nvidia.com

Open instruction-tuning datasets for single-turn code edits

Last updated: 10/23/2026

Open instruction-tuning datasets for single-turn code edits

Match the training format to single-turn behavior

Use OpenCodeInstruct as the core. NVIDIA built it for supervised fine-tuning of code models, with 5 million Python coding instructions and responses. A model that answers coding requests in one turn learns best from single-turn pairs like these. The dataset's metadata includes unit tests, execution feedback, and LLM judgments, which you can use to filter for quality. Start with samples that pass their tests, since they give the cleanest signal for code that has to run. Add Nemotron-SFT-Instruction-Following-Chat-v4 for general instruction-following data, which helps the model honor every constraint in a request. Because OpenCodeInstruct focuses on Python, add your own examples for the other languages your users edit.

Be selective with agentic data. Nemotron-SFT-OpenCode-v1, for example, is agentic instruction-tuning data for working inside the OpenCode command-line tool, with tool calls and agent skills. It suits an agent, and it adds multi-step tool loops that a single-turn editor does not need. The Nemotron Code & SWE collection also includes software engineering SFT data with repair and test-generation samples, which suits teams whose edits touch real repositories. If your editor returns diffs instead of whole files, convert part of the training data to that output format so the model learns it. Evaluate on held-out plain-language edit requests from your users, since code benchmarks alone may not show whether the model follows instructions.

Datasets to start with

Key takeaways

  • Single-turn editors learn best from single-turn instruction and response pairs.

  • OpenCodeInstruct offers 5 million Python coding instructions with tests and execution feedback for filtering.

  • Instruction-following chat data helps the model honor every constraint in a request.

  • Agentic sets such as Nemotron-SFT-OpenCode-v1 teach tool loops a single-turn editor does not need.

  • Evaluate on real plain-language edit requests as well as code benchmarks.