ynklab/Qwen2.5-7B-d2d
ynklab/Qwen2.5-7B-d2d is a 7.6 billion parameter language model, fine-tuned from Qwen/Qwen2.5-7B-Instruct, specifically designed for multilingual document-level machine translation. It excels at translating entire documents between English and nine other languages, including German, Spanish, French, and Chinese, supporting bidirectional translation. This model is optimized for length-consistent document translation using a fixed-range chunking approach, making it suitable for applications requiring accurate and coherent long-form text translation.
Loading preview...
Overview
ynklab/Qwen2.5-7B-d2d is a 7.6 billion parameter model, fine-tuned from Qwen/Qwen2.5-7B-Instruct, for multilingual document-level machine translation. It was developed as part of the research for the paper "Doc2FRC: Length-Consistent Document-Level Machine Translation via Fixed-Range Chunking" and is trained directly on the document-to-document (d2d) translation task.
Key Capabilities
- Document-Level Translation: Optimized for translating entire documents, ensuring length consistency and coherence.
- Multilingual Support: Supports bidirectional translation between English and nine other languages: German, Spanish, French, Italian, Korean, Dutch, Portuguese, Russian, and Chinese.
- High Context Length: Capable of handling up to 32,768 tokens, allowing for translation of substantial document segments.
Training Details
The model underwent full-parameter supervised fine-tuning on the sardinelab/DocBlocks dataset for 2 epochs. It utilized a maximum sequence length of 32,768 tokens and bfloat16 precision.
Usage
Users should refer to the Doc2FRC GitHub repository for exact inference scripts, prompting setup, and evaluation procedures. The recommended prompt format follows a ChatML-style structure, specifying source and target languages and text.
License
This model is released under the Apache License 2.0. Users should also consult the DocBlocks dataset and original data sources for their respective licensing conditions.