菜单
extract structured benchmark information from academic papers when the input includes multiple pdfs, abstracts, or links and the user needs chinese notes or table-ready fields for tasks, datasets, metrics, baselines, sota claims, code release, or evaluation settings.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | benchmark-extractor |
| description | extract structured benchmark information from academic papers when the input includes multiple pdfs, abstracts, or links and the user needs chinese notes or table-ready fields for tasks, datasets, metrics, baselines, sota claims, code release, or evaluation settings. |
用这个 skill 从论文中抽取 benchmark、dataset、metric、baseline、SOTA 等结构化信息，方便后续整理成表格、CSV 或 JSON。
references/extraction_schema.md 统一字段名和缺失值写法。references/comparison_table_template.md 输出中文说明版或结构化表格版。未知，不要臆造。references/extraction_schema.md。references/comparison_table_template.md。summarize machine learning experiment logs in chinese when the input includes training logs, eval results, hyperparameter changes, user notes, or multiple runs and the user needs a grounded experiment summary, error analysis, best configuration recap, or a weekly update ready abstract.
produce a chinese deep reading note for a single academic paper when the input is a pdf, arxiv link, title with abstract, or paper excerpts and the user needs a grounded reading card, contribution summary, reproduction risks, or reading priority judgment.
monitor recent arxiv papers and produce a chinese digest when the user needs a filtered paper watchlist, a ranked update for agent or rag related topics, or an optional feishu webhook push from recent cs.ai or cs.cl submissions.
analyze topic coverage, bottlenecks, controversies, and plausible research gaps when the input includes a research topic, a set of papers, or the user's early ideas and the user needs a grounded gap analysis, problem framing, small entry points, or experiment hypotheses without forced novelty claims.
analyze academic reviews and draft a professional rebuttal when the input includes reviewer comments, a meta-review, paper abstract, or user supplied experiment status and the user needs concern classification, response strategy, risk warnings, or a rebuttal draft without inventing evidence.
write a chinese survey draft around a research topic when the input includes a topic, multiple papers, abstracts, links, pdfs, or a user curated related-work list and the user needs a short survey, long survey, taxonomy, or method evolution summary.