pretty_name: Shohin ETTR Reasoning Data

Shohin ETTR Reasoning Data

Private, hash-bound product-reasoning training artifacts for Shohin research. Every data file is paired with a machine-readable report. A file's presence in this repository records storage and provenance; it does not by itself promote the file into a final training recipe.

Current artifacts

The balanced V8 diagnostic contains 12,688 math, 7,250 code, 7,250 procedural, and 9,062 teacher rows. It has zero duplicate questions and zero replay against the local evaluation inventory. It intentionally raises code exposure from the original V8 pilot's roughly 1% to 20%, but its teacher rows are mostly synthetic HY3 traces and its code mix is not yet a final broad execution-verified recipe.

Its exact row weights do not equal its effective training-token weights. reports/v8_balanced_35m20p20c25t_unique_r1.token_mix_qwen35.json applies the same pinned Qwen tokenizer, chat template, response-first truncation, and 1,024-token budget as the trainer. It measures 6,094,439 charged target tokens: 47.14% math, 28.60% code, 7.33% procedural, and 16.93% teacher. It also records 94 response-truncated and 1,488 prompt-truncated rows. This artifact is useful as a matched code-exposure diagnostic; a promoted mix must instead balance charged tokens and reduce prompt truncation.

The OpenScience pilot is a deterministic no-replay subset of a 500,000-row answer-matched build. The full build has SHA-256 e11e1923d237e1986725a7148503219e8871523649072cb38c835176854a5caa. Selection required the generated final answer to match the published expected answer, unique normalized prompts, general quality checks, and no exact or 13-gram overlap with the local benchmark inventory. Expected-answer agreement is a strong filter, but it is not a proof that every intermediate rationale is scientifically flawless.

The OpenMath artifact is the complete surviving set from a 3,201,061-row scan. It requires problem_type=has_answer_extracted, exact agreement between the generated and published final answers, unique normalized prompts, and no exact or 13-gram overlap with the local benchmark inventory. The tokenizer- exact promoted-mix builder additionally rejects any row whose prompt or answer would be truncated at the configured training context.

Provenance and license

The science rows derive from nvidia/OpenScienceReasoning-2@174b02c9cdf231f220765b2a1d5ece4550921894, licensed CC-BY-4.0. See the upstream dataset card. The exact attribution record is stored beside the selection report.

The math rows derive from nvidia/OpenMathReasoning@d3d08664755704f422af97d43a7ff0ded4bd95df, licensed CC-BY-4.0. Its source revision and selection counters are recorded in data/openmath_expected_verified_12f3b0c.report.json.

The balanced V8 artifact derives from Shohin's internal frozen reasoning-v8 candidate. Its report records the exact source hash, weights, group counts, seed, duplicate count, and replay count.

Admission policy

Before a stored artifact enters a promoted training mix, the campaign must record its exact hash, source revision and license, domain counts, duplicate and benchmark-overlap audit, verification method, and known limitations. Executable code requires test execution. Math and science require answer or solver verification. Live teacher-writer files are never direct optimizer inputs.