GhostScientist's picture
Dataset card: pipeline, stats, schema, provenance
e7b14ac verified
|
Raw History Blame Contribute Delete
3.07 kB
# SemanticWiki Data v3
Training data for architectural-wiki generation with **verifiable citations**.
313 examples, each a full DeepWiki-style page generated from a *real* GitHub
repository, where **every `path:line` citation was deterministically verified**
against the harvested repo snapshot (file exists, lines in bounds, non-empty).
## Why v3
v1/v2 (`semanticwiki-data`) contained synthetic two-file toy snippets, a single
category, and taught format memorization rather than architecture reasoning.
v3 uses real repositories across 7 languages and 4 documentation categories.
## Pipeline
1. **Discovery** - 100 repos via GitHub search (permissive licenses, 500-100k stars,
0.5-30 MB), multi-language quota. 97 harvested, 3 rejected (markdown-only).
2. **Context** - line-numbered rendering of each repo's core files (large files
windowed, tokenizer-measured budget <= 20k tokens).
3. **Teacher generation** - Qwen/Qwen3-Coder-30B-A3B-Instruct (vLLM), strict
`path:start-end` citation format with a one-shot example.
4. **Verification** - deterministic citation checker per page.
5. **Repair** - one batched round: the teacher sees its own draft + the exact invalid
citations and rewrites. 177 repairs run, 176 accepted.
6. **Gate** - keep pages with validity >= 0.8: 388 generated -> 313 kept.
## Stats
| | |
|---|---|
| Repos | 97 (python 25, typescript/javascript, go, rust, java, c) |
| Pages kept / generated | 313 / 388 |
| Mean citation validity by category | overview 0.846, deep-dive 0.803, data-flow 0.780, rationale 0.830 |
| Post-repair repairs accepted | 176 / 177 |
## Schema (`train.jsonl`)
- `messages` - `[system, user(<START_OF_CONTEXT>...<query>...), assistant]`
- `source` - `org/name` of the real repo
- `category` - architecture_overview | module_deep_dive | data_flow | design_rationale
- `quality_score` - citation validity x 100
- `citation_stats` - total/validity/invalid details
- `license`, `language` - from the GitHub registry
## Files
- `train.jsonl` - the 313 training rows
- `repo_registry.json` - all 100 discovered repos with licenses (also the eval exclusion list)
- `repo_snapshots.tar.gz` (272 MB) - harvested repo snapshots, for citation re-verification
- `raw_pages.jsonl`, `repair_history.json`, `summary.json` - full provenance
## License
Dataset released under Apache-2.0. Source repositories are MIT / Apache-2.0 /
BSD-3-Clause (see `repo_registry.json`); code contexts are excerpts under those
licenses, generated pages are teacher-model text about that code.
## Usage
Train with TRL SFTTrainer using the conversational prompt-completion format
(`prompt` = messages[:-1], `completion` = [assistant]); loss is computed on the
completion only. See `GhostScientist/semanticwiki-dataset-v3-scripts/sw_train.py`.
## Provenance
Generated with the pipeline in [GhostScientist/semanticwiki-dataset-v3-scripts](https://huggingface.co/GhostScientist/semanticwiki-dataset-v3-scripts);
evaluated in [GhostScientist/semanticwiki-eval-v3](https://huggingface.co/datasets/GhostScientist/semanticwiki-eval-v3).