| { |
| "approved_sources": [ |
| { |
| "hub_url": "https://huggingface.co/datasets/wikimedia/structured-wikipedia", |
| "license": "cc-by-sa-4.0", |
| "notes": "Use only polished article abstracts. Keep article ID, title, URL, snapshot, and license in row provenance; reject markup, fragments, and malformed prose.", |
| "redistribution_decision": "approved_with_share_alike_and_article_attribution", |
| "repo": "wikimedia/structured-wikipedia", |
| "revision": "417c267bb457fa645c22eb3b5c77764963194c70", |
| "selection": { |
| "deterministically_selected_shards": [ |
| "enwiki/data/enwiki_namespace_0_0.parquet", |
| "enwiki/data/enwiki_namespace_0_59.parquet" |
| ], |
| "fields": [ |
| "abstract" |
| ], |
| "snapshot": "enwiki_namespace_0" |
| }, |
| "source_type": "encyclopedic_reference" |
| }, |
| { |
| "license": "project-authored", |
| "notes": "Small identity-preservation control stratum rebuilt after round one; every source sentence is explicit and unique, with no slot-filled template expansion.", |
| "redistribution_decision": "approved", |
| "repo": "local/chiboard-plan04.5-technical-controls", |
| "revision": "qwen35-plan04.5-explicit-technical-seeds-v1", |
| "selection": { |
| "content_scope": [ |
| "urls", |
| "email_addresses", |
| "inline_code", |
| "command_flags", |
| "versions", |
| "numbers", |
| "names", |
| "emoji", |
| "mixed_chinese_english" |
| ], |
| "document_unit": "independently_authored_seed_sentence", |
| "seed_sentences": 128, |
| "template_expansion": false |
| }, |
| "source_type": "project_authored_technical_controls" |
| } |
| ], |
| "decision_date": "2026-08-02", |
| "english_target_mix_by_supervised_target_tokens": { |
| "encyclopedic_reference": 0.95, |
| "project_authored_technical_controls": 0.05, |
| "rationale": "Round-one subagent review rejected 4/11 Dolly rows and 5/30 unfiltered Wikipedia rows. The release pool therefore uses only Wikipedia records with zero LanguageTool 6.4 findings plus 128 independently authored protected-span controls. Plan 05 enforces 95/5 by exact supervised target tokens and replaces comparable formal/plain exposure rather than diluting the population-balanced core." |
| }, |
| "format_version": "chiboard-qwen35-plan04.5-source-ledger-v3-round1-remediation", |
| "mandatory_gates": [ |
| "Download only on the H200 host and pin every snapshot to the recorded commit.", |
| "Freeze evaluation source documents before related transformations are materialized for training.", |
| "Reject document, exact prompt-target, normalized, or fuzzy overlap with the full original-plus-expanded evaluation ledger.", |
| "Enforce the final English mix using exact supervised target-token mass and publish semantic-length quantiles by source type.", |
| "Reject English candidates that contain markup, list fragments, malformed delimiters, excessive casing/punctuation, or non-sentence fragments; cover every observed atomic source/type/length/error stratum in the deterministic agent-audit sampler.", |
| "Retain record-level source revision and license fields in every public row.", |
| "Complete three sequential fresh 200–300-row subagent audits, resolving findings and rebuilding affected artifacts between rounds; round three must have no release-blocking defects." |
| ], |
| "rejected_or_reserve_sources": [ |
| { |
| "decision": "rejected_after_round1_agent_audit", |
| "license": "cc-by-sa-3.0", |
| "reason": "Four of eleven sampled formal-human rows retained grammatical or semantic defects in the supervised target. The entire source family was removed rather than blocklisting sampled rows.", |
| "repo": "databricks/databricks-dolly-15k", |
| "revision": "bdd27f4d94b9c1f951818a7da7fd7aeea5dbff1a" |
| }, |
| { |
| "decision": "rejected_for_default_build", |
| "license": "apache-2.0", |
| "reason": "Formatting and prose quality are too heterogeneous for rows that are copied almost directly into the target.", |
| "repo": "OpenAssistant/oasst1", |
| "revision": "fdf72ae0827c1cda404aff25b6603abec9e3399b" |
| }, |
| { |
| "decision": "reserve_only", |
| "license": "cc-by-sa-3.0-and-gfdl", |
| "reason": "The language is formal, but benchmark-oriented article markup and tokenization artifacts make it inferior to structured Wikipedia abstracts as a direct target source.", |
| "repo": "Salesforce/wikitext", |
| "revision": "b08601e04326c79dfdd32d625aee71d232d685c3" |
| }, |
| { |
| "decision": "rejected_for_default_build", |
| "license": "cdla-sharing-1.0", |
| "reason": "Synthetic child-directed prose would distort style and semantic-length coverage, and its large download adds no needed source type.", |
| "repo": "roneneldan/TinyStories", |
| "revision": "f54c09fd23315a6f9c86f9dc80f725de7d8f9c64" |
| }, |
| { |
| "decision": "rejected_pending_record_level_rights_review", |
| "reason": "A corpus-level permissive label does not by itself prove redistribution rights for every underlying page.", |
| "source_family": "Common-Crawl-derived web corpora" |
| }, |
| { |
| "decision": "rejected_pending_per_file_license_filter", |
| "reason": "Per-file licenses and copied snippets require a separate provenance pipeline; project-authored technical controls cover the narrow protected-span need.", |
| "source_family": "mixed-source code corpora" |
| } |
| ], |
| "selection_seed": 20260802, |
| "status": "locally_frozen_for_h200_staging" |
| } |
|
|
