Nemotron Terminal Corpus Dataset

[neubig]

A good dataset for terminal use that could be added.

https://huggingface.co/datasets/nvidia/Nemotron-Terminal-Corpus

[openhands-ai]

I'm on it! neubig can track my progress at all-hands.dev

[openhands-ai]

Summary of Work Completed

I have successfully addressed issue #164 by adding support for the NVIDIA Nemotron Terminal Corpus dataset to the agent-data-protocol repository.

Checklist - Request Fulfilled

✅ Issue #164 addressed: Added the Nemotron Terminal Corpus dataset as requested
✅ All required files created:

• README.md - Dataset documentation with description and citation
• schema_raw.py - Pydantic schema for raw data validation
• extract_raw.py - HuggingFace data extraction script
• raw_to_standardized.py - Conversion to standardized format
• sample_raw.json - 3 example raw trajectories
• sample_std.json - Standardized data
• sample_sft.json - OpenHands SFT format

✅ Tests pass: All 4 relevant test suites pass for the new dataset
✅ Branch pushed: openhands/add-nemotron-terminal-corpus-dataset
✅ PR created: #165
✅ PR mentions issue: Includes "Fixes #164" to auto-close the issue

Changes Summary

The implementation follows the existing dataset structure patterns in the repository. The dataset contains ~366k terminal interaction trajectories designed to scale LLM terminal capabilities. The raw-to-standardized conversion handles the dataset's unique format with <think> blocks and JSON command structures, converting them to CodeAction and TextObservation objects.

All changes are scoped only to the new datasets/nemotron_terminal_corpus/ directory with no extraneous modifications to existing files.