Downloads · 30 days
8
14% of all-time downloads
n0rths7ar/bioevidence-qwen3-4b-extraction-lora-v2
bioevidence-qwen3-4b-extraction-lora-v2 is a text generation model from n0rths7ar. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as apache-2.0.
This is the second QLoRA adapter for query-focused structured extraction from one PubMed title and abstract. It targets the versioned ModelEvidenceExtraction JSON schema used by BioEvidence Copilot.
Downloads · 30 days
8
14% of all-time downloads
All-time downloads
57
Public
Repo size
77.6 MB
Likes
0
Public
Click a slice to open those files.
.safetensors66.1 MB · 85%
From the Hugging Face model README
This is the second QLoRA adapter for query-focused structured extraction from
one PubMed title and abstract. It targets the versioned
ModelEvidenceExtraction JSON schema used by BioEvidence Copilot.
The adapter is an engineering experiment in strict structured output and local GPU deployment. It is not a clinical model and its draft annotations are not an expert biomedical benchmark.
unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit7744afa8566e264af1a92a806d8d9aae00cc7c782e-4All systems used the same 13-row PMID-held-out draft test set and deterministic decoding.
| Metric | Rules | Prompted base | Adapter v1 | Adapter v2 |
|---|---|---|---|---|
| Strict JSON parse rate | 1.000 | 1.000 | 1.000 | 1.000 |
| Schema validity | 1.000 | 0.923 | 1.000 | 1.000 |
| Grounding rate | n/a | 0.615 | 0.923 | 0.923 |
| Evidence-status accuracy | 0.308 | 0.769 | 0.462 | 0.615 |
| Study-design accuracy | 0.769 | 0.769 | 0.692 | 0.923 |
| Semantic-field token F1 | 0.487 | 0.512 | 0.595 | 0.681 |
| Outcome-name token F1 | 0.231 | 0.326 | 0.632 | 0.526 |
| Outcome-direction accuracy | 0.308 | 0.410 | 0.692 | 0.692 |
| Evidence-span token F1 | 0.409 | 0.342 | 0.577 | 0.515 |
| Evidence-span support rate | 1.000 | 0.737 | 0.923 | 0.923 |
| Mean generation time | n/a | 19.58 s | 6.81 s | 11.23 s |
Adapter v2 fixed the observed indirect -> none collapse on this test: all four
indirect rows remained indirect. It did not solve direct-status calibration;
all three direct rows were classified as indirect. It also traded lower
outcome-name/span metrics and higher latency for improved status, study-design,
and semantic-field metrics over adapter v1.
BioEvidence Copilot validates every generated object against its Pydantic schema and requires verbatim evidence spans. Invalid model output falls back to the deterministic rules backend, with attempted backend, used backend, and fallback reason exposed per evidence row.
A three-paper product smoke completed with three model successes and no
fallbacks. The v1 schema failure for PMID 41772161 was resolved by adapter v2.
The source repository contains the schema, annotations, split manifest, training script, evaluation code, aggregate result summary, and optional product backend. The adapter package contains LoRA weights and tokenizer/chat-template files, but does not redistribute the training dataset or base-model weights.