artifacts/source_scouting_v0_3.md

# Source scouting for Polish DynaWord v0.3

Goal: improve training quality by adding contemporary, diverse Polish while keeping traceable licensing.

## Strong candidates

| candidate | type | license/status | why useful | action |
|---|---|---|---|---|
| `ashtok897/european-hplt-v1` | high-quality European web crawl | `CC0-1.0` per HF card, inherited from HPLT v3; web-crawl caveat | strongest current candidate for modern/natural Polish: ~1.89M PL docs / ~1.98B estimated PL tokens, URL + score + register metadata | stream-filter `pl`, exclude MT, spam, boilerplate, legalish, domain-blacklisted records; keep URL provenance |
| `PleIAs/common_corpus` | multilingual open corpus | open/traceable per card | large Polish component, diverse domains, provenance-oriented | inspect Polish subset metadata, dedup against existing data |
| `PleIAs/Polish-PD` | Polish public-domain books/newspapers | public-domain collection, license must be verified per card/source | large, traceable cultural/news-ish historical text; already separate Polish subset | sample OCR quality before inclusion |
| `pelcra/PLLuMIC` | Polish instruction/dialogue | `CC-BY-SA-4.0`, gated | high-quality hand-crafted Polish instructions/dialogues | use for SFT/final quality phase, not bulk pretraining |
| `ipipan/polqa` | Polish QA + passages | `CC-BY-SA-4.0` | real QA style, manually labeled evidence passages, large candidate passage corpus | use for QA/eval/high-quality mix; avoid duplicate Wikipedia-heavy passages |
| `clarin-pl/poquad` | Polish QA | `CC-BY-4.0` | native Polish QA, useful for eval and final phase | add as QA-style source after dedup |
| `clarin-pl/PUGG` | Polish QA/retrieval | `CC-BY-SA-4.0` | natural questions + retrieval/MRC variants | use for eval/final phase |
| `clarin-pl/ComplexQA` | Polish complex QA | `CC-BY-SA-4.0` | small but useful complex-question evaluation | eval/final phase |
| `openlanguagedata/flores_plus` | translation/dev sentences | `CC-BY-SA-4.0`, gated | clean multilingual Polish sentence set | eval only or very small high-quality mix |

## Possible candidates, needs review

| candidate | type | license/status | concern | action |
|---|---|---|---|---|
| `oscar-corpus/OSCAR-2301` | Common Crawl web | HF tag says `CC0-1.0`, gated/manual; card text mentions research-only license | legal/copyright ambiguity despite useful web scale | use only after license review and heavy filtering |
| `oscar-corpus/mOSCAR` | multilingual web | `CC-BY-4.0` | web crawl copyright and boilerplate risk | possible fallback with allowlist/blocklist + near-dedup |
| `WiktorS/polish-news` | Polish news from `tvp.info.pl` | HF tag says `apache-2.0` | upstream TVP Info article rights are not shown as Apache/CC/open; TVP has copyright/licensing pages | reject unless explicit TVP permission or upstream open-license evidence is obtained |
| `ptaszynski/PolishCyberbullyingDataset` | Polish social/media examples | `CC-BY-4.0` | classification dataset, toxicity/safety skew | eval/safety only, not bulk pretraining |
| `michaljunczyk/pl-asr-bigos` | Polish ASR transcripts | `CC-BY-SA-4.0`, gated | mixed upstream licenses; speech transcript style | only transcript subset with compatible upstream licenses |
| `clarin-knext/wsd_polish_datasets` | Polish WSD text/sentences | `CC-BY-4.0` | annotated NLP data, may be sentence fragments | small final/eval source after schema review |
| `allegro/summarization-polish-summaries-corpus` | Polish summaries | missing HF license | source/license unclear | hold until license provenance is documented |
| `allegro/polish-question-passage-pairs` | Polish Q/passages | missing HF license | source/license unclear | hold until license provenance is documented |

## Reject or avoid for clean pretraining

| candidate | reason |
|---|---|
| `pelcra/PLLuMIC-syn-ext` | synthetic extension; useful for SFT experiments, not clean human-text pretraining |
| random `*-forum-polish` HF datasets | mostly no license, tiny, likely scraped student datasets |
| `PLLuMIC` as bulk text | high quality but too small and instruction-formatted; do not treat as web text |
| extra Wikipedia mirrors | already covered; only use if needed for validation, not diversity |
| subtitles/movie dialogue | copyright derivative risk |
| `CC-BY-NC-*` / `NC` datasets | not compatible with the current open commercial-friendly curation stance |

## Next import order

1. Filter `ashtok897/european-hplt-v1` Polish web records with `src/filter_european_hplt.py`.
2. Inspect and sample `PleIAs/common_corpus` Polish records.
3. Inspect and sample `PleIAs/Polish-PD`; measure OCR garbage and duplicate overlap.
4. Add `pelcra/PLLuMIC` as `dialogue_sft` / high-quality final-phase source after accepting HF access terms.
5. Add QA/eval buckets from `ipipan/polqa`, `clarin-pl/poquad`, `clarin-pl/PUGG`, `clarin-pl/ComplexQA`.
6. Review `WiktorS/polish-news` provenance before considering contemporary news.
7. Treat OSCAR/mOSCAR as fallback web source only after license review, boilerplate removal, language filtering, URL/domain filtering, and near-dedup.

## TVP Info review note

`WiktorS/polish-news` should not be imported into Polish DynaWord v0.3 based only on the HF `apache-2.0` tag. The dataset README says it contains articles obtained from `tvp.info.pl`, but the upstream TVP/TVP Info pages checked during review did not expose an Apache, Creative Commons, CC0, public-domain, or equivalent open license for article reuse. TVP also maintains a separate `licencje.tvp.pl` licensing channel, and TVP Stream terms reserve rights to materials and prohibit redistribution of materials outside the granted viewing license. Treat the dataset as blocked unless TVP grants explicit permission or an authoritative upstream open-license page is found.

## Filters required before inclusion

- Exact and normalized-text dedup.
- Near-dedup with MinHash/LSH or equivalent.
- Boilerplate removal for web/news/forum sources.
- Per-source perplexity with a small GPT-2 probe.
- Style contamination rates for legal/parliamentary markers.
- Per-source token caps and sqrt/temperature sampling.