13274243a0
Keep K3 suite selection and report-schema scoring in bash, merge K3/vision dataset_args into dpv4 yamls, and pin EvalScope at 735d920ee911 with local patches. Co-authored-by: Cursor <cursoragent@cursor.com>
67 lines
9.1 KiB
JSON
67 lines
9.1 KiB
JSON
{
|
||
"meta": {
|
||
"pretty_name": "BrowseComp",
|
||
"dataset_id": "evalscope/browse_comp",
|
||
"paper_url": "https://arxiv.org/abs/2504.12516",
|
||
"tags": [
|
||
"Agent",
|
||
"Knowledge",
|
||
"QA"
|
||
],
|
||
"metrics": [
|
||
"is_correct",
|
||
"is_incorrect"
|
||
],
|
||
"primary_metric": {
|
||
"name": "is_correct",
|
||
"aggregation": null,
|
||
"dimensions": {}
|
||
},
|
||
"few_shot_num": 0,
|
||
"eval_split": "test",
|
||
"train_split": "",
|
||
"subset_list": [
|
||
"default"
|
||
],
|
||
"description": "\n## Overview\n\nBrowseComp is an OpenAI benchmark for evaluating browsing and search agents. It contains 1,266 hard-to-find, fact-seeking questions with short, verifiable answers. EvalScope loads the mirrored dataset from ModelScope (`evalscope/browse_comp`).\n\n## Task Description\n\n- **Task Type**: Search-agent factual question answering\n- **Input**: Challenging natural-language question that generally requires persistent web browsing\n- **Output**: Explanation, exact answer, and confidence\n- **Grading**: LLM judge compares the final answer against the reference answer\n\n## Key Features\n\n- Tests persistence, creative search, and multi-hop evidence gathering\n- Uses short answers to keep grading tractable\n- Official data is distributed as encrypted CSV rows and decrypted at evaluation time\n- Classified as an Agent benchmark and compatible with EvalScope agent loop modes\n- Supports single-turn model evaluation by default and native/external agent execution when `TaskConfig.agent_config` is provided\n\n## Evaluation Notes\n\n- Default evaluation loads `evalscope/browse_comp` from ModelScope through the standard EvalScope dataset loader.\n- Use `TaskConfig.agent_config` to evaluate BrowseComp with EvalScope agent loop capabilities such as native tool-use or external agent runners.\n- The primary metric is `is_correct`; `is_incorrect` is also reported.\n- LLM judge is enabled by default. `JudgeStrategy.RULE` falls back to normalized exact match.\n",
|
||
"prompt_template": "{question}\n\nYour response should be in the following format:\nExplanation: {{your explanation for your final answer}}\nExact Answer: {{your succinct, final answer}}\nConfidence: {{your confidence score between 0% and 100% for your answer}}",
|
||
"system_prompt": "",
|
||
"few_shot_prompt_template": "",
|
||
"aggregation": "mean",
|
||
"extra_params": {},
|
||
"sandbox_config": {},
|
||
"category": "agent"
|
||
},
|
||
"statistics": {
|
||
"total_samples": 1266,
|
||
"subset_stats": [
|
||
{
|
||
"name": "default",
|
||
"sample_count": 1266,
|
||
"prompt_length_mean": 811.02,
|
||
"prompt_length_min": 424,
|
||
"prompt_length_max": 2219,
|
||
"prompt_length_std": 226.87,
|
||
"target_length_mean": 17.46
|
||
}
|
||
],
|
||
"prompt_length": {
|
||
"mean": 811.02,
|
||
"min": 424,
|
||
"max": 2219,
|
||
"std": 226.87
|
||
},
|
||
"target_length_mean": 17.46,
|
||
"computed_at": "2026-06-17T22:07:18.747363"
|
||
},
|
||
"sample_example": {},
|
||
"readme": {
|
||
"en": "# BrowseComp\n\n\n## Overview\n\nBrowseComp is an OpenAI benchmark for evaluating browsing and search agents. It contains 1,266 hard-to-find, fact-seeking questions with short, verifiable answers. EvalScope loads the mirrored dataset from ModelScope (`evalscope/browse_comp`).\n\n## Task Description\n\n- **Task Type**: Search-agent factual question answering\n- **Input**: Challenging natural-language question that generally requires persistent web browsing\n- **Output**: Explanation, exact answer, and confidence\n- **Grading**: LLM judge compares the final answer against the reference answer\n\n## Key Features\n\n- Tests persistence, creative search, and multi-hop evidence gathering\n- Uses short answers to keep grading tractable\n- Official data is distributed as encrypted CSV rows and decrypted at evaluation time\n- Classified as an Agent benchmark and compatible with EvalScope agent loop modes\n- Supports single-turn model evaluation by default and native/external agent execution when `TaskConfig.agent_config` is provided\n\n## Evaluation Notes\n\n- Default evaluation loads `evalscope/browse_comp` from ModelScope through the standard EvalScope dataset loader.