exolabs/terminal-bench-2-hermes-eval-data
收藏资源简介:
该数据集包含Terminal-Bench 2.0编码基准的每运行评估工件,通过Margin harness(在每案例docker容器中的OpenCode代理)针对自托管的vLLM策略运行生成。每个模型调用都被捕获并重塑为矩阵模式,其中reward = 1.0当且仅当案例的测试通过。avg_reward / pass@1仅基于SCORED(模型结果)行计算;基础设施/实现失败被隔离(从率中排除,在infra_failures下跟踪),因此运行仅在complete时计入头条。每个运行目录包含rollouts.jsonl、overall_summary.{json,md}、domain_summary.json(按案例类别)、all_tasks_usage.{json,csv}、proxy_calls.jsonl(每请求捕获,包括缓存/未缓存的提示令牌分割)、case_id_map.json、DONE.json和telemetry.tar.gz(每案例的agent_requests.jsonl + trajectory.json)。数据集主要用于评估模型在终端编码任务中的性能,包括奖励、通过率和令牌使用情况。
This dataset contains per-run evaluation artifacts of the Terminal-Bench 2.0 coding benchmark, generated via the Margin harness (the OpenCode agent operating within per-case Docker containers) against self-hosted vLLM policies. Every model invocation is captured and reshaped into a matrix format, where reward = 1.0 if and only if the test case passes. avg_reward and pass@1 are calculated exclusively based on the SCORED (model outcome) rows; infrastructure/implementation failures are isolated (excluded from the evaluation metrics and tracked under infra_failures), so runs are only counted toward headline metrics when they are complete. Each run directory contains rollouts.jsonl, overall_summary.{json,md}, domain_summary.json (grouped by case category), all_tasks_usage.{json,csv}, proxy_calls.jsonl (captured per request, including cached/uncached prompt token splits), case_id_map.json, DONE.json, and telemetry.tar.gz, which holds agent_requests.jsonl and trajectory.json for each individual case. This dataset is primarily used to evaluate model performance on terminal coding tasks, including rewards, pass rates, and token usage.




