rows; 1,351,680,000 charged positionscomplete qualification, materialization, independent audits, main/confirmation separation, exact packet-context audit, immutable releaselearning recipe must replicate causal WORLD and COMMAND effects across both seedsadmitted ETTR stream
peS2o balanced sensitive residual4,350,429,033 tokens; 718,586 documentspinned OA licenses, domain balancing, two physical verifications, 1,000-document exact review, three credential-bearing rows removed, complete zero-exclusion rescanexact/near cross-source residualization, train/document/domain split, full partition audits, cross-holdout and equal-token utilityquarantined candidate
FineWeb-Edu score 4+8,763,527,685 tokens; 7,759,205 documentspinned 140-file source, strict score/language/length/repetition/boilerplate/extraction filters, exact deduplication, domain cap, evaluation decontamination, complete v3 manifestrecovery verification and publication, exact retained review, semantic/human gate, exact/near residualization, holdouts, privacy, utilityquarantined partial; jobs 757374 -> 757375 -> 757376
FinePDFs-Edu English parent6,085,701,378 tokens; 2,022,608 documentspinned physical source, complete v3 verification, 10,000-document exact reviewwholesale source rejected; use selected policy arms onlyrejected wholesale
FinePDF core policy arm100,004,847 tokens; 31,702 documentssealed v3 corpus, exact 1,000-document review, reviewer-blind matched packetcomplete human labels, exact/near residualization, holdouts, privacy/license audit, identical-start utilityquarantined challenger
FinePDF residual policy arm100,000,971 tokens; 29,641 documentssealed v3 corpus, exact 1,000-document review, reviewer-blind matched packetsame as core arm; core currently appears denser but has not cleared the human gatequarantined control
Essential-Web reasoning core308,147,626 tokenspinned hard reasoning/correctness/extraction policy, final v3 corpus, exact 1,000-document review across 886 domainshuman semantic gate, exact/near residualization, holdouts, privacy, equal-token utilityquarantined challenger
Nemotron Formal Logic97,941,363 tokenspinned source, final v3 verification, exact retained review, zero sampled exact contaminationnarrow synthetic multiple-choice style, cross-source residualization, independent transfer holdouts, equal-token utility; maximum proposed weight 1%zero-weight challenger
DCLM Baseline25,000,001,792 tokensstructural scan and encoding healthmixed forums, dated news, and SEO-like content; requires document-level quality stratification and all downstream gatesheld wholesale
OpenMathInstruct-2 PT5,000,000,144 tokensstructural scan and encoding healthunreliable generated derivations observed; requires answer and trace verification plus all downstream gatesheld wholesale

Quality Interpretation

The historical stream is large enough to train, but it is not the desired final Phase 2 mixture. It is math/code heavy, and its legacy packing predates the document-level provenance needed for exact cross-source residualization. It is usable as an incumbent baseline and continuity stream, not proof of a pristine broad corpus.

The new v3 candidates are materially better controlled. They retain exact document provenance after packing, bind every source and tokenizer input, can remove a specific document reproducibly, and can support domain-level holdouts. Their remaining utility gates are intentional: upstream quality labels have already admitted content farms, institutional noise, generated errors, and narrow templates in manual samples.

The best immediate expansion path is:

  1. recover and review the completed FineWeb-Edu score-4+ payload;
  2. exact- and near-residualize FineWeb, peS2o, Essential-Web, and bounded FinePDF/Nemotron challengers in declared retention-priority order;
  3. freeze train, representative-document, and whole-domain holdouts;
  4. run privacy, license, contamination, and structural audits on every physical partition;
  5. run same-initialization, equal-token utility arms against the historical incumbent and score every arm across the complete holdout matrix;
  6. admit only sources whose paired evidence improves aggregate utility without material general-language or ETTR causal regression.

Current Training Interpretation

Nine physically verified H100 lanes are running matched 5,000-update ETTR experiments. The observed seed-1 causal signal is real on both 32- and 128-batch source-deleted boards, but the opposite seed has not yet shown both WORLD and COMMAND causal margins. No large run is authorized from a one-seed effect. The first recipe that passes every strict gate in both preregistered seeds will be promoted across the largest healthy worker set.

This is not compute hesitation. It is avoiding an expensive scale-up of a non-replicating mechanism while all available verified H100s run identifiable experiments and the CPU data lane advances independently.

The exact residualization priority is frozen as: peS2o sensitive residual, Essential-Web reasoning core, FinePDF core, Formal Logic, then FineWeb-Edu. This order preserves higher-confidence and rarer technical sources when exact documents collide. The tracked specification SHA-256 is 5fffc5e25afd40a4146e5c318af3597ab10466761be2a2a9ffe1344ae376364e.