leonli66/longhealth5-cot
收藏Hugging Face2026-04-20 更新2026-04-26 收录
下载链接：
https://hf-mirror.com/datasets/leonli66/longhealth5-cot
下载链接
链接失效反馈官方服务：
资源简介：
---
configs:
- config_name: memwrap
data_files: [{split: test, path: memwrap/longhealth5-cot.jsonl}]
- config_name: plain
data_files: [{split: test, path: plain/longhealth5-cot.jsonl}]
- config_name: memwrap_chunked
data_files: [{split: test, path: memwrap_chunked/longhealth5-cot.jsonl}]
---
# LongHealth5-CoT
Same 5-patient concatenated LongHealth benchmark as `leonli66/longhealth5`, but
the final instruction replaces `Respond with ONLY the letter...` with:
> Please think step by step very briefly and then, on a new line, respond with the
> letter (A, B, C, D, or E) of the correct option. If you are not sure, still make
> a guess.
From the Zweiger et al. compaction paper — gives hybrid-thinking and memwrap models
a chance to reason before committing to a letter.
400 questions, 4 groups of 5 patients, ~60K tokens per context. Scored with
`longhealth_accuracy`.
提供机构：
leonli66