\n- Use `TaskConfig.agent_config` to evaluate BrowseComp with EvalScope agent loop capabilities such as native tool-use or external agent runners.\n- The primary metric is `is_correct`; `is_incorrect` is also reported.\n- LLM judge is enabled by default. `JudgeStrategy.RULE` falls back to normalized exact match.\n\n\n## Properties\n\n| Property | Value |\n|----------|-------|\n| **Benchmark Name** | `browsecomp` |\n| **Dataset ID** | [evalscope/browse_comp](https://modelscope.cn/datasets/evalscope/browse_comp/summary) |\n| **Paper** | [Paper](https://arxiv.org/abs/2504.12516) |\n| **Tags** | `Agent`, `Knowledge`, `QA` |\n| **Metrics** | `is_correct`, `is_incorrect` |\n| **Default Shots** | 0-shot |\n| **Evaluation Split** | `test` |\n\n\n## Data Statistics\n\n| Metric | Value |\n|--------|-------|\n| Total Samples | 1,266 |\n| Prompt Length (Mean) | 811.02 chars |\n| Prompt Length (Min/Max) | 424 / 2219 chars |\n\n## Sample Example\n\n*Sample example not available.*\n\n## Prompt Template\n\n**Prompt Template:**\n```text\n{question}\n\nYour response should be in the following format:\nExplanation: {{your explanation for your final answer}}\nExact Answer: {{your succinct, final answer}}\nConfidence: {{your confidence score between 0% and 100% for your answer}}\n```\n\n## Usage\n\n### Using CLI\n\n```bash\nevalscope eval \\\n --model YOUR_MODEL \\\n --api-url OPENAI_API_COMPAT_URL \\\n --api-key EMPTY_TOKEN \\\n --datasets browsecomp \\\n --limit 10 # Remove this line for formal evaluation\n```\n\n### Using Python\n\n```python\nfrom evalscope import run_task\nfrom evalscope.config import TaskConfig\n\ntask_cfg = TaskConfig(\n model='YOUR_MODEL',\n api_url='OPENAI_API_COMPAT_URL',\n api_key='EMPTY_TOKEN',\n datasets=['browsecomp'],\n limit=10, # Remove this line for formal evaluation\n)\n\nrun_task(task_cfg=task_cfg)\n```\n\n\n",
|
||
"zh": "# BrowseComp\n\n\n## 概述\n\nBrowseComp 是 OpenAI 推出的一项用于评估浏览和搜索智能体的基准测试。它包含 1,266 个难以查找、以事实为导向的问题，每个问题都有简短且可验证的答案。EvalScope 从 ModelScope 加载该数据集的镜像版本（`evalscope/browse_comp`）。\n\n## 任务描述\n\n- **任务类型**：搜索智能体的事实型问答\n- **输入**：具有挑战性的自然语言问题，通常需要持续的网页浏览才能解答\n- **输出**：解释、精确答案和置信度\n- **评分方式**：由大语言模型（LLM）裁判将最终答案与参考答案进行比对\n\n## 主要特性\n\n- 测试智能体的持久性、创造性搜索能力以及多跳证据收集能力\n- 使用简短答案以简化评分过程\n- 官方数据以加密 CSV 行的形式分发，并在评估时解密\n- 被归类为智能体（Agent）基准测试，兼容 EvalScope 的智能体循环模式\n- 默认支持单轮模型评估；当提供 `TaskConfig.agent_config` 时，也支持原生或外部智能体执行\n\n## 评估说明\n\n- 默认评估通过标准 EvalScope 数据集加载器从 ModelScope 加载 `evalscope/browse_comp`。\n- 使用 `TaskConfig.agent_config` 可启用 EvalScope 智能体循环功能（如原生工具调用或外部智能体运行器）来评估 BrowseComp。\n- 主要指标为 `is_correct`；同时也会报告 `is_incorrect`。\n- 默认启用 LLM 裁判；若未启用，则回退到 `JudgeStrategy.RULE`，即归一化精确匹配。\n\n## 属性\n\n| 属性 | 值 |\n|----------|-------|\n| **基准测试名称** | `browsecomp` |\n| **数据集 ID** | [evalscope/browse_comp](https://modelscope.cn/datasets/evalscope/browse_comp/summary) |\n| **论文** | [Paper](https://arxiv.org/abs/2504.12516) |\n| **标签** | `Agent`, `Knowledge`, `QA` |\n| **指标** | `is_correct`, `is_incorrect` |\n| **默认示例数** | 0-shot |\n| **评估划分** | `test` |\n\n\n## 数据统计\n\n| 指标 | 值 |\n|--------|-------|\n| 总样本数 | 1,266 |\n| 提示词长度（平均） | 811.02 字符 |\n| 提示词长度（最小/最大） | 424 / 2219 字符 |\n\n## 样例示例\n\n*样例示例不可用。*\n\n## 提示模板\n\n**提示模板：**\n```text\n{question}\n\n你的回答应采用以下格式：\nExplanation: {{你对最终答案的解释}}\nExact Answer: {{你简洁明确的最终答案}}\nConfidence: {{你对答案的置信度，介于 0% 到 100% 之间}}\n```\n\n## 使用方法\n\n### 使用 CLI\n\n```bash\nevalscope eval \\\n --model YOUR_MODEL \\\n --api-url OPENAI_API_COMPAT_URL \\\n --api-key EMPTY_TOKEN \\\n --datasets browsecomp \\\n --limit 10 # 正式评估时请删除此行\n```\n\n### 使用 Python\n\n```python\nfrom evalscope import run_task\nfrom evalscope.config import TaskConfig\n\ntask_cfg = TaskConfig(\n model='YOUR_MODEL',\n api_url='OPENAI_API_COMPAT_URL',\n api_key='EMPTY_TOKEN',\n datasets=['browsecomp'],\n limit=10, # 正式评估时请删除此行\n)\n\nrun_task(task_cfg=task_cfg)\n```",
|
||
"content_hash": "346f1a49bd0fd4d124615da5519916e4",
|
||
"needs_translation": false
|
||
},
|
||
"updated_at": "2026-08-11T15:30:53.614759",
|
||
"translation_updated_at": "2026-06-17T22:07:21"
|
||
}
|