事实抽取、指令遵循、多轮一致性、中文惯用语与结构化摘要的统一评测框架。当前为测试计划，指标填入前保持"未实测"。 A unified framework for extraction, instruction following, multi-turn consistency, idioms and structured summarization. Plan only — cells stay "Not tested" until measured.
固定 5 类任务，每类 4 组提示词，温度 0.2；同一网络环境，24 小时窗口内重复 20 次，取中位数。事实抽取按精确匹配计分，多轮一致性按会话内的自洽度计分。 5 fixed task types x 4 prompts each, temperature 0.2, same network, 20 runs over a 24h window, median reported. Extraction scored by exact match; consistency by intra-session coherence.
| 任务Task | 说明Description |
|---|---|
| 事实抽取 Fact extraction | 从给定段落中抽取实体、时间、数字与关系Extract entities, dates, numbers and relations from a passage |
| 指令遵循 Instruction following | 按约束条件完成改写、摘要与格式要求Rewrite, summarize and format under explicit constraints |
| 多轮一致性 Multi-turn consistency | 10 轮对话中保持事实与立场一致Keep facts and stance consistent across 10 turns |
| 中文惯用语 Idioms | 成语、谚语、口语的准确理解与运用Correct understanding and use of idioms and colloquialisms |
| 结构化摘要 Structured summary | 输出指定结构的 JSON 摘要Produce JSON summaries with a fixed structure |
| Model ID | 事实抽取Extraction | 指令遵循Instruction | 多轮一致性Multi-turn | 惯用语Idioms | 综合Overall |
|---|---|---|---|---|---|
gpt-5.6-sol | 未实测 | 未实测 | 未实测 | 未实测 | 未实测 |
gpt-5.6-luna | 未实测 | 未实测 | 未实测 | 未实测 | 未实测 |
gpt-5.4 | 未实测 | 未实测 | 未实测 | 未实测 | 未实测 |
kimi-k3 | 未实测 | 未实测 | 未实测 | 未实测 | 未实测 |
kimi-k2.7-code | 未实测 | 未实测 | 未实测 | 未实测 | 未实测 |
claude-opus-4-7 | 未实测 | 未实测 | 未实测 | 未实测 | 未实测 |
claude-sonnet-4-6 | 未实测 | 未实测 | 未实测 | 未实测 | 未实测 |
