Dataset · nvidia
SWE agent trajectories for distillation
nvidia/Open-SWE-Traces
Agent rollouts across various frameworks for software engineering tasks, intended for distillation and training SWE agents.
What's inside
This is a collection of agent trajectories for software engineering tasks collected by NVIDIA. Version 2.0 adds rollouts based on Qwen3.8-Flash-Next with varying reasoning effort levels across OpenHands, SWE-agent, mini-swe-agent, Claude-Code, and OpenCode. Previously added Qwen3.8-27B trajectories; metadata indicates a size of 1M–10M and text modality.
What you can do with it
- Fine-tuning an SWE agent or LoRA for agentic tasks
- Distilling the behavior of strong models into smaller ones
- Comparing agent frameworks on identical trajectories
- Size
- 1M–10M rows
- Downloads this month
- 37K
- Likes
- 183
Models trained on it
- nicosuter/Qwen3.8-27B-AWQImage text → text 436.5K
- Jitheswar/Qwen3.5-4B-SalieriImage text → text 1.1K
- mradermacher/Qwen3.5-4B-Salieri-i1-GGUF 1K
- mradermacher/Qwen3.5-4B-Salieri-GGUF 1K
- mradermacher/Mellum2.1-12B-A2.5B-Thinking-DevSec-merged-bf16-GGUF 863
- shadowpa0327/Qwen3.8-Flash-Next-DSpark 489
- ahmetggg/Luck-Qwen3-4b-Code-FineTuneText generation 177
- SousiOmine/Lumi-v2-llm-jp-4-8bText generation 98
How to load it
from datasets import load_dataset
ds = load_dataset("nvidia/Open-SWE-Traces")