nemotron-swe-v1

Cleaned version of nvidia/Nemotron-SWE-v1 converted to a uniform OpenAI-compatible tool-calling format.

Source

• Original dataset: nvidia/Nemotron-SWE-v1 (59k SWE-agent trajectories via OpenHands, synthesised with Qwen3-Coder-480B-A35B-Instruct)
• Split: r2e_gym

Schema

Column	Type	Description
messages	JSON string	List of message dicts (roles: system, user, assistant, tool)
tools_json	JSON string	List of tool definitions in OpenAI function-calling format

Message fields

Field	Present on roles	Notes
role	all	system / user / assistant / tool
content	all	null when assistant emits tool calls
reasoning_content	assistant	Chain-of-thought moved here when tool_calls is non-empty
tool_calls	assistant	OpenAI-format list; arguments is a JSON string
tool_call_id	tool	Matches the id in the triggering tool call

Processing Rules

1. id → tool_call_id: The raw dataset stores tool-result identity in a top-level id field; this is renamed to tool_call_id on tool role messages.
2. content / tool_calls mutual exclusivity: When an assistant message has both content (reasoning) and tool_calls, the content is moved to reasoning_content and content is set to null.
3. tools_json is taken from the per-row tools field of the source dataset (each trajectory may have a different tool set).

Statistics

Split	Rows
r2e_gym	51,029

Loading

from datasets import load_dataset
import json

ds = load_dataset("tuandunghcmut/nemotron-swe-v1", split="r2e_gym")
sample = ds[0]

messages = json.loads(sample["messages"])
tools    = json.loads(sample["tools_json"])

for msg in messages[:3]:
    print(msg["role"], "->", str(msg.get("content") or msg.get("reasoning_content", ""))[:80])

Citation

@misc{nemotron_swe_v1,
  title  = {Nemotron-SWE-v1},
  author = {NVIDIA},
  year   = {2025},
  url    = {https://huggingface.co/datasets/nvidia/Nemotron-SWE-v1}
}