Building Supervised Fine-Tuning Data from NVIDIA Open-SWE-Traces: Trajectory Parsing, Patch Analysis, Token Budgets, and Tool-Use Metrics

Short summary

Learn to build supervised fine-tuning datasets from NVIDIA's Open-SWE-Traces by streaming from Hugging Face, normalizing agent trajectories, parsing code patches, and filtering by success. The tutorial covers trajectory metrics, tool usage analysis, and dataset curation for training agentic software-engineering models with proper token budgets and language filtering.

• •Stream NVIDIA's Open-SWE-Traces dataset from Hugging Face without local downloads
• •Parse code patches and normalize agent trajectories for supervised fine-tuning
• •Filter datasets by success labels, token budgets, and language to optimize training

Generated with AI, which can make mistakes.