supervisio/herbert-base-ner-clinical-pl

A continued fine-tune of pczarnik/herbert-base-ner at commit 8770e1d2b8c273e4dababfaa7d30a7443005ff70 for Polish clinical-note pseudonymization: plural family names (u Kowalskich, z Nowakami), sentence-initial surnames, diminutives and names that are also common nouns (Marek/marka, Lis, Kowal). Same label set (PER, LOC, ORG, BIO), same tokenizer, same head. Licence CC-BY-4.0, carried through from the base model.

Trained by therapist-copilot issue M4-18 (#596). The training data is synthetic — template sentences over declension tables and Morfeusz (SGJP) paradigms — and contains no clinical text.

Evidence status. The numbers below were taken on an arm64 laptop against this checkpoint before it was pinned. They are a candidate's numbers, not a deployment's; the x86 measurement that decides whether this checkpoint ships is recorded in the repository's docs/testing/pseudonymization-eval.md § M4-18.

Training

base	pczarnik/herbert-base-ner @ 8770e1d2b8c273e4dababfaa7d30a7443005ff70
data	scripts/gen-ner-train.py, seed 20260916, SGJP pl.sgjp.sgjp-2026.06.01
train / dev	6522 / 784 sentences (dev over disjoint names)
train sha256	54c60fbc8ffc56cc2530073eeb30ec5be34bbca3c7cec08b2943265da62d0dc5
held out by lemma	Kowalska, Kowalski, Nowak, Wiśniewska, Wiśniewski, Zielińska, Zieliński
frozen	embeddings + 6 encoder layers (42532615 of 123857671 parameters trainable)
optimizer	AdamW, lr 2e-05, weight decay 0.01, grad clip 1.0
schedule	3 epochs, linear warmup, linear decay, warm-up 0.1
batch / length	16 / 128
seed	20260916
hardware	Apple M3 Pro, mps, 176 s
versions	torch 2.13.0, transformers 5.14.1
synthetic dev, last epoch	PER F1 1.0, LOC F1 1.0, ORG F1 1.0
model.safetensors sha256	e538c2498eedb5cf774844dcbf73c4ee032f89522a79ebe8bcd09e28fab8d0f1

Files

config.json, model.safetensors, the tokenizer files and train-record.json. No pickle is published: the consuming service loads safetensors only.

Reproduce

cd services/nlp-sidecar
python scripts/gen-ner-train.py                 # the data, from the seed above
python scripts/finetune-ner.py                  # the checkpoint
python -m pl_eval --ner real --model-path .train/checkpoint