lighteternal/biodecision-tev1-4b
BioDecision-4B by lighteternal is a 4.5 billion parameter System-1 decision model, based on Qwen3.5-4B with a 32K context length, specifically designed for biomedical, pharmaceutical, and clinical trial decision-making. It processes a source, question, and predefined answers to return calibrated probabilities for each option in a single pass, supporting various answer types without retraining for new tasks. The model excels in clinical-trial outcome forecasting, patient-trial matching, and judging AI-written answers against sources, demonstrating significant performance gains over general-purpose models in these specialized biomedical applications.
Loading preview...
BioDecision-4B: A Specialized Biomedical Decision Model
BioDecision-4B, developed by lighteternal, is a 4.5 billion parameter System-1 decision model built upon the Qwen3.5-4B architecture, optimized for rapid and calibrated decision-making in biomedicine, pharma, and clinical trials. Unlike general-purpose LLMs, it processes a source text, a question, and a set of possible answers to return a calibrated probability for each answer in a single forward pass. This design allows for flexible decision tasks (Choice, Yes/No, Score) without requiring retraining for new scenarios.
Key Capabilities
- Biomedical Decision-Making: Achieves 70.2% accuracy on 46,199 held-out decisions from 28 biomedical benchmarks, outperforming Qwen3.5-4B and Together's Tev1-4B by over 7 points.
- Clinical Trial Forecasting: Demonstrates strong performance in clinical-trial outcome forecasting (CT Open Endpoint), with 72.6% and 62.0% macro-F1, comparable to or exceeding larger models like GPT-5 and Claude Opus 4.5.
- Patient-Trial Matching: Improves NDCG@10 to 0.833 in TREC 2022 patient-trial matching, up from 0.713.
- AI-Answer Grounding: Scores 87.9% on HaluBench PubMedQA for judging AI-written answers against a source, on par with GPT-4.1.
- Efficiency: Delivers decisions rapidly at 44 ms per decision, achieving 93 decisions/s on an A100 GPU, with a low cost of $0.0074 per 1,000 decisions.
Good for
- Automated Biomedical Analysis: Ideal for tasks requiring quick, probabilistic decisions based on medical literature, clinical trial data, or pharmaceutical information.
- Clinical Trial Optimization: Useful for forecasting trial outcomes, matching patients to trials, and evaluating trial designs.
- Fact-Checking and Grounding: Excellent for verifying AI-generated content against source material in biomedical contexts.
- High-Throughput Decision Systems: Its speed and efficiency make it suitable for applications requiring a large volume of rapid, calibrated decisions.