pico-extractor

BIO token classifier tagging PICO elements (Population, Intervention, Comparator, Outcome) in clinical-trial abstracts; used by Mineris for PICO span extraction.

This is a format conversion of zuzab/pico-extractor for Mineris' MLX inference stack (mineris_inference) — not a fine-tune. Weights are the full upstream BertForTokenClassification (encoder + classification head) cast to fp16 and stored as model.safetensors; a tokenizer.json is included so the model loads without transformers or torch.

Labels

id	label
0	O
1	B-P
2	B-I
3	B-C
4	B-O
5	I-P
6	I-I
7	I-C
8	I-O

Provenance

Upstream repo	zuzab/pico-extractor
Upstream commit	ed4df86e482d9c53d953791a71c15caa5a687d96
License	apache-2.0 (declared by Mineris, the upstream owner)
Conversion notebook	notebook/notebooks/converter/convert_pico_extractor.ipynb (mineris repo)
Mineris commit	9817fe7bbeb2a4b23fd683144a28a5843af94af9
Version	v1.0.0
dtype	fp16
Tensors / size	199 / 215.3 MB

Conversion parity (fp16 mirror vs fp32 upstream, PyTorch)

Metric	Value
Texts	18
Tokenizer id mismatches (tokenizer.json via tokenizers vs upstream tokenizer)	0
Token argmax label agreement, per text — min / mean	1.000000 / 1.000000
Logits, non-padding tokens — max abs diff / median relative diff	9.82e-03 / 5.09e-04
Threshold (min token argmax label agreement)	0.995

MLX-runtime parity is tracked separately in the mineris notebook/notebooks/benchmark/ notebooks.

Versioning

Releases are semver git tags on this repo (vMAJOR.MINOR.PATCH); consumers pin a tag. MAJOR = output-incompatible change, MINOR = compatible re-conversion, PATCH = card/metadata only.

CHANGELOG

• v1.0.0 — initial conversion from zuzab/pico-extractor@ed4df86.