pfnet/Preferred-MedRECT-32B
Preferred Networks' Preferred-MedRECT-32B is a 32 billion parameter language model, fine-tuned from Qwen3-32B using LoRA, specifically for medical error detection and correction. It excels at identifying errors, extracting erroneous sentences, and providing corrections in clinical texts across both Japanese and English. The model was trained on bilingual medical reasoning data that includes explicit reasoning processes, making it highly specialized for medical text analysis.
Loading preview...
Overview
Preferred-MedRECT-32B is a 32 billion parameter model developed by Preferred Networks, Inc., fine-tuned from Qwen3-32B using LoRA. Its primary specialization is medical error detection and correction in clinical texts. The model was trained on a unique dataset of bilingual (Japanese/English) medical reasoning data, which includes explicit reasoning processes to enhance its ability to identify and rectify errors.
Key Capabilities
- Medical Error Detection: Accurately identifies errors within clinical texts.
- Sentence Extraction: Pinpoints and extracts sentences containing errors.
- Error Correction: Provides corrected versions of erroneous clinical text segments.
- Bilingual Support: Optimized for both Japanese and English medical texts.
- Reasoning-Enhanced: Benefits from training data that includes explicit reasoning processes, improving its analytical capabilities in medical contexts.
Performance Highlights
On the MedRECT-ja benchmark, Preferred-MedRECT-32B achieved an Error Detection F1 score of 0.743, Sentence Extraction Accuracy of 81.5%, and an EC Avg. Score of 0.627, outperforming its base model Qwen3-32B (think) across all three metrics. For MedRECT-en, it showed strong performance with an Error Detection F1 of 0.728, Sentence Extraction Accuracy of 90.9%, and an EC Avg. Score of 0.718, demonstrating robust cross-lingual capabilities.
Limitations
This model is developed for research purposes only and is not intended for clinical diagnosis. Users are responsible for ensuring compliance with applicable regulations when deploying the model.