---

name: qwen-2-5-vl description: "Route Qwen2.5-VL JSON extraction and COCO object-detection fine-tuning, inference, conversation formatting, pixel controls, and troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Qwen2.5-VL

Use this sub-skill for Qwen2.5-VL workflows that need Maestro's Qwen-specific training, loading, inference, or detection-format helpers.

Route here when

• You need maestro qwen_2_5_vl train, load_model, save_model, predict, or the Qwen collators.
• You need to format a Qwen chat conversation with system_message, an image turn, and a prefix/suffix pair.
• You need Qwen detection JSON suffixes, min_pixels / max_pixels, or LoRA / QLoRA / none configuration.
• You need a safe config builder or formatter smoke for this model family.
• You need Qwen-specific troubleshooting for prompt formatting, pixel sizing, or qwen-vl-utils compatibility.

Route away when

• The task is about JSONL / COCO dataset layout, Roboflow identifiers, generic metrics, split validation, or loader helpers; use datasets-and-metrics.
• The task is Florence-2 or PaliGemma 2; use the sibling sub-skill for that model.

Defaults and invariants

• Model ID: Qwen/Qwen2.5-VL-3B-Instruct
• Revision: refs/heads/main
• Learning rate: 2e-4
• Optimisation strategies: lora, qlora, none
• Default pixel bounds: min_pixels=256*28*28, max_pixels=1280*28*28
• Max new tokens: 1024
• The source recipes often override min_pixels to 512*28*28 for object detection; keep train, load, inference, and parsing bounds aligned.

Prompt and chat handling

• format_conversation(...) inserts an optional system message first, then the image plus user prefix, then an optional assistant suffix for supervised training.
• processor.apply_chat_template(...) renders that message list into the model prompt string.
• predict(...) adds a generation prompt before generation so the model completes the assistant turn.
• Use the same system_message in training and inference when you want stable output formatting.
• Generic data layout, metrics, and COCO / JSONL wiring live in the sibling dataset skill; this sub-skill owns the Qwen-specific prompt, resize, and detection formatting logic.

Bundled assets

• references/api-reference.md
• references/workflows.md
• references/detection-formats.md
• references/troubleshooting.md
• scripts/smoke_qwen_detection_format.py
• scripts/build_qwen_config.py

Safe checks

• python scripts/smoke_qwen_detection_format.py
• python scripts/build_qwen_config.py --help
• python scripts/build_qwen_config.py --emit cli --dataset /path/to/dataset

Notes

• Keep qwen-vl-utils pinned to 0.0.8 if smart_resize(...) starts failing against newer releases.
• Do not route dataset or metric questions here unless the Qwen-specific prompt or detection format is the real issue.