lighteternal/biodecision-tev1-4b
BioDecision-4B by lighteternal is a 4.5 billion parameter System-1 decision model built on Qwen3.5-4B, specifically designed for biomedicine, pharma, and clinical trials. It excels at tasks like clinical-trial outcome forecasting, patient-trial matching, and judging AI-written answers against sources, achieving 70.2% accuracy across 28 biomedical benchmarks. The model provides calibrated probabilities for predefined answers (Choice, Yes/No, Score) in a single forward pass, making it efficient for rapid decision-making in specialized biomedical contexts.
Loading preview...
BioDecision-4B: A Specialized Biomedical Decision Model
BioDecision-4B, developed by lighteternal, is a 4.5 billion parameter model built upon the Qwen3.5-4B architecture, specifically engineered for rapid, System-1 decision-making in biomedical, pharmaceutical, and clinical trial applications. Unlike general-purpose LLMs, it focuses on providing calibrated probabilities for predefined answer options (Choice, Yes/No, Score) in a single forward pass, eliminating text generation.
Key Capabilities and Performance
- Biomedical Benchmark Excellence: Achieves 70.2% accuracy across 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 macro-F1 scores of 72.6 and 62.0, comparable to or exceeding larger models like GPT-5 and Claude Opus 4.5.
- Patient-Trial Matching: Significantly improves NDCG@10 to 0.833 for patient-trial matching (TREC 2022).
- AI Answer Grounding: Excels at judging AI-written answers against source texts (HaluBench PubMedQA) with 87.9% accuracy, on par with GPT-4.1.
- Efficiency: Delivers decisions rapidly at 44 ms per decision (93 decisions/s) on an A100 GPU, offering cost-effective inference.
Unique Approach
BioDecision-4B is fine-tuned using Together AI's Tev1 recipe, training the model to emit only the letter of the correct option, directly translating the letter distribution into the answer probability distribution. It is designed for research use and is not validated for patient care.