BioJev-SystemOne

Biomedical decision models for native Ollama System One inference.

BioJev-SystemOne is a family of research models built on Qwen3.5, specialized for structured decisions over biomedical and general-language evidence.

Unlike the original BioJev NLI classifiers, BioJev-SystemOne uses a causal language-model architecture to select decision candidates directly through Ollama’s native /v1/systemone endpoint, without requiring an external Python inference bridge.

Available models

Installation

Pull your preferred model:

ollama pull Gabriel382/biojev-systemone:0.8b
ollama pull Gabriel382/biojev-systemone:4b
ollama pull Gabriel382/biojev-systemone:9b

A version of Ollama supporting native System One decision models is required.

Usage

BioJev-SystemOne supports three structured decision types:

• choice — Select the best-supported option among candidates.
• noul — Estimate support for a binary decision.
• score — Score an ordered set of alternatives.

Example request:

curl http://localhost:11434/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Gabriel382/biojev-systemone:4b",
    "state": {
      "case": "A 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 response contains a structured decision, candidate probabilities and confidence information.

These models are designed for structured decision inference, rather than conventional open-ended chat.

Training methodology

The models follow a three-stage research pipeline:

1. Base model

Qwen3.5-0.8B-Base, Qwen3.5-4B-Base or Qwen3.5-9B-Base.

2. Biomedical domain adaptation

BioQwen domain-adaptive pretraining on biomedical text. The 4B and 9B models used approximately 100 million tokens from PubMed and PMC, while Nano used a smaller pilot adaptation.

3. System One decision fine-tuning

All three models were fine-tuned using 147,048 examples combining general decision tasks, natural language inference, BioNLI and NLI4CT.

The training uses QLoRA and completion-only supervision over candidate letters (A–Z), enabling native structured decision scoring.

Evaluation results

The following results come from a common development evaluation of 6,768 examples across 11 datasets using Ollama’s native System One API.

Key findings

BioJev-SystemOne-4B achieved the highest overall development accuracy, offering a strong balance between model size and performance.

BioJev-SystemOne-9B achieved the highest biomedical NLI accuracy, particularly on clinical-trial inference tasks.

BioJev-SystemOne-Nano achieved substantially lower inference latency while retaining competitive decision performance.

The default context is 2,048 tokens. Longer NLI4CT examples were evaluated using separate, selectively extended context profiles without changing the original model weights.

These are development-set results, not a fully independent test evaluation. Further independent benchmarking is required before making state-of-the-art claims.

Limitations

BioJev-SystemOne is a research model family and has not been clinically validated.

It must not be used as the sole basis for medical diagnosis, treatment selection or other high-stakes clinical decisions.

Decision scores and confidence estimates should not be interpreted as calibrated probabilities of clinical correctness.

Links

Source code: https://github.com/Gabriel382/BioJev

Ollama: https://ollama.com/Gabriel382/biojev-systemone

Authors

Gabriel Henrique Alencar Medeiros

Research supervision: Lina F. Soualmia

LITIS — Université de Rouen Normandie, France.