README.md

---
library_name: peft
base_model: Qwen/Qwen3.5-9B-Base
pipeline_tag: text-classification
tags:
- biojev
- biomedical
- qwen3.5
- peft
- qlora
- natural-language-inference
- biomedical-nlp
- system-one
language:
- en
---

# BioJev-9B

**BioJev-9B** is the 9B member of the BioJev biomedical decision-model family.

Built from **Qwen/Qwen3.5-9B-Base**, it follows the full BioJev pipeline:

```text
Qwen3.5-9B-Base
        ↓
Biomedical DAPT — PubMed + PMC
        ↓
General NLI — SNLI + MNLI + ANLI
        ↓
Biomedical NLI — BioNLI + NLI4CT
        ↓
BioJev-9B
```

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

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

## Links

- BioJev-9B: https://huggingface.co/Gabriel382/BioJev-9B
- BioJev-4B: https://huggingface.co/Gabriel382/BioJev
- BioJev-Nano: https://huggingface.co/Gabriel382/BioJev-Nano
- Source code: https://github.com/Gabriel382/BioJev

## Authors

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

Developed at **LITIS / Université de Rouen Normandie**.

## Training recipe

### Biomedical DAPT

```text
100M biomedical tokens
80% PubMed / 20% PMC
sequence length 2048
QLoRA
bfloat16
4-bit quantization
```

### General NLI

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

Dev results:

```text
Accuracy: 89.20%
Macro-F1: 89.15%
Loss:     0.2958
```

### Biomedical NLI

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

Dev results:

```text
Accuracy: 92.77%
Macro-F1: 92.30%
Loss:     0.1996
```

## Frozen evaluation

| Dataset | Role | Macro-F1 |
|---|---|---:|
| BioNLI | held-out biomedical NLI | **94.73** |
| NLI4CT | validation seen during full training | **72.50** |
| ChemProt | zero-shot relation typing | **31.13** |
| DDI2013 | zero-shot relation typing | **36.35** |
| BioRED | zero-shot relation typing | **34.32** |

**Zero-shot relation-transfer mean:** **33.93 macro-F1**

**All-5 mean:** **53.80 macro-F1**

### Reliability

| Dataset | Accuracy (%) | Macro-F1 (%) | ECE (%) | NLL | Mean confidence (%) |
|---|---:|---:|---:|---:|---:|
| BioNLI | 95.14 | 94.73 | 2.20 | 0.146 | 97.27 |
| NLI4CT | 72.50 | 72.50 | 14.70 | 0.693 | 86.30 |
| ChemProt | 32.24 | 31.13 | 10.34 | 1.995 | 41.00 |
| DDI2013 | 39.33 | 36.35 | 9.17 | 1.372 | 32.27 |
| BioRED | 50.11 | 34.32 | 6.52 | 1.156 | 54.64 |

### Scaling context

```text
Relation-transfer mean macro-F1

BioJev-Nano   21.01
BioJev-4B     30.83
BioJev-9B     33.93
```

The 9B model improves aggregate transfer relative to 4B, with strong gains on ChemProt and BioRED, while DDI2013 does not improve monotonically with model size.

This is an empirical model-size comparison, not a formal scaling-law study.

# Loading BioJev-9B

Install:

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

Optional 4-bit inference:

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

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

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

MODEL = "Gabriel382/BioJev-9B"
LABEL2ID = {"contradiction": 0, "entailment": 1, "neutral": 2}
ID2LABEL = {v: k for k, v in LABEL2ID.items()}

cfg = PeftConfig.from_pretrained(MODEL)

base = AutoModelForSequenceClassification.from_pretrained(
    cfg.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()
```

# System One compatibility API

The main BioJev repository exposes BioJev checkpoints through:

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

Supported types:

```text
choice
noul
score
```

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

Serve a local copy with:

```bash
python scripts/serve_systemone.py \
  --checkpoint /path/to/BioJev-9B \
  --model-name biojev-9b \
  --load-in-4bit \
  --host 0.0.0.0 \
  --port 8000
```

## Scientific status

BioJev is a research project. The model 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-9B was created by **Gabriel Henrique Alencar Medeiros** under the supervision of **Lina F. Soualmia** at **LITIS / Université de Rouen Normandie**.