biored-ner-gemma-4-E2B-it

LoRA adapter for NER span extraction on BioRED, fine-tuned from google/gemma-4-E2B-it. Gemma 4 is gated -- request access on its model page before loading this adapter.

For reproducing results, refer to this GitHub repository (TBD).

Getting Started

from transformers import AutoModelForMultimodalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForMultimodalLM.from_pretrained("google/gemma-4-E2B-it", device_map="auto")
model = PeftModel.from_pretrained(base, "kiwiki196/biored-ner-gemma-4-E2B-it")
tokenizer = AutoTokenizer.from_pretrained("kiwiki196/biored-ner-gemma-4-E2B-it")

See repo for more details.

Prompt

Inference uses the following system prompt:

Extract text naming ChemicalEntity, DiseaseOrPhenotypicFeature, GeneOrGeneProduct, OrganismTaxon, SequenceVariant, or CellLine from the input. The text has positional markers already embedded (`<unused0>N<unused1>`, e.g. `<unused0>58<unused1>`) — each marker's id is arbitrary, not a reading-order count, so markers do not necessarily appear in ascending numeric order. These markers are not part of the content; never extract one as a match.

Copy each match verbatim from the input, including any irregular internal whitespace or embedded line breaks. Include negated, uncertain, and hedged mentions, but not the negation/hedging words themselves. Do not include anything outside the categories above.

Before every match, write the id of the nearest `<unused0>N<unused1>` marker before it, followed by `<unused2>` — repeat this prefix before EVERY match, even consecutive matches under the same marker; never omit or dedupe it — then the match text, then its `<CATEGORY_TOKEN>`, as `N<unused2>TEXT<CATEGORY_TOKEN>` (text first, category token last — not the other way around, and no other characters — never write markup like `<span>`). `<CATEGORY_TOKEN>` is one of: `<unused3>` ChemicalEntity, `<unused4>` DiseaseOrPhenotypicFeature, `<unused5>` GeneOrGeneProduct, `<unused6>` OrganismTaxon, `<unused7>` SequenceVariant, or `<unused8>` CellLine. Tag every occurrence separately, even repeats of the same text — do not deduplicate. Output only this, with no explanation or markdown fences.

Results

Category + text (multiset)

category	P	R	F1
micro	0.730	0.737	0.733
macro	0.724	0.673	0.689
CellLine	0.786	0.449	0.571
ChemicalEntity	0.677	0.785	0.727
DiseaseOrPhenotypicFeature	0.691	0.749	0.719
GeneOrGeneProduct	0.833	0.732	0.779
OrganismTaxon	0.760	0.785	0.772
SequenceVariant	0.598	0.537	0.566

Positional (char-level)

category	P	R	F1
micro	0.814	0.838	0.826
macro	0.731	0.694	0.704
CellLine	0.693	0.435	0.535
ChemicalEntity	0.706	0.839	0.767
DiseaseOrPhenotypicFeature	0.767	0.812	0.789
GeneOrGeneProduct	0.793	0.725	0.758
OrganismTaxon	0.683	0.764	0.721
SequenceVariant	0.746	0.586	0.656

Note: 2/201 (1.0%) of test set generations were unparseable (no credit).

Citation

Training data: Luo et al., "BioRED: a rich biomedical relation extraction dataset," Briefings in Bioinformatics, 2022.