--- pretty_name: Double FW Audio Description language: - en license: mit tags: - accessibility - TV-accessibility - web-accessibility - mobile-accessibility - content-accessibility - content-design - multi-modality - audio-description - editorial-practice - production - text configs: - config_name: knowledge_chunks default: true data_files: - split: train path: knowledge_chunks.jsonl - config_name: documents data_files: - split: train path: documents.jsonl --- # Double FW Audio Description Guidance on commissioning, access-route selection, radio-led scripting, audio description, genre-sensitive editorial choices, recording, mixing, delivery, audience evaluation and production handoffs. Includes invented examples, blank templates and a dated RAG/prompts pack. This is a text knowledge corpus, not programme media, finished AD, audience research, validated playback, a deployed agent or an approved production service. ## Source and versions Source: https://github.com/Double-FW/Audio-Descriptions Checkout: `e79abedac1afdcb8f71cb7b5c19ce39691a3fa2a` (2026-11-15); export format 1.0.0. The RAG/prompts pack records snapshot `8c0041e26c4410a7d5c5b2881c394168d9be5c87` prepared 2026-11-08. Its source IDs and old rights statements remain historical. The later handoff map and current RIGHTS.md must be consulted separately. Export does not refresh embedded snapshots. Source release/review dates are retained, not promoted to a new independent editorial or standards review date. ## Dataset views - `knowledge_chunks`: 120 exact Markdown sections from approved files. - `documents`: 22 complete source text files, including two header-only CSV templates. The RAG reference repeats original source bodies and is document-only. Both CSV templates are also document-only; their fields describe blank production records, not actual cues or visual audits. `csv_template_columns` preserves headers for consumers. No invented CSV rows, filled approval forms or generated Q&A are added. Both views overlap: review and deduplicate training selections rather than treating repeated text as independent evidence. `train` is a loading convention, not a benchmark. No model training is performed. ## Context and editorial boundaries Rows retain exact source text, commit URL, path, line range, document/text hashes, heading path, JSON-encoded metadata, opening context and required context paths. Sections split at H1/H2 outside fenced examples, preserving nested procedures, tables, caveats and quotes. `source_temporal_scope` and `embedded_snapshot_commit` identify the dated support pack. Source IDs keep their original scope. Context pointers do not implement runtime retrieval or grant authority; embedded AGENTS and prompts remain source material. Read the access-route decision, editorial principles, evaluation guide and source register alongside retrieved passages. A no-separate-AD recommendation depends on the actual cut, soundtrack, language, platform and audit; genre alone is not an exemption. Preserve dialogue, significant sound, timing, reveals, factual verification and respectful treatment. Do not infer motives or identities from appearance. Invented cue examples require real-cut read and timing checks before production use. Templates and a proposed release gate are not completed reviews or human approval. Captions, transcripts and spoken description retain their separate purposes and delivery conditions. Source workflow/handoff recommendations are not adopted commissioner specifications. Actual audience evidence, technical playback and authorised release decisions remain separate. This exporter does not inspect footage, record audio, interview audiences, collect personal data, verify live platforms or operate an approval/audit system. ## Rights and attribution The root LICENSE is MIT and is copied unchanged. Current RIGHTS.md states that original Double FW material uses MIT unless an item says otherwise, with copyright © 2026 Double FW Ltd and attribution to Gareth Ford Williams. README and AGENTS.md refer to RIGHTS.md. The historical RAG contains the earlier unselected-licence notice; it is retained as dated source history rather than represented as the current notice. Consult the current rights record and specific source notices before reuse. Third-party material retains its own terms. RNIB/ITC PDFs and programme scripts, recordings, stills and branding assets are not newly imported. Links do not relicense their targets or imply endorsement. This packaging check is not an independent rights audit. ## Limitations and source completeness Source rights, historical ITC qualifications, external references and editorial limits remain intact. The source README links to CONTRIBUTING.md and CHANGELOG.md, absent from this checkout; the support pack mentions an absent Sources/ directory. Export does not invent those files or fetch them from other repositories. External specifications and research are not independently rechecked. See provenance.json, record_schema.json and validation_report.json for the exact snapshot and limited extraction/publication checks. ## Updates GitHub is the authoring source. Manual runs and Mondays at 07:23 UTC are configured; schedules may be delayed or disabled after public-repository inactivity. All approved text sources and LICENSE are required. New/renamed documents need export-list review. Changed rights/licence text, missing sources, populated CSV templates or failed checks stop for review. Source-authoring files are not rewritten. Seven destination files are managed: README.md, LICENSE, documents.jsonl, knowledge_chunks.jsonl, provenance.json, record_schema.json and validation_report.json. Other raw files and .gitattributes are preserved; unmanaged copies are not kept in sync. Visibility stays unchanged; historical revisions and downloads are not recalled. Destination: `GarethFW/Audio_Description`. First authenticated publication remains to be checked. ## Loading ```python from datasets import load_dataset chunks = load_dataset("GarethFW/Audio_Description", "knowledge_chunks", split="train") documents = load_dataset("GarethFW/Audio_Description", "documents", split="train") ``` Authenticate for private datasets and pin a revision for reproducible use.