AIAny
Icon for item

Nemotron-SFT-SWE-v3.5

Provides agentic instruction‑tuning trajectories for software‑engineering tasks, formatted for supervised fine‑tuning and agent training. Contains multi‑file edits, tests, docs and structured agent traces (≈5,115 records, 1.9 GiB). Intended for commercial use; licensed CC‑BY 4.0 with additional permissive licenses.

Introduction

Most code-focused SFT blends target single-shot repairs or test generation; this dataset emphasizes agentic, multi-step workflows and repository-aware edits so models learn planning, tool use, and cross-file reasoning. It packages structured agent traces collected with the OpenCode harness and curated for supervised fine‑tuning of software-engineering agents.

What Sets It Apart
  • Agentic trajectories rather than isolated Q&A: includes stepwise agent actions, tool calls and patch-style edits so fine‑tuned models can learn multi-step workflows and intermediate state management (so what: better support for agents that must plan and execute sequences across files).
  • Multi-artifact, multi-file focus: examples include source, tests, docs and config changes together, not just single-file patches (so what: improves models’ repository-aware reasoning and regression‑free edits).
  • Compact, curated SFT target: ~5,115 records (1.9 GiB) designed for supervised fine‑tuning rather than large‑scale pretraining (so what: faster iteration for teams training moderate-size models or distillation pipelines).
  • Clear commercial-use signal and mixed permissive licensing: distributed under CC‑BY 4.0 with additional permissive licenses noted (so what: suitable for product integration but check organiational legal requirements).
Who It's For and Tradeoffs

Great fit if you are training or distilling LLMs to act as autonomous coding agents, need cross-file repair/test-generation examples, or want agent traces that reflect tool use and stepwise reasoning. Look elsewhere if you need very large-scale SWE corpora (tens or hundreds of thousands of examples) for pretraining, raw repository snapshots, or if your project policies prohibit the dataset's commercial-use terms. In short: efficient SFT material for repository-aware agent behavior, with explicit tradeoffs around scale and licensing.

More Items

Hugging Face

Provides about 10,000 OpenUSD objects for robotics simulation, including articulated and static assets with physics properties and robot-interaction annotations.

Hugging Face

Pairs metric-scale 3D scans of real environments with egocentric task recordings, estimated body and hand motion, and hierarchical annotations for physical AI research. The first release contains 126 scenes; the dataset is licensed CC BY-NC 4.0.

Turns LLMs’ physical knowledge into video training data using simulator constraints, physical verification, and iterative correction. It also repairs physical errors in generated videos and releases 100K videos spanning 2,500 physical events.