vukrosic/blueberry-1B-sft
收藏Hugging Face2025-12-17 更新2025-12-20 收录
下载链接：
https://hf-mirror.com/datasets/vukrosic/blueberry-1B-sft
下载链接
链接失效反馈官方服务：
资源简介：
这是用于**Blueberry-Nano**模型的监督微调（SFT）数据集，包含高质量、遵循指令的对话，专为训练而格式化。数据集采用ChatML格式（`<|im_start|>user...<|im_end|>`），并采用**仅助手预测**的掩码策略，即用户和系统令牌被掩码（标签设置为-100），以便模型仅学习生成响应而非提示。序列被打包为2048个令牌以提高效率。数据来源于**[SmolTalk](https://huggingface.co/datasets/HuggingFaceTB/smoltalk)**的精选子集，包括**Magpie-Ultra**（高质量合成指令）和**Everyday-Conversations**（自然对话）。
This is the Supervised Fine-Tuning (SFT) dataset for the **Blueberry-Nano** model. It contains high-quality instruction-following conversations formatted for training. The dataset uses ChatML format (`<|im_start|>user...<|im_end|>`) and employs **Assistant-Only Prediction** masking, where User and System tokens are masked (labels set to -100) so the model only learns to generate responses, not prompts. Sequences are packed to 2048 tokens for efficiency. The data is a curated subset of **[SmolTalk](https://huggingface.co/datasets/HuggingFaceTB/smoltalk)**, including **Magpie-Ultra** (high quality synthetic instructions) and **Everyday-Conversations** (natural dialogue).
提供机构：
vukrosic
