OmkarShewale/clintrial-qwen2.5-7b-sft
OmkarShewale/clintrial-qwen2.5-7b-sft is a 7.6 billion parameter Qwen2.5-7B-Instruct model fine-tuned using QLoRA for clinical trial understanding. It specializes in tasks like eligibility criteria extraction into structured JSON, plain-language summaries, condition Q&A, and phase classification. This model demonstrates significantly improved performance and format discipline on clinical trial-specific tasks compared to its base model.
Loading preview...
ClinTrial-Qwen2.5-7B-SFT: Clinical Trial Understanding Model
This model is a QLoRA-fine-tuned Qwen2.5-7B-Instruct variant, specifically designed for tasks related to clinical trial understanding. It was trained on approximately 26,000 instruction examples derived from the ClinicalTrials.gov registry, focusing on extracting and summarizing information from trial descriptions.
Key Capabilities
- Eligibility Extraction: Converts free-text inclusion/exclusion criteria into structured JSON, achieving 100% schema-valid JSON output and high F1 scores (0.968).
- Plain-Language Summary: Generates patient-friendly summaries from complex trial descriptions.
- Condition Q&A: Answers questions about the medical conditions studied in a trial.
- Phase Classification: Identifies the phase of a clinical trial with high accuracy (0.794 exact match).
Performance Highlights
Evaluated on a held-out test set, the fine-tuned model significantly outperforms the base Qwen2.5-7B model across all specialized tasks. For instance, it achieves a criterion F1 of 0.968 for eligibility extraction and an exact match of 0.794 for phase classification, where the base model scored 0.000 due to format adherence issues. This fine-tuning ensures not only improved accuracy but also strict adherence to specified output formats.
Training Details
The model was trained using QLoRA (4-bit NF4 + LoRA) targeting key modules, with only ~0.5% of total parameters being trainable. Best-checkpoint selection based on eval_loss prevented overfitting, ensuring optimal performance. The training data was sourced from 8,000 real ClinicalTrials.gov studies, with labels derived from the registry's own structured fields, ensuring auditable and reliable targets.
Limitations
- Not medical advice: Outputs are for research/education only and require human expert review.
- Trained exclusively on English registry text; performance on other formats or languages is not guaranteed.
- Summary quality (ROUGE-L 0.29) is the weakest task.
- Generalization could be improved with more training data, as only 8k of ~500k available trials were used.