AlienKevin/terminal-bench-2-sft-traces

Terminal-Bench 2.0监督微调轨迹比较数据集包含三个不同检查点（Marin/Qwen3 SFT）在Terminal-Bench 2.0基准测试中的每个试验的代理轨迹。该数据集用于调试Marin-32B-base SFT中的70%退化故障模式，并比较不同检查点在相同评估设置下的表现。数据集包含三个分割：marin_step858、qwen3_step1000和qwen3_step500，分别对应不同的检查点和训练步骤。每个试验行包含模型名称、任务名称、是否解决、奖励、输入/输出令牌数等字段，以及详细的聊天格式对话记录和验证器输出。

The Terminal-Bench 2.0 SFT Trajectory Comparison dataset contains per-trial agent trajectories from running the Terminal-Bench 2.0 benchmark against three supervised fine-tuned 32B-parameter checkpoints. It was generated to debug a 70%-degeneration failure mode in `marin-32b-base` SFT and to compare the performance of different checkpoints under the same evaluation setup. The dataset includes three splits: marin_step858, qwen3_step1000, and qwen3_step500, corresponding to different checkpoints and training steps. Each trial row contains fields such as model name, task name, whether it was solved, reward, input/output tokens, and detailed chat-format conversation records and verifier output.