thangvip/qwen3-4b-vietnamese-legal-grpo-phase-2

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 24, 2025License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The thangvip/qwen3-4b-vietnamese-legal-grpo-phase-2 is a 4 billion parameter Qwen 3 model developed by thangvip, specifically fine-tuned for advanced Vietnamese legal reasoning. This Phase 2 model utilizes Group Relative Policy Optimization (GRPO) on hard-difficulty legal question-answering data, excelling at complex syllogistic reasoning and generating structured legal arguments with citations. It supports an extended context length of 8192 tokens, making it suitable for intricate legal analysis in Vietnamese.

Loading preview...

Model Overview

This model, thangvip/qwen3-4b-vietnamese-legal-grpo-phase-2, is a 4 billion parameter Qwen 3 architecture developed by thangvip. It represents the Phase 2 of a specialized training regimen, focusing on hard difficulty Vietnamese legal reasoning. The model is fine-tuned using Group Relative Policy Optimization (GRPO) to handle complex syllogistic arguments, generating structured legal responses with citations.

Key Capabilities

  • Advanced Legal Reasoning: Specifically trained on hard-difficulty Vietnamese legal Q&A scenarios.
  • Syllogistic Structure: Generates responses following a Major Premise → Minor Premise → Conclusion format.
  • Vietnamese Legal Domain: Optimized for legal texts and questions in Vietnamese.
  • Citation Support: Includes relevant legal citations in its outputs.
  • Extended Context: Supports up to 8192 tokens for processing and generating complex reasoning chains.
  • Structured Output: Utilizes XML-like tags for organized and parseable responses.

Training Approach

This model builds upon a Phase 1 GRPO-trained model, progressing to more challenging legal problems. Its training incorporates a sophisticated reward system that prioritizes correctness, format compliance, citation accuracy, and reasoning quality, while penalizing hallucination and excessive length.

Use Cases

  • Complex Legal Education: Teaching advanced legal reasoning.
  • Advanced Legal Research: Preliminary analysis of intricate legal questions.
  • Multi-step Legal Analysis: Generating structured arguments for complex scenarios.
  • Legal Training: Demonstrating proper syllogistic reasoning for difficult legal problems.

Finetunes
1 models