Step-by-step guide to multilingual fine-tuning
The following steps are required to realize multilingual reasoning:
- Data preparation: Load HuggingFace multilingual dataset (
load_dataset('HuggingFaceH4/Multilingual-Thinking')), the dataset contains English/Spanish/French language samples - LoRA Configuration: Settings
LoraConfig(r=8, lora_alpha=32)Specify adapter parameters to focus on tuningq_projcap (a poem)v_projplane of projection (in perspective drawing) - Model loading: Use
PeftModelPacking the original model, keeping the 95% parameters frozen and only fine-tuning the adaptation layer - Training control: Set by TRL library
max_seq_length=2048cap (a poem)batch_size=4The use of gradient checkpoints saves video memory. - Language Designation: add to the system prompt when reasoning
'Reasoning language: Spanish'isometric instruction
A full example can be found in the repositoryfinetune.ipynb, the entire process takes about 6 hours on a single 24GB GPU.
This answer comes from the articleCollection of scripts and tutorials for fine-tuning OpenAI GPT OSS modelsThe
May not be reproduced without permission:AI productivity tools " How to achieve fine-tuning of multilingual reasoning capabilities for the gpt-oss-20b model?
