laion/sera-subset-mixed-316
收藏Hugging Face2026-04-29 更新2026-05-03 收录
下载链接：
https://hf-mirror.com/datasets/laion/sera-subset-mixed-316
下载链接
链接失效反馈官方服务：
资源简介：
`sera-subset-mixed-316`是从`ethanlshen/sera-subset`中随机抽取的316行数据，混合了两个上游阶段（stage1未解决和stage2已解决）的数据并进行了确定性洗牌。数据集每行包含`messages`（角色、内容、训练标记）和`instance_id`，其中训练标记仅在助手回合为`True`。Hermes的`<tool_call>`/`<tool_response>`令牌已预先渲染到内容中。数据集用于`laion/sera-subset-mixed-316-axolotl__Qwen3-8B-v8`（基于Qwen3-8B的SFT）。
`sera-subset-mixed-316` is a random subset of 316 rows drawn from `ethanlshen/sera-subset`, mixed across the two upstream stages (stage1 unresolved + stage2 resolved) and shuffled deterministically. Each row is JSON with `messages` (list of {role, content, train}) and `instance_id`. The training mask is `train: True` only on assistant turns. Hermes `<tool_call>`/`<tool_response>` tokens are pre-rendered into content. Used by `laion/sera-subset-mixed-316-axolotl__Qwen3-8B-v8` (SFT on Qwen3-8B base).
提供机构：
laion
