nvidia.com

Open datasets for training coding agents on full trajectories

Last updated: 10/23/2026

Open datasets for training coding agents on full trajectories

Train on full agent runs, then reward the runs that pass

Start with Open-SWE-Traces. NVIDIA publishes it as agentic instruction-tuning data for software engineering agents, collected on real repository issues across open agent frameworks including OpenHands, SWE-agent, and mini-swe-agent. Each record keeps the full message history: the agent's reasoning and actions, every tool call, and what the environment returned. It also includes a resolved flag that marks whether the agent fixed the issue, along with the model's patch and the reference patch. The tasks come from permissively licensed sources, and the data is released under CC BY 4.0. NVIDIA also removed trajectories that showed git hacking, a known way agents game software engineering benchmarks.

A diff shows where a change ended up. A trajectory shows how the agent got there, which is what a coding agent has to learn. Fine-tune on resolved runs first, and keep unresolved runs for preference or error-recovery training. Then move to RL, so the reward reflects whether the code passes its tests. Nemotron-RL-Agentic-SWE-Pivot-v1 provides RL data for software engineering agents, and Nemotron-RL-Agentic-Terminal-Pivot-v1, released with Nemotron 3.5 Lightning, covers agents working in a terminal. NVIDIA NeMo Gym provides open environments and verifiers for running those rollouts. For broader tool-use skills, Nemotron-SFT-Agentic-v4 adds 634,664 single-turn and multi-turn tool-use trajectories. Check each card's trajectory format and tool definitions against your agent's harness before you train.

Datasets to start with

Key takeaways

  • Trajectories teach the full loop: plan, act, observe, and verify.

  • Open-SWE-Traces publishes complete agent runs on real repository issues, with a resolved flag on each.

  • Fine-tune on resolved runs first, then run RL against executable tests.

  • NVIDIA's agentic SWE and terminal RL datasets and NeMo Gym support the RL stage.

  • Confirm that the trajectory format and tool definitions match your agent's harness.