DangIT02/qwen3vl-flowchart-to-mermaid_v2

Qwen3-VL-8B Flowchart-to-Mermaid (v2)

This model, developed by DangIT02, is an 8 billion parameter vision-language model fine-tuned from unsloth/Qwen3-VL-8B-Instruct. Its primary function is to transcribe flowchart diagrams from images into valid Mermaid flowchart code, preserving nodes, edges, labels, and direction.

Key Capabilities & Features

• Flowchart-to-Mermaid Conversion: Converts visual flowcharts into Mermaid syntax.
• Canonicalized Node IDs: Outputs Mermaid code with standardized node IDs (A, B, C, etc.) for deterministic and tool-compatible results.
• High Parse Success: Achieves 100% parse success, ensuring all generated Mermaid code is syntactically valid.
• Optimized for Medium Complexity: Performs best on flowcharts with 10-20 nodes, achieving a node F1 score of 0.796.
• Efficient Training: Fine-tuned using LoRA (rank 16, alpha 16) on the DangIT02/flowchart-to-mermaid_v2 dataset, with only 0.50% of parameters trainable.

Limitations & Considerations

• Hallucination on Small Diagrams: Tends to over-generate nodes for flowcharts with fewer than 10 nodes.
• Direction Detection: Direction accuracy can degrade on very large diagrams (20+ nodes), sometimes defaulting to graph TD.
• Canonical IDs Only: Does not reproduce descriptive node IDs from the original diagram; post-processing is required for semantic IDs.
• English Labels Only: Performance with non-English labels is untested.
• Single Image Input: Processes one flowchart image per prompt.
• Output Length: Maximum output is approximately 2048 tokens, which may truncate very large flowcharts.