Dataset · llm-jp
Magpie Japanese Synthetic SFT Dataset
llm-jp/magpie-sft-v1.0
100k to 1M Japanese instruction-response pairs for SFT generated by Calm3 and Qwen2.5 models.
What's inside
Prepared by the LLM-jp Japanese consortium, this dataset contains between 100,000 and 1 million instruction-response text pairs in Japanese. The data was collected entirely synthetically using the Magpie method: prompts were generated by cyberagent/calm3-22b-chat and responses were written by Qwen/Qwen2.5-32B-Instruct.
What you can do with it
- SFT fine-tuning of language models for Japanese language tasks
- Knowledge distillation from Qwen2.5-32B into smaller compact models
- Adapting multilingual LLMs to Japanese instructions and communication style
- Tasks
- Text generation
- Languages
- Japanese
- Size
- 100K–1M rows
- Downloads this month
- 915
- Likes
- 23
Models trained on it
- Kendamarron/Qwen2.5-1.75B-A1.1B-Instruct-jaText generation 636
- hmwri/mlab-LLM-MoE-900MText generation 545
- nagohachi/tiny-lm-v3-japanese-1.5b-a500m-sftText generation 526
- nagohachi/Llama-3.2-tiny-lm-v2-japanese-500m-sftText generation 430
- nagohachi/tiny-lm-v1-japanese-500m-sftText generation 349
- pokutuna/llm2024-competition 219
- shivamjha98/Agnes-8BText generation 56
- tokutsu/japanese-answer-13b-8bit 48
How to load it
from datasets import load_dataset
ds = load_dataset("llm-jp/magpie-sft-v1.0")