Release: Oct 7, 2024

Llama-JPSFT

Supervised fine-tuning was performed on meta-llama/Llama-3.1-8B-Instruct on a select ~140,000 query-response pairs from a diverse corpus of anonymized, scraped chat data, with a priority on casual conversation. BF16 mixed precision training was executed on NVIDIA A100 Tensor Core GPUs and precision reduction/quantization from safetensors to GGUF was then completed for q8_0, q6_k, and q4_k_m models. This project targets standard SFT as well as instruction-tuning for GPT-based architectures to generate significantly improved coherent and context-aware responses in multi-speaker conversations in casual Japanese.

約14万件の多様な匿名化されたチャットデータを基に、教師あり学習による微調整が実施されました。

Instruct Model: https://huggingface.co/ai-net/Llama-JPSFT-2.0

Precision Reduction and Quantization: https://huggingface.co/ai-net/Llama-JPSFT-2.0-GGUF

例えば:

{{user}}
軽率に話しかけてくれる人が増えて嬉しいです！

llama-jpsft-2.0-q4_k_m.gguf
それはいいことだね