Building Supervised Fine-Tuning Data from NVIDIA Open-SWE-Traces: Trajectory Parsing, Patch Analysis, Token Budgets, and Tool-Use Metrics
NVIDIA's Open-SWE-Traces just became the blueprint for fine-tuning agentic models. Here's how to build production-grade SFT datasets in Colab.
Why it matters
Open datasets for agent fine-tuning are becoming the competitive moat. This tutorial shows how to transform NVIDIA's agentic trajectories into high-quality supervised training data, enabling any team to build specialized software-engineering agents without proprietary data.
The key facts
14 to knowNVIDIA Open-SWE-Traces dataset enables agentic software-engineering trajectory analysis
Multi-turn agent conversation normalization for SFT data curation
Token budget constraints used for data filtering and curation
Tool-use metrics tracked across agent trajectories
Resolution outcomes labeled for supervised fine-tuning subset creation
Language distribution analysis across patches
Hugging Face integration for efficient cloud-based processing
Success-label filtering for high-quality training data
NVIDIA Open-SWE-Traces dataset now available for fine-tuning workflows
Focus on agentic software-engineering trajectories and tool-use metrics
Supervised fine-tuning approach using success labels and token budgets
Multi-turn agent conversation normalization and patch analysis included
Streamable from Hugging Face for efficient processing without local downloads
Data curation covers trajectory length, tool usage, patch size, language distribution, and resolution outcomes
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: In this tutorial, we work with NVIDIA's Open-SWE-Traces dataset to study agentic software-engineering trajectories for fine-tuning. We stream the data directly from Hugging Face, so we can process it efficiently in Google Colab without downloading everything locally. We normalize multi-turn agent…