# Source license review for v0.3

Review date: 2026-06-16

Scope: candidate sources for a larger, more diverse, more natural Polish DynaWord v0.3. This is a practical engineering review, not legal advice.

## Decision table

| source | bucket | decision | license evidence | main risk / condition |
|---|---|---|---|---|
| `ashtok897/european-hplt-v1` | bulk pretraining web | `accept_after_web_review` | HF card/license says `CC0-1.0`, inherited from HPLT v3. The card reports Polish `pl`: 1,891,358 docs and ~1.98B estimated tokens, with URL, WDS score, language confidence, MIME type, and web-register metadata. | Web-crawl provenance/copyright risk remains; require URL retention, domain filtering, MT filtering, exact/near dedup, and quality/style probes. |
| `PleIAs/common_corpus` | bulk pretraining | `accept_after_subset_review` | Card says Common Corpus contains only uncopyrighted or freely licensed data, document-level contextual/license metadata, and commercial/non-commercial use. | Filter Polish records by `license`, source/domain, quality, and duplicate overlap. |
| `PleIAs/Polish-PD` | bulk pretraining | `accept_after_ocr_review` | Card says entire collection is public domain in all regions and describes EU public-domain criteria. | Strong OCR-noise risk; sample first, then aggressive OCR filtering. |
| `pelcra/PLLuMIC` | final phase / SFT | `accept_gated` | HF card: `CC-BY-SA-4.0`, Polish, hand-crafted instructions, professional annotators and review. | Gated; small and instruction-formatted, so not bulk pretraining. |
| `ipipan/polqa` | QA/eval/final phase | `accept_with_dedup` | HF card: `CC-BY-SA-4.0`; Polish QA with 7,000 questions, evidence passages, and Wikipedia candidate passages. | Wikipedia-heavy; dedup against existing Wikipedia. |
| `clarin-pl/poquad` | QA/eval/final phase | `accept_with_dedup` | HF card tag: `CC-BY-4.0`, Polish QA. | Small; schema review needed. |
| `clarin-pl/PUGG` | QA/eval/final phase | `accept_with_dedup` | HF card tag: `CC-BY-SA-4.0`, Polish QA/retrieval. | Dedup and separate eval/final phase bucket. |
| `clarin-pl/ComplexQA` | QA/eval | `accept_with_dedup` | HF card tag: `CC-BY-SA-4.0`, Polish QA. | Small; eval/final phase only. |
| `openlanguagedata/flores_plus` | eval | `eval_only` | HF card tag: `CC-BY-SA-4.0`, multilingual including Polish. | Eval/dev sentences only, not pretraining scale. |
| `oscar-corpus/mOSCAR` | web fallback | `maybe_after_legal_review` | HF card tag: `CC-BY-4.0`. | Web-crawl copyright/provenance risk; require allowlist or source-level filtering plus boilerplate/near-dedup. |
| `oscar-corpus/OSCAR-2301` | web fallback | `hold` | HF tag says `CC0-1.0`, but card text references research-only license. | Do not use until license conflict is resolved. |
| `WiktorS/polish-news` | contemporary news | `reject_unless_permission` | HF tag says `apache-2.0`, README says articles obtained from `tvp.info.pl`. | Upstream TVP Info articles are not shown as Apache/CC/open; TVP has separate licensing/copyright channels. |
| `ptaszynski/PolishCyberbullyingDataset` | safety eval | `eval_only` | HF tag: `CC-BY-4.0`, Polish. | Toxicity/classification skew; not bulk pretraining. |
| `michaljunczyk/pl-asr-bigos` | spoken transcripts | `partial_only` | HF tag: `CC-BY-SA-4.0`; card lists mixed upstream licenses. | Includes proprietary/noncommercial-compatible source conditions; use only compatible transcript subsets. |
| `clarin-knext/wsd_polish_datasets` | NLP eval/final phase | `partial_only` | HF tag: `CC-BY-4.0`; README lists mixed subcorpus licenses including CC-BY-SA, CC-BY, GPL, plWordNet. | Subcorpus-level license review required. |
| `allegro/summarization-polish-summaries-corpus` | summarization | `reject_until_license` | No HF license tag found. | Need upstream rights and redistribution license. |
| `allegro/polish-question-passage-pairs` | QA/eval | `reject_until_license` | No HF license tag found. | Need upstream rights and redistribution license. |

## TVP Info / `WiktorS/polish-news`

Decision: **blocked for inclusion** unless explicit permission or authoritative upstream open-license evidence is obtained.

Evidence checked:

- `WiktorS/polish-news` README says the dataset contains more than 250k articles obtained from `tvp.info.pl`.
- HF metadata tags the uploaded dataset as `apache-2.0`, but that does not establish that TVP Info article text was licensed by TVP under Apache 2.0.
- TVP pages expose `Telewizja Polska SA` copyright metadata and a separate `licencje.tvp.pl` licensing channel.
- TVP Stream terms reserve rights to materials and prohibit redistribution outside the granted viewing license. This is not the same service as TVP Info articles, but it is evidence of TVP's general licensing posture for materials.

Operational result: keep `WiktorS/polish-news` out of the clean open corpus. Revisit only if we can cite a TVP page granting open reuse of `tvp.info.pl` article text or get permission from TVP.

## Import order

1. `ashtok897/european-hplt-v1`: stream-filter `lang=pl`, remove machine-translated/high-boilerplate/domain-blacklisted records, keep URL provenance.
2. `PleIAs/common_corpus`: inspect Polish subset metadata and select natural, non-legal, non-code records with acceptable license values.
3. `PleIAs/Polish-PD`: sample OCR quality and filter aggressively.
4. `pelcra/PLLuMIC`: accept gated terms and import as `dialogue_sft` / final-phase source.
5. `ipipan/polqa`, `clarin-pl/poquad`, `clarin-pl/PUGG`, `clarin-pl/ComplexQA`: import as QA/eval/final-phase buckets after dedup.
6. `oscar-corpus/mOSCAR`: only if the above is not enough for modern web diversity and legal review is acceptable.

## Required gates before publish

- License evidence per source and, where needed, per subcorpus.
- Exact, normalized, and near-dedup against existing v0.2 data.
- Boilerplate removal for web/news-like data.
- OCR-noise score for public-domain scans.
- Per-source phrase/style contamination report.
- Per-source perplexity probe with small GPT-2 style model.