Support numind/NuExtract3

[ariefwijaya]

Checklist / 检查清单

• I have searched existing issues, and this is a new feature request. / 我已经搜索过现有的 issues，确认这是一个新的 Feature Request。

Feature Request Description / Feature Request 描述

Checklist / 检查清单

Feature Request Description / Feature Request 描述

Please add support for fine-tuning NuExtract3 from NuMind.

NuExtract3 is a 4B vision-language model based on Qwen3.5-4B (qwen3_5 architecture) specialized for structured information extraction and document-to-Markdown conversion. It supports text, image, and multimodal inputs, and is useful for SFT on document understanding tasks (invoices, receipts, forms, contracts, etc.).

• ModelScope: https://www.modelscope.ai/models/numind/NuExtract3
• HuggingFace: https://huggingface.co/numind/NuExtract3

Pull Request / Pull Request 信息

No response

[hiSandog]

For NuExtract3 support, the implementation scope probably needs to cover more than registering the model name. Since it is based on a Qwen3.5-style VLM but used for structured extraction/document-to-Markdown tasks, good acceptance tests would include processor loading, image+text SFT sample formatting, multimodal collator behavior, and a tiny inference/export smoke test. That would make it clear whether support is only for loading weights or for the full fine-tuning path.