Dataset Summary

SWE-rebench-OpenHands-Trajectories is a dataset of multi-turn agent trajectories for software engineering tasks, collected using Qwen/Qwen3-Coder-480B-A35B-Instruct with OpenHands (v0.54.0) agent scaffolding. This dataset captures complete agent execution traces as they attempt to resolve real GitHub issues from nebius/SWE-rebench. Each trajectory contains the agent's step-by-step reasoning, actions, and environmental observations.

Table 1: Comparison of statistics across different datasets containing multi-turn trajectories of agent’s interactions with executable SWE environments.

Note: Entries marked with asterisk (*) indicate statistics whose values couldn't be derived from available data.

For more details see our report in Nebius blog.

---

How to use

import json

from datasets import load_dataset


ds = load_dataset("nebius/SWE-rebench-openhands-trajectories", split="train")

role2field_names = {
 "system": ["role", "content"],
 "assistant": ["role", "content", "tool_calls"],
 "user": ["role", "content"],
 "tool": ["role", "content", "name", "tool_call_id"],
}

def filter_and_deserialize(row):
    trajectory = []
    for msg in row["trajectory"]:
        msg = {
            field_name: msg[field_name] for field_name in role2field_names[msg["role"]]
        }
        if (msg["role"] == "assistant") and (msg["tool_calls"] is not None):
            for i, tool_call in enumerate(msg["tool_calls"]):
                if "arguments" in tool_call.get("function", {}):
                    msg["tool_calls"][i]["function"]["arguments"] = json.loads(
                        tool_call["function"]["arguments"]
                    )
        trajectory.append(msg)
    return row | {"trajectory": trajectory}

first_trajectory = filter_and_deserialize(ds[0])["trajectory"]

for msg in first_trajectory:
    print(msg)

---

Dataset Structure

Each row contains the following information about trajectory:

Table 2: Dataset field descriptions and schema.

To our knowledge, no other publicly available agent trajectory dataset includes evaluation of agent-generated tests.

Important Note: arguments field inside tool calls present in assistant steps of trajectory is serialized to string format for storage efficiency. When training on this data, you may want to deserialize it first to ensure chat templates apply the same formatting during training and inference.

---

Citation

@article{trofimova2025openhandstrajs,
  title={OpenHands Trajectories with Qwen3-Coder-480B-A35B-Instruct},
  author={Trofimova, Maria and Shevtsov, Anton and Ibragim, Badertdinov and Pyaev, Konstantin and Karasik, Simon and Golubev, Alexander},
  year={2025},
  journal={Nebius blog},
  note={}
}