README.md

---
library_name: peft
base_model: Qwen/Qwen3.5-4B-Base
pipeline_tag: text-classification
tags:
- biojev
- biomedical
- qwen3.5
- peft
- qlora
- natural-language-inference
- biomedical-nlp
- system-one
language:
- en
---

# BioJev-4B

**BioJev-4B** is the main 4B biomedical decision model of the BioJev project.

It is built from **Qwen/Qwen3.5-4B-Base** and follows the full BioJev pipeline:

```text
Qwen3.5-4B-Base
        ↓
Biomedical DAPT — PubMed + PMC
        ↓
General NLI — SNLI + MNLI + ANLI
        ↓
Biomedical NLI — BioNLI + NLI4CT
        ↓
BioJev-4B
```

The released checkpoint is a **PEFT/QLoRA sequence-classification adapter** with three NLI labels:

```text
0 → contradiction
1 → entailment
2 → neutral
```

## Links

- BioJev-4B weights: https://huggingface.co/Gabriel382/BioJev
- BioJev-Nano weights: https://huggingface.co/Gabriel382/BioJev-Nano
- Source code: https://github.com/Gabriel382/BioJev

## Authors

**Creator:** Gabriel Henrique Alencar Medeiros  
**Supervisor:** Lina F. Soualmia

BioJev is developed at **LITIS / Université de Rouen Normandie**.

## Training recipe

### Biomedical DAPT

```text
Biomedical token budget: 100M
Sequence length:          2048
Corpus:                   80% PubMed / 20% PMC
Method:                   QLoRA
```

### General NLI

```text
SNLI      35,000
MNLI      45,000
ANLI R1   20,000
----------------
Total    100,000
```

### Biomedical NLI

```text
BioNLI    30,000
NLI4CT    10,000
----------------
Total     40,000
```

The released checkpoint is the full configuration:

```text
DAPT → General NLI → Biomedical NLI
```

## Evaluation

| Dataset | Macro-F1 |
|---|---:|
| BioNLI | 94.35 |
| NLI4CT | 70.97 |
| ChemProt | 23.60 |
| DDI2013 | 42.01 |
| BioRED | 26.89 |

Relation-transfer mean over ChemProt, DDI2013 and BioRED: **30.83 macro-F1**.

Selected reliability results:

| Dataset | Accuracy (%) | Macro-F1 (%) | ECE (%) | AURC |
|---|---:|---:|---:|---:|
| BioNLI | 94.78 | 94.35 | 1.74 | 0.01 |
| NLI4CT | 71.00 | 70.97 | 16.00 | 0.17 |
| ChemProt | 24.74 | 23.60 | 35.92 | 0.68 |
| DDI2013 | 42.80 | 42.01 | 10.23 | 0.58 |
| BioRED | 42.44 | 26.89 | 16.03 | 0.50 |

BioNLI and NLI4CT participate in the biomedical decision-training pipeline and should not be interpreted as clean zero-shot transfer tasks for the full model.

# Loading BioJev-4B

Because BioJev is a **three-class PEFT sequence classifier**, reconstruct the base model with `num_labels=3` before loading the adapter.

```bash
pip install -U torch transformers peft accelerate
```

For optional quantized inference:

```bash
pip install -U bitsandbytes
```

```python
import torch
from peft import PeftConfig, PeftModelForSequenceClassification
from transformers import AutoModelForSequenceClassification, AutoTokenizer

MODEL = "Gabriel382/BioJev"

LABEL2ID = {
    "contradiction": 0,
    "entailment": 1,
    "neutral": 2,
}
ID2LABEL = {v: k for k, v in LABEL2ID.items()}

peft_config = PeftConfig.from_pretrained(MODEL)

base = AutoModelForSequenceClassification.from_pretrained(
    peft_config.base_model_name_or_path,
    num_labels=3,
    label2id=LABEL2ID,
    id2label=ID2LABEL,
    dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
    device_map="auto" if torch.cuda.is_available() else None,
)

model = PeftModelForSequenceClassification.from_pretrained(
    base,
    MODEL,
    is_trainable=False,
)

tokenizer = AutoTokenizer.from_pretrained(MODEL)

if tokenizer.pad_token_id is None:
    tokenizer.pad_token = tokenizer.eos_token

model.config.pad_token_id = tokenizer.pad_token_id
model.eval()

premise = "The clinical report describes a bacterial pneumonia."
hypothesis = "The patient has an infectious pulmonary disease."

inputs = tokenizer(
    premise,
    hypothesis,
    return_tensors="pt",
    truncation=True,
    max_length=2048,
)

device = next(model.parameters()).device
inputs = {k: v.to(device) for k, v in inputs.items()}

with torch.inference_mode():
    logits = model(**inputs).logits[0].float()
    probs = torch.softmax(logits, dim=-1)

for label, idx in LABEL2ID.items():
    print(f"{label:14s}: {probs[idx].item():.4f}")
```

# Jev / System One compatibility API

The main BioJev repository exposes existing BioJev checkpoints through:

```text
POST /v1/systemone
```

Supported decision types:

```text
choice
noul
score
```

This is an **engineering compatibility bridge** over BioJev's NLI classifier. BioJev-4B is not a native Ollama/System-One checkpoint.

Clone the main project:

```bash
git clone https://github.com/Gabriel382/BioJev.git
cd BioJev
python -m venv .venv
source .venv/bin/activate
pip install -e .
pip install fastapi uvicorn
```

Then serve a local copy of this model repository:

```bash
python scripts/serve_systemone.py   --checkpoint /path/to/BioJev   --model-name biojev-4b   --load-in-4bit   --host 0.0.0.0   --port 8000
```

Check the service:

```bash
curl http://127.0.0.1:8000/health
```

Example decision request:

```bash
curl http://127.0.0.1:8000/v1/systemone   -H "Content-Type: application/json"   -d '{
    "model": "biojev-4b",
    "state": "The patient has fever, productive cough, and a new lobar infiltrate.",
    "questions": {
      "diagnosis": {
        "type": "choice",
        "instructions": "Which diagnosis is best supported?",
        "criteria": {
          "pneumonia": "Community-acquired pneumonia",
          "asthma": "Acute asthma exacerbation",
          "migraine": "Migraine"
        }
      }
    }
  }'
```

The bridge computes:

```text
support(candidate) = entailment_logit - contradiction_logit
probabilities = softmax(candidate_supports)
```

Candidates are scored sequentially with `batch_size=1` for robust Qwen3.5/PEFT compatibility.

## Repository files

```text
BioJev/
├── adapter_config.json
├── adapter_model.safetensors
├── tokenizer.json
├── tokenizer_config.json
├── chat_template.jinja
├── training_manifest.json
├── README.md
├── example_inference.py
├── requirements.txt
└── SYSTEMONE_API.md
```

## Scientific status

BioJev is a research project.

The model, confidence estimates, and System One compatibility layer are **not validated clinical decision systems** and must not be used as the sole basis for diagnosis, treatment, or other high-stakes medical decisions.

## Citation

A formal paper citation will be added when the BioJev publication is available.

Until then, please cite the BioJev repository and the specific Hugging Face checkpoint used.

## Acknowledgements

BioJev-4B was created by **Gabriel Henrique Alencar Medeiros** under the supervision of **Lina F. Soualmia** at **LITIS / Université de Rouen Normandie**.