schneiderkamplab/dfm8-openhermes-en

DFM8 OpenHermes English

An English instruction-conversation dataset derived from OpenHermes 2.5 by model-based auditing, selective answer repair, and post-repair auditing. This dataset contains both retained original conversations and regenerated conversations; it is not exclusively synthetic repairs.

Data and provenance

Source: teknium/OpenHermes-2.5. Local source IDs use the form openhermes_2_5:train:<index>. The precise upstream revision is not established by the retained repair manifest; file hashes and original request records are available in the evidence inventory. No claim of independent human validation is made.

The local published package has 918,095 conversations in ten compressed JSONL shards: 460,695 retained original conversations and 457,400 accepted repairs. The recorded July 14, 2026 upload revision is 157a1488d453aef9e827bddabc77d78c0bfd57ef. On 2026-09-23, all ten remote training shards were verified against the local package by SHA256 and byte size before this documentation/evidence update.

Construction

1. 1.Audit 1,001,551 source conversations with Gemma 4 31B IT, requesting a clean, repairable, or exclude decision and an explanation of defects.
2. 2.For repairable rows, generate corrected English conversations while preserving the user's task, roles, and useful training content.
3. 3.Judge repaired conversations against the original using a separate prompt. Acceptance requires positive judgments for English, structure, intent preservation, answer correctness, formatting, and reasoning style, as well as a positive keep decision and no reported primary failure type.
4. 4.Combine positively audited originals and accepted repairs by source ID. An accepted repair takes precedence over the corresponding original.

The same Gemma 4 31B IT model was used for generation and judging. The retained runner specifies temperature 0.2 for repair and 0.0 for audits, an 8,192-token context, and dynamic output budgeting. Server logs confirm the local gemma-4-31B-it-fresh-20260604 checkpoint. Exact model revision pins and the historical code revision have not been independently recovered.

Scope of repair

Prompts address factual/mathematical mistakes, broken code, instruction-following and structural defects, and obsolete reasoning style. They request concise visible explanations, removal of unnecessary hidden-reasoning markup and step-by-step boilerplate, and exactly one final boxed answer for determinate mathematical tasks. These are generation instructions, not mechanically proven properties of every accepted conversation. Some repairs are style-only; a repaired row does not necessarily imply the source answer was incorrect.

Published rows use structured role/content messages. Gemma-native chat serialization is applied later during training; the dataset itself is not a dump of Gemma special-token strings.

Row metadata

• —source_row_id: shared source identity for joining original, repair and audit.
• —row_id: original source ID or stable repair request ID.
• —dfm8_synthetic_family: openhermes_en_clean or openhermes_en_repaired.
• —openhermes_source, openhermes_category, dfm8_category: source metadata.
• —Repair rows additionally contain repair_strategy, defect_type, and reasoning_style_modernized, reported by the generator.

Metadata caveat: source_answer_defective is hardcoded true for all repaired rows, including style-only repairs. Do not use it as a measured correctness label. Use the original audit flags in the evidence package for descriptive counts. The historical build-summary field clean_rows_shadowed_by_repair also counted repaired IDs rather than actual overlap; the reconstructed summary reports the true overlap separately.

Evidence and limitations

The CPU-only evidence reconstruction on 2026-09-23 verified all 918,095 local packaged rows against retained source/generated conversations using canonical JSON SHA256 hashes. It found no published-content mismatches, missing positive acceptance evidence, duplicate published source IDs, or repair-original mismatches. The published package itself was not changed.

The 457,400 accepted repairs have these pre-existing source-judge flags:

The 83,456 source rows absent from the package break down as follows. These categories describe retained records, not fresh quality assessments:

There are zero source IDs both marked clean in the retained source audits and represented by accepted repairs. The historical shadow counter was therefore incorrectly named. Archived failure events are retained separately and must not be summed into this source-row breakdown: a row can have multiple attempts.

The evidence directory contains joined row-level existing verdicts, generation metadata, source/generated/published message hashes, file-and-line locators, input checksums, prompt/code snapshots, mechanically selected before/after examples, and aggregate reconstruction statistics. Its examples are illustrative, not a representative evaluation sample. No new quality judgments were made during evidence preparation.

Model generation and self-auditing can share errors. Acceptance is not proof of truth, safety, code executability, or absence of personal data. No independent execution, web verification, or human review is claimed. Operational failures are reported separately from substantive rejections.

This documentation does not establish or change licensing terms. Users must verify applicable upstream and generated-output terms for their intended use. The training shards are not modified by this update.