lighteternal/biodecision-v2-4b
BioDecision-4B v2 by lighteternal is a 4.5 billion parameter System-1 decision model, retrained from Qwen3.5-4B, specifically designed for biomedicine, pharma, and clinical trials. It processes a source, question, and answer options to return calibrated probabilities without generating text. This model excels in domain-specific decision tasks, outperforming its predecessor and general models like TypeSafe's Jev 1.13 on biomedical datasets, and supports a 32768-token context length.
Loading preview...
BioDecision-4B v2: Specialized Biomedical Decision Model
BioDecision-4B v2 is a 4.5 billion parameter model developed by lighteternal, retrained from Qwen3.5-4B, and optimized for System-1 decision-making in biomedicine, pharmaceuticals, and clinical trials. Unlike generative LLMs, it reads a source, a question, and a set of possible answers, returning a calibrated probability for each answer without generating text. This version significantly improves upon BioDecision v1, incorporating 22 new data sources and 6 new use cases, leading to a 1.75 million decision training pool.
Key Capabilities
- Biomedical Decision Accuracy: Achieves 75.3% accuracy on 63,432 held-out test decisions, an improvement from BioDecision v1's 69.4%. It outperforms TypeSafe's Jev 1.13 on 17 out of 21 biomedical, pharma, and clinical-trial datasets.
- Specialized Domain Expertise: Demonstrates strong performance in areas like drug-protein and drug-drug relations (DrugProt F1 71.7, DDI-2013 F1 68.6), genomics, diagnosis, and regulatory work.
- Confidence Calibration: Provides calibrated probabilities for answers, with 98.5% accuracy when the model is at least 90% confident, covering 60% of answers.
- Efficient Inference: Achieves a median latency of 39 ms per decision and 95 decisions/s in batches on an A100 80GB GPU, making it suitable for high-throughput applications.
- Qwen3.5-4B Architecture: Built on the Qwen3.5-4B architecture with a LoRA adapter, supporting a context length of up to 8,192 tokens for training.
Good For
- Biomedical Research: Analyzing drug and device safety, regulatory compliance, clinical data standards, and genomics.
- Clinical Trial Management: Forecasting outcomes and matching, though BioDecision v1 may be better for some CT Open tasks.
- Automated Decision Support: Applications requiring rapid, calibrated probabilistic answers in specialized biomedical contexts.
- Relation Extraction: Identifying drug-protein and drug-drug interactions from text.
- Benchmarking: Evaluating and comparing performance against other models on specific biomedical decision tasks.