bibocat/qwen3-ner-grpo-v2-merged
bibocat/qwen3-ner-grpo-v2-merged is an 8 billion parameter Qwen3 model fine-tuned for Vietnamese medical Named Entity Recognition (NER). This BF16 merged checkpoint integrates a Qwen3 SFT model with a GRPO v2 LoRA adapter, specifically trained to identify medical entities. It excels at extracting structured information like entities, types, positions, assertions, and candidates from Vietnamese medical text, outputting results as a JSON array.
Loading preview...
Overview
The bibocat/qwen3-ner-grpo-v2-merged model is an 8 billion parameter Qwen3 variant specifically designed for Vietnamese medical Named Entity Recognition (NER). It represents a BF16 merged checkpoint, combining a Qwen3 Vietnamese medical SFT (Supervised Fine-Tuning) model with a grpo_v2 LoRA adapter.
Key Capabilities
- Specialized NER: Fine-tuned to accurately identify and extract medical entities from Vietnamese text.
- GRPO v2 Adapter: Utilizes a GRPO v2 LoRA adapter, trained for one epoch on a specific
grpo_v2dataset with a span-aware v2 reward function, enhancing its NER performance. - Structured Output: Generates output as a JSON array, including fields such as
entity,type,position,assertions, andcandidates, making it suitable for downstream processing. - Integrated Tokenizer: Designed to be used with its bundled tokenizer and the standard Qwen3 chat template.
Good For
- Vietnamese Medical Text Analysis: Ideal for applications requiring precise entity extraction from medical documents, clinical notes, or research papers written in Vietnamese.
- Structured Data Extraction: Useful for converting unstructured medical text into structured, machine-readable formats for databases or analytical tools.
- Research and Development: Provides a specialized tool for researchers working on natural language processing in the medical domain for the Vietnamese language.