maveryn/trace-qwen2.5-vl-3b
The maveryn/trace-qwen2.5-vl-3b is a 3 billion parameter vision-language model, a GRPO checkpoint derived from Qwen/Qwen2.5-VL-3B-Instruct. It is specifically fine-tuned on 64,000 grounded visual-reasoning examples from the maveryn/trace dataset. This model is optimized for multimodal reasoning tasks, demonstrating a 3.51% improvement in overall score on the trace_eval_v1 benchmark compared to its base model.
Loading preview...
TRACE Qwen2.5-VL 3B Overview
This model, developed by Maveryn, is a 3 billion parameter vision-language (VL) model, specifically a GRPO (Grounded Reasoning Policy Optimization) checkpoint. It is built upon the Qwen/Qwen2.5-VL-3B-Instruct base model and has been extensively fine-tuned on the maveryn/trace dataset, which comprises 64,000 grounded visual-reasoning examples across 1,000 tasks.
Key Capabilities & Training
- Enhanced Visual Reasoning: The model is specialized in processing and reasoning over visual inputs, trained to produce structured JSON answers for visual queries.
- GRPO Fine-tuning: It underwent 500 steps of GRPO training, focusing on generating exact and valid JSON answers, with a reward structure prioritizing precise JSON output.
- Performance Improvement: On the
trace_eval_v1benchmark, it achieves an overall score of 42.85 ± 0.39, representing a +3.51 ± 0.25 improvement over its base model,Qwen2.5-VL-3B-Instruct.
Intended Use Cases
- Research in Multimodal Reasoning: Ideal for researchers exploring advanced multimodal reasoning capabilities and verifiable-reward post-training methods.
- Structured Output Generation: Particularly suited for applications requiring structured, JSON-formatted answers based on visual and textual prompts.
Limitations
It's important to note that this model is intended for research and may produce incorrect answers, unsupported reasoning, or unreliable grounding. It has not been validated for safety-critical, medical, legal, or autonomous decision-making applications.