shisa-v2-sharegpt

# shisa-v2-sharegpt

This is an updated version of the original shisa-v1 dataset [augmxnt/ultra-orca-boros-en-ja-v1](https://huggingface.co/datasets/augmxnt/ultra-orca-boros-en-ja-v1) and retains the same `conversations` field and sharegpt formatting to facilitate its use as drop-in replacement for the original dataset.

The shisa-v2 revision filters a few entries, but largely retains the exact composition and prompts of the original.

- All responses have been entirely regenerated from open weight models ([Athene V2](https://huggingface.co/Nexusflow/Athene-V2-Chat), [Llama 3.3 70B](meta-llama/Llama-3.3-70B-Instruct), and [Tulu 3 405B](https://huggingface.co/allenai/Llama-3.1-Tulu-3-405B))
- Outputs from the models were rated via [UltraFeedback](https://github.com/OpenBMB/UltraFeedback)-inspired scoring and treated as a best-of-n selection
- While non-standard system prompts were retained, default system prompts (which should be usually be masked in training anyway) were removed for improved dataset efficiency and general cleanup

# shisa-v2-sharegpt

本数据集为原始shisa-v1数据集[augmxnt/ultra-orca-boros-en-ja-v1](https://huggingface.co/datasets/augmxnt/ultra-orca-boros-en-ja-v1)的更新版本，保留了与原版一致的`conversations`字段与ShareGPT格式，可作为原版数据集的即插即用替代方案，便于直接复用。

shisa-v2版本对少量条目进行了筛选，但整体保留了原始数据集的核心构成与提示词内容。

- 所有回复均由开源权重模型重新生成，包括[Athene V2](https://huggingface.co/Nexusflow/Athene-V2-Chat)、[Llama 3.3 70B](meta-llama/Llama-3.3-70B-Instruct)以及[Tulu 3 405B](https://huggingface.co/allenai/Llama-3.1-Tulu-3-405B)
- 模型输出通过基于UltraFeedback的评分机制进行筛选，并采用最佳n选（best-of-n）的方式选取最终结果
- 尽管保留了非标准化的系统提示词，但为提升数据集训练效率并完成整体清理，移除了通常在训练中需进行掩码的默认系统提示词。