opus-gpt-swe-frontier-core

<div align="center"> # SWE Base ### Repository-level software engineering trajectories for training coding agents. **2,459 chat trajectories** · **48,499 API calls** · **$837.57 recorded generation cost** `SWE-bench` · `debugging` · `patching` · `tools` · `agents` </div> --- ## Overview SWE Base is a software-engineering dataset centered on real repository issues. Each training example gives an agent a problem statement and captures the multi-turn process of inspecting a codebase, reasoning about a bug, using development tools, editing files, running tests, and working toward a patch. The collection is designed for supervised fine-tuning and research on coding agents that need to operate across complete repositories rather than answer isolated programming questions. ## What's inside | Capability | Examples include | |---|---| | Repository exploration | Searching files, reading code, and locating relevant components | | Bug diagnosis | Connecting issue reports to implementation behavior | | Patch generation | Making targeted changes across one or more files | | Tool use | Shell commands, test runners, and iterative environment interaction | | Verification | Running tests and interpreting failures or regressions | | Long-horizon reasoning | Multi-step trajectories from issue description to attempted resolution | Tasks span widely used open-source Python projects represented in SWE-bench, including Django, SymPy, scikit-learn, Matplotlib, pytest, Sphinx, Astropy, and others. Additional SWE-bench Pro artifacts broaden coverage to larger and more varied repositories. ## Training-ready chat data | Config | Rows | Resolved | Resolution rate | API calls | Recorded cost | |---|---:|---:|---:|---:|---:| | `small` | 500 | 371 | 74.20% | 8,134 | $129.19 | | `medium` | 1,959 | 855 | 43.64% | 40,365 | $708.38 | | **Total** | **2,459** | **1,226** | **49.86%** | **48,499** | **$837.57** | Resolution statistics are taken from the stored `resolved` field. They describe these recorded runs and should not be interpreted as official model benchmark scores. ## Format The `small` and `medium` configurations contain one conversation per JSONL row: ```json { "instance_id": "project__repository-12345", "problem_statement": "Repository issue description...", "messages": [ {"role": "system", "content": "Agent instructions..."}, {"role": "user", "content": "Issue and environment state..."}, {"role": "assistant", "content": "Analysis and tool action..."} ], "model": "generator identifier", "resolved": true, "instance_cost": 0.12, "api_calls": 14 } ``` | Field | Type | Description | |---|---|---| | `instance_id` | `string` | SWE-bench task identifier | | `problem_statement` | `string` | Repository issue to solve | | `messages` | `list` | Ordered agent/environment conversation | | `model` | `string` | Recorded generator identifier | | `resolved` | `bool` | Whether the stored run resolved the task | | `instance_cost` | `float` | Recorded API cost for the run | | `api_calls` | `int` | Number of API calls in the run | Environment observations can appear as `user` messages, so the role sequence may differ from a conventional two-party chat transcript. ## Load with 🤗 Datasets After uploading this folder to the Hub: ```python from datasets import load_dataset small = load_dataset("YOUR_USERNAME/swe-base", "small", split="train") medium = load_dataset("YOUR_USERNAME/swe-base", "medium", split="train") print(medium[0]["instance_id"]) print(medium[0]["messages"]) ``` Load the local files directly: ```python from datasets import load_dataset dataset = load_dataset( "json", data_files="gpt5.2swe-medium.jsonl", split="train", ) ``` ## Raw trajectory artifacts The repository also contains raw SWE-bench Verified and SWE-bench Pro artifacts from multiple agent systems: | File | Artifact rows | |---|---:| | `claude-sonnet-4-5_swebench_pro_traj.jsonl` | 733 | | `claude-sonnet-4-5_swebench_verified_traj.jsonl` | 502 | | `gemini_3_pro_swebench_verified_traj.jsonl` | 502 | | `gpt-5_swebench_verified_traj.jsonl` | 502 | | `gpt-5-2_swebench_verified_traj.jsonl` | 501 | | `gpt-5-mini_swebench_verified_traj.jsonl` | 502 | | `kimi-k2-thinking_swebench_verified_traj.jsonl` | 501 | These are preservation-oriented artifacts, not normalized chat splits. Rows may contain per-instance trajectories, patches, evaluation outputs, or aggregate result maps. Inspect and normalize their structures before using them for training. Model names in filenames identify the recorded trajectory generators. No model weights are included. ## Intended use SWE Base is suitable for research and experimentation involving: - supervised fine-tuning of repository-level coding agents; - tool-use and terminal-action prediction; - debugging and patch-generation research; - resolved-versus-unresolved trajectory analysis; - process supervision and long-context agent behavior; - trajectory filtering, ranking, and curriculum design. ## Limitations - A `resolved` label does not guarantee that every intermediate action is optimal or desirable. - Unresolved trajectories may contain useful negative examples but require intentional handling during training. - Agent messages can include verbose reasoning, failed commands, repeated attempts, and environment-specific paths. - Tool and message conventions can vary across trajectory generators. - Recorded costs reflect the stored runs and may not reproduce current API pricing. - Benchmark contamination and overfitting are important risks when training on benchmark trajectories. - No security, secrets, license, or personally identifiable information audit is claimed. Evaluate and filter the data for the target model, chat template, and deployment context. ## License No license is declared in this card. Review the repository terms and the licenses of the represented codebases and benchmark materials before redistribution or commercial use. --- <div align="center"> Built for agents that read the repo, find the bug, make the patch, and run the tests. </div>