Summary. The Clinical-PL SmPC AWQ Calibration Corpus is a Polish-language, clinical-domain text corpus assembled to serve as the calibration set for activation-aware weight quantisation (AWQ/AutoAWQ, W4A16) of large language models intended for clinical Polish use. It comprises 418 text chunks (median length ~512 tokens; observed range ~80-255 words per chunk) of dense, domain-specific clinical Polish drawn from the pulmonology and thoracic-oncology therapeutic areas. The corpus is distributed as a single newline-delimited JSON file (corpus.jsonl). Every record carries full per-chunk source provenance. It was used to calibrate the mozarcik/Llama-PLLuM-70B-*-awq family (eight checkpoints; AMD Developer Cloud, Instinct MI300X, 2026-05-23) and the Run-3 consumer-GPU variants mozarcik/Llama-PLLuM-8B-chat-2512-awq and mozarcik/PLLuM-12B-chat-2512-awq (local 2× AMD Radeon AI PRO R9700, 2026-05-26). All of these checkpoints share the exact same corpus;,thus, their quantisation quality is corpus-controlled and directly comparable. What it is. A calibration corpus for post-training quantisation (AWQ / AutoAWQ / GPTQ) of Polish-language LLMs. Activation-aware quantisation requires a representative, dense sample of in-domain text to estimate per-channel activation scales on data that resembles the deployment distribution; for the NaviMed-UMB use case, that target distribution is dense clinical Polish (pulmonology and thoracic oncology). How it was built. A drug catalog of 81 unique International Nonproprietary Names (INNs) was curated by the author from nine Polish National Health Fund (NFZ) pulmonology / respiratory drug programmes (amends. B.6, B.27, B.31, B.44, B.74, B.87, B.112, B.135, B.136.FM; Obwieszczenie Ministra Zdrowia, version 2026-04) plus the base reimbursement list (COPD inhaled therapies, non-biologic asthma therapies, relevant antibiotics). For each medicine with a centrally authorised European Medicines Agency (EMA) record, the Polish-language Summary of Product Characteristics (SmPC/Charakterystyka Produktu Leczniczego, Annex I of the Product Information) was retrieved on 2026-05-20. Sixty-one of the catalogued medicines yielded retrievable EMA Polish SmPC Annex I prose; these are the medicines represented in the released chunks. Extraction (script extract_corpus.py, PyMuPDF) was restricted to Annex I clinical prose, deliberately excluding Annex II/III labelling and leaflet text and the Section 6 pharmaceutical-particulars tail (excipients, shelf-life, marketing-authorisation boilerplate). Text was chunked at a soft window of ~150 words (hard cap 255), sentence-boundary preferred, and sampled per drug proportionally to that drug's clinical-prose volume with a section weighting biased toward clinical-efficacy (SmPC §5.1), pharmacokinetics (§5.2), dosing (§4.2), and special-warnings (§4.4) content. The extraction workflow is reproducible and included in the deposit. Per-chunk provenance schema (every record in corpus.jsonl). { "text": "… (Polish SmPC clinical prose, ~512 tokens) …", "source_authority": "EMA", "source_document_type": "SmPC / Product Information", "source_url": "https://www.ema.europa.eu/en/medicines/human/EPAR/…", "medicine": "aclidinium", "brand_name": "Eklira Genuair", "language": "pl", "retrieved_at": "2026-05-20", "chunk_id": "EMA_aclidinium_pl_0001", "license_note": "EMA reproduction policy; source attribution required" } All 418 released chunks have source_authority = "EMA", source_document_type = "SmPC / Product Information" and language = "pl". Intended use. Calibration data for post-training quantisation (AWQ / AutoAWQ / GPTQ) of Polish-language and Polish-clinical LLMs; reuse as a corpus-controlled calibration set enabling like-for-like quantisation-quality comparison across model scales (e.g. 8B / 12B / 70B of the same family). What it is NOT. This is not a training set or an evaluation/benchmark set; it is not a question-answering, instruction-tuning, or clinical-decision-support dataset and must not be used to fine-tune clinical behaviour or to evaluate clinical accuracy. It contains no patient data or protected health information (PHI): SmPC documents describe medicinal products (indications, dosing, adverse reactions, pharmacokinetics, aggregate clinical trial data), not individuals. The absence of PHI was confirmed by an automated pattern scan and manual sampling. The corpus is not clinical advice and confers no clinical authority; the canonical, legally-authoritative product information remains the EMA-published SmPC. Rights and source attribution. This dataset is a calibration corpus that contains text chunks derived from official medicinal product information documents. Parts of the source text are derived from EMA-published Polish SmPC / Product Information documents. EMA source text: © European Medicines Agency. EMA-published documents are reproduced and distributed under EMA's content-reproduction policy, which permits reproduction and/or distribution, in whole or in part, for non-commercial and commercial purposes, provided that EMA is always acknowledged as the source. Parts of the source text may derive from Polish national medicinal product documentation available via URPL / Polish public-sector sources; reuse of those parts is subject to the applicable source-specific public-sector information reuse rules and the original document provenance. Compilation, drug selection, extraction workflow, chunking, dataset structuring and metadata: Łukasz Minarowski / navimed-umb. No claim is made that the underlying SmPC / ChPL source text is licensed under CC-BY-4.0, CC-BY-NC-4.0, MIT, Apache-2.0 or any other open-source software licence. Users are responsible for preserving source attribution and for checking source-specific restrictions before redistribution or downstream use. The licence on this record is therefore set to "other" (per the EMA public-reproduction policy and source-specific reuse terms); the author's compilation contribution is the author's own work but does not relicense the underlying source text. Context. Produced for NaviMed-UMB (https://github.com/kicrazom/navimed-umb), a local LLM benchmarking and clinical-decision-support feasibility project at the Medical University of Białystok. This corpus was first published on HuggingFace alongside the 2026-05-23 Llama-PLLuM-70B AWQ release (inherited under Zenodo DOI 10.5281/zenodo.20317011) and is referenced by the NaviMed-UMB v0.3.0 and v0.4.0 software records; this deposit makes the corpus independently citable with its own DOI. AI assistance disclosure. Project documentation was prepared with assistance from large language models (Claude, Anthropic; GPT, OpenAI; Gemini, Google), with a locally-served Bielik-11B-v3.0-instruct-AWQ used for Polish-language editorial work. All drug selection, extraction design, provenance assignment and scientific claims are the author's. See AI_USAGE_DISCLOSURE.md in the repository.

Corpus text is EMA-derived and is NOT released under CC-BY or any open-source software licence; licence is "other" (EMA public-reproduction policy + source-specific reuse terms). EMA must always be acknowledged as the source. The author's compilation/extraction/metadata are the author's own work. No patient data / no PHI. Not a training or evaluation set — calibration only. No throughput/latency/benchmark numbers are included (those remain under the NaviMed-UMB METHODOLOGY publication embargo).