BiomedBERT-base-uncased-abstract

BiomedBERT (formerly PubMedBERT) pretrained on PubMed abstracts; used by Mineris for mean-pooled abstract and claim embeddings.

This is a format conversion of microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract for Mineris' MLX inference stack (mineris_inference) — not a fine-tune. Weights are the upstream BertModel encoder trunk (+ pooler) cast to fp16 and stored as model.safetensors; a tokenizer.json is included so the model loads without transformers or torch.

Provenance

Upstream repo	microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract
Upstream commit	d673b8835373c6fa116d6d8006b33d48734e305d
License	mit (inherited from upstream)
Conversion notebook	notebook/notebooks/converter/convert_pubmedbert.ipynb (mineris repo)
Mineris commit	9817fe7bbeb2a4b23fd683144a28a5843af94af9
Version	v1.0.0
dtype	fp16
Tensors / size	199 / 219.0 MB

Conversion parity (fp16 mirror vs fp32 upstream, PyTorch)

Metric	Value
Texts	18
Tokenizer id mismatches (tokenizer.json via tokenizers vs upstream tokenizer)	0
Mean-pooled embedding cosine, per text — min / mean	1.000000 / 1.000000
Last hidden state, non-padding tokens — max abs diff / median relative diff	5.26e-02 / 1.19e-03
Threshold (min mean-pooled embedding cosine)	0.999

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 microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract@d673b88.