JetBrains-Research/agent-trajectories-swe-bench-test-minus-verified
收藏资源简介:
--- license: mit task_categories: - text-generation language: - en tags: - swe-bench - code - agent-trajectories - mixed-teachers pretty_name: Agent Trajectories on SWE-bench Test \ Verified (Mixed Teachers) size_categories: - 1K<n<10K configs: - config_name: default data_files: - split: train path: data/train-* dataset_info: features: - name: messages list: - name: content dtype: string - name: role dtype: string - name: instance_id dtype: string - name: n_turns dtype: int64 - name: n_messages dtype: int64 - name: selected_models sequence: string - name: resolved dtype: 'null' - name: exit_status dtype: string splits: - name: train num_bytes: 102279077 num_examples: 1785 download_size: 40000657 dataset_size: 102279077 --- # Agent Trajectories: SWE-bench Test \ Verified — Mixed Teachers (gpt-5.2 / gpt-5-mini) ## Summary Full multi-turn agent trajectories collected from the **SWE-bench Test minus Verified** split (i.e., SWE-bench Test instances that are **not** part of SWE-bench Verified). Intended for SFT of agent models on coding tasks. ## Data Collection Each trajectory was produced by a **GT-aware lookahead agent** that, at every turn: 1. Sampled a candidate response from **both** `gpt-5.2` and `gpt-5-mini`. 2. Had a **gemini-3-pro-preview** router evaluate both candidates and select the better one. 3. Executed the selected step in the environment and appended the observation. This "best-of-two" selection is oracle-informed (access to the ground-truth patch during routing), producing near-optimal mixed-teacher trajectories. ## Format Each row contains a single complete trajectory in **OpenAI messages format**: ```json { "messages": [ {"role": "system", "content": "You are a helpful assistant..."}, {"role": "user", "content": "<pr_description>...</pr_description>"}, {"role": "assistant", "content": "THOUGHT: ...\nACTION: ..."}, {"role": "user", "content": "<returncode>0</returncode>..."}, "...", {"role": "assistant", "content": "...submit..."} ], "instance_id": "sympy__sympy-24152", "n_turns": 8, "n_messages": 17, "selected_models": ["litellm_proxy/openai/gpt-5.2", "litellm_proxy/openai/gpt-5-mini", "..."], "resolved": true, "exit_status": "Submitted" } ``` ## Statistics | Split | Instances | |------------|-----------| | train | 1607 | | validation | 178 | | **total** | **1785** | ## Source - **SWE-bench subset**: SWE-bench Test \ Verified (instances in Test that are not in Verified) - **Candidate models**: `gpt-5.2`, `gpt-5-mini` - **Router (step selector)**: `gemini-3-pro-preview` (GT-aware lookahead routing) - **Langfuse session**: `data-collection-run_id-2eb77197e7c72634d36b53890e775004-gt_aware_llm-2026-03-13_11-14-52`




