The bottleneck in training coding agents isn’t compute — it’s trajectories: full records of an agent reading an issue, navigating a repo, and shipping a fix. Everyone hoards them. NVIDIA just published 550k+ on Hugging Face, permissively licensed.
What’s in the release
Three software-engineering datasets. Open-SWE-Traces: 207,489 trajectories across nine languages, sourced from 20,000 real-world PRs, collected with the SWE-agent and OpenHands harnesses, synthesized by Minimax-M2.5 (thinking traces) and Qwen3.5-122B (non-thinking). Plus SWE-Zero (318k) and SWE-Hero (34k). All load directly via the Hugging Face datasets library, with open training recipes and an eval framework on GitHub — enough to SFT your own coding agent end to end.
Why NVIDIA gives this away
Labs treat trajectory data as their moat; that’s why open models lag frontier coding agents. NVIDIA doesn’t sell agents — it sells the training runs. More teams fine-tuning coding agents means more GPU demand, so torching everyone else’s data moat is just good business. If you’re building on open models, this is the strongest starting hand you’ve been dealt this year.
You Might Also Like
- Hugging Face Speech to Speech Open Source Local Voice Agents is the Openai Realtime Clone you can run on Your own gpu
- Nvidia Agrees to Acquire Hugging Face for 12 9b the Github of ai Just Picked a Side
- Stripe Coinbase and Ramp Built Internal Coding Agents Langchain Open swe Gives you the Same Architecture for Free
- Agent Reach Hits 12k Github Stars by Solving ai Agents Biggest Blind Spot
- Addy Osmani Open Sources Agent Skills 19 Workflows That Make ai Agents Code Like Google Engineers
Top AI Product 
Leave a comment