exolabs/swe-bench-pro-hermes-eval-data
收藏资源简介:
该数据集是SWE-Bench Pro Hermes Agent vLLM评估追踪数据集,用于记录在swe-bench-pro编码基准测试中,通过Margin harness(使用OpenCode代理在每案例docker容器中运行)针对自托管vLLM策略的每轮评估工件。数据集捕获了所有模型调用(通过代理角色和服务策略的日志代理),并重塑为矩阵模式。评估奖励(reward)设置为1.0当且仅当案例测试通过。数据集包含多个文件,如rollouts.jsonl、overall_summary.{json,md}、domain_summary.json(按案例类别分类)、all_tasks_usage.{json,csv}、proxy_calls.jsonl(包含缓存/未缓存提示令牌拆分的每请求捕获)、case_id_map.json、DONE.json和telemetry.tar.gz(每案例的agent_requests.jsonl和trajectory.json)。此外,数据集还提供了模型性能指标(如平均奖励、中位数奖励、通过率和零奖励计数)、MTP草案令牌接受率(按位置)和令牌使用情况(代理侧每轮总计)的统计表格。该数据集主要用于评估人工智能代理在软件工程任务中的性能。
This dataset is the SWE-Bench Pro Hermes Agent vLLM Evaluation Tracking Dataset, used to record per-round evaluation artifacts for self-hosted vLLM policies during the SWE-Bench-Pro coding benchmark test, executed via the Margin harness (using the OpenCode Agent running within per-case Docker containers). The dataset captures all model invocations (via log agents for agent roles and service policies) and reshapes them into a matrix format. The evaluation reward is set to 1.0 if and only if the corresponding task case passes all tests. The dataset includes multiple files, such as rollouts.jsonl, overall_summary.{json,md}, domain_summary.json (categorized by task case categories), all_tasks_usage.{json,csv}, proxy_calls.jsonl (captures per-request breakdowns of cached/uncached prompt tokens), case_id_map.json, DONE.json, and telemetry.tar.gz (containing agent_requests.jsonl and trajectory.json for each individual case). Additionally, the dataset provides statistical tables for model performance metrics including average reward, median reward, pass rate, and zero-reward count, MTP draft token acceptance rate (by position), and token usage statistics (total per round on the agent side). This dataset is primarily used to evaluate the performance of AI Agents on software engineering tasks.




