Niraya666/Qwen3-VL-4B-Instruct-WMVLM-RL-0213
Niraya666/Qwen3-VL-4B-Instruct-WMVLM-RL-0213 is a 4 billion parameter vision-language model, part of the WaferSAGE framework, specifically designed for semiconductor wafer map defect analysis. This model is initialized from a WaferSAGE supervised fine-tuned Qwen3-VL and further optimized with curriculum-based reinforcement learning using rubric-aligned rewards. It excels at defect type identification, spatial distribution analysis, and hallucination reduction in wafer map understanding. The model is primarily intended for visual question answering tasks within the semiconductor domain, offering domain-adapted terminology and local deployment capabilities.
Loading preview...
WaferSAGE-RL: Rubric-Guided Qwen3-VL for Wafer Map Defect Analysis
This model, part of the WaferSAGE framework, is a 4 billion parameter vision-language model (VLM) specialized in semiconductor wafer map understanding. It is built upon a WaferSAGE supervised fine-tuned Qwen3-VL base and enhanced through curriculum-based reinforcement learning (RL) using rubric-aligned rewards.
Key Capabilities & Differentiators
- Domain-Specific Optimization: Tailored for wafer map defect analysis, including defect identification, spatial distribution, and morphological description.
- Rubric-Guided Reinforcement Learning: Utilizes structured rubrics with positive and negative criteria as reward signals during RL, significantly reducing hallucination and encouraging accurate, evidence-based responses.
- Curriculum Learning: Training progresses from easier recognition tasks to more complex multi-defect reasoning and root-cause hypothesis generation.
- Hallucination Reduction: Explicitly penalizes incorrect locations, defect types, or unsupported root-cause explanations.
- Local Deployment: Supports local deployment, making it suitable for privacy-sensitive semiconductor environments.
- Performance: Achieved a 6.493 LLM-Judge score, closely approaching Gemini-3-Flash, while enabling local operation.
Recommended Use Cases
- Wafer map Visual Question Answering (VQA) research.
- Industrial VLM evaluation and semiconductor AI assistant prototypes.
- Local on-premise proof-of-concept deployments.
- Research into synthetic data and rubric-reward mechanisms.
Limitations
It's crucial to note that this model provides candidate hypotheses for root causes and does not perform definitive diagnosis. It operates without external data like lot history or metrology, and its outputs should always be reviewed by experts, especially for high-cost manufacturing decisions.