Cross-Language Migration Implementation Program
To realize the extension of the model's multilingual capabilities, it can be advanced in three phases:
- Data preparation::
- Constructing a parallel corpus (combinations of Chinese/English/Chinese/Japanese etc. are recommended)
- exist
data/Catalog New Creationmultilingual.jsonThe field containslanguage_tag
- blended training::
- Keep the original model word list and add it with SFT scripts
--lang_loss_weight 0.3parameters - Recommended mixed multilingual samples within batch (supported by project dataloader)
- Keep the original model word list and add it with SFT scripts
- capability testing::
- Specify during interaction testing
--language enParameters such as switching language - Quantitative assessment using indicators such as BLEU
- Specify during interaction testing
Note: Smaller size models (1.7B) are recommended to focus on single language pairs, while models above 4B can try joint multi-language training.
This answer comes from the articleQwen3-FineTuning-Playground: a ready-to-use code base for fine-tuning Qwen3's big models.The
May not be reproduced without permission:AI productivity tools " How to achieve capability migration for Qwen3 fine-tuning models in cross-language scenarios?
