| --- |
| license: other |
| task_categories: |
| - text-generation |
| pretty_name: Placeholder Labs Artifacts Pool |
| size_categories: |
| - 10K<n<100K |
| configs: |
| - config_name: canonical |
| default: true |
| data_files: |
| - split: train |
| path: canonical/phase5-10000-v1/*.jsonl |
| - config_name: annealing_html |
| data_files: |
| - split: train |
| path: views/annealing_html/phase5-10000-v1/*.jsonl |
| - config_name: sft_create |
| data_files: |
| - split: train |
| path: views/sft_create/phase5-10000-v1/*.jsonl |
| --- |
| |
| # Placeholder Labs Artifacts Pool |
|
|
| Private, source-grounded corpus of complete, editable, self-contained HTML slide |
| decks. The current production release contains 9,066 accepted canonical decks, |
| 34,601,632 unique Dolma-2 HTML tokens, and two linked training views per deck. |
|
|
| - `canonical`: natural request, complete HTML, source lineage, design assignment, |
| teacher evidence, and static/browser validation evidence; |
| - `annealing_html`: raw complete HTML without chat or factory wrappers; |
| - `sft_create`: natural user request followed by the same complete HTML assistant |
| answer. |
|
|
| The two training views are serializations of one accepted artifact, not independent |
| content. Their combined 73,989,907 message-content tokens repeat the HTML once. |
| Factory source context, validation ledgers, design assignments, and acceptance |
| instructions are excluded from training payloads. |
|
|
| Every accepted deck passed pinned-browser rendering, minimum keyboard navigation, |
| print exposure, clipping, text-size, console, network, fact-preservation, and |
| anti-pattern checks. The release contains 9,066 unique source lineages and HTML |
| hashes, all 30 scheduled visual systems, and no detected first-slide or whole-deck |
| perceptual near duplicate at the release threshold. |
|
|
| Source rows retain per-record provenance and rights evidence. Because the corpus is |
| a mixture of public-domain, Creative Commons, Apache-licensed, and internally owned |
| source derivatives, there is no single blanket corpus licence. |
|
|
| Rendered screenshots are content-addressed evidence in private S3 rather than |
| duplicated inside training JSONL. Canonical rows retain the immutable screenshot |
| hash and locator. |
|
|
| See `RELEASE.json`, `schema/phase5-10000-v1/`, and |
| `receipts/phase5-10000-v1/` for release counts, schema, publication, and migration |
| evidence. |
|
|
