Open
Conversation
This PR adds support for the NVIDIA Nemotron Terminal Corpus dataset, a large-scale SFT dataset designed to scale terminal interaction capabilities of LLMs. Dataset includes: - README.md with dataset documentation and citation - schema_raw.py defining the raw data schema - extract_raw.py for downloading data from HuggingFace - raw_to_standardized.py for converting to standardized format - sample_raw.json with 3 example trajectories - sample_std.json with standardized data - sample_sft.json in OpenHands SFT format Closes #164 Co-authored-by: openhands <openhands@all-hands.dev>
This change adds proper support for chain-of-thought reasoning and think tags following conventions from Harbor ATIF and Agent Client Protocol: Schema changes: - Add reasoning_content field to base Action class - Update SCHEMA.md documentation to explain the new field Dataset updates: - nemotron_terminal_corpus: Use reasoning_content for <think> blocks - toucan_1_5m: Use reasoning_content for reasoning_content in raw data SFT conversion updates: - openhands/std_to_sft.py: Handle reasoning_content in CodeAction, ApiAction, MessageAction - sweagent/std_to_sft.py: Handle reasoning_content with helper function The reasoning_content field is separate from the description field: - reasoning_content: Extended chain-of-thought reasoning (<think> blocks) - description: Brief action description/summary Co-authored-by: openhands <openhands@all-hands.dev>
Co-authored-by: openhands <openhands@all-hands.dev>
…ample files Changes: - agents/openhands/std_to_sft.py: Add _build_thought_text() helper that wraps reasoning_content in <think> tags while keeping description as plain text - agents/sweagent/std_to_sft.py: Same helper function for consistency - datasets/nemotron_terminal_corpus/sample_sft.json: Regenerated with proper function_call and observation labels (not converted to gpt/human) - datasets/nemotron_terminal_corpus/sample_sft/: Agent-specific SFT samples The <think> tags preserve the training signal so models learn to produce chain-of-thought reasoning in the expected format. Co-authored-by: openhands <openhands@all-hands.dev>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
This PR adds support for the NVIDIA Nemotron Terminal Corpus dataset, a large-scale Supervised Fine-Tuning (SFT) dataset designed to scale terminal interaction capabilities of Large Language Models (LLMs).
About the Dataset
Terminal-Corpus was developed by NVIDIA using the Terminal-Task-Gen pipeline, which combines dataset adaptation with synthetic task generation across diverse domains. The dataset contains approximately 366k high-quality execution trajectories for terminal agents.
Source: https://huggingface.co/datasets/nvidia/Nemotron-Terminal-Corpus
Files Added
README.md- Dataset documentation including description and citationschema_raw.py- Pydantic schema for raw data validationextract_raw.py- Script for downloading data from HuggingFaceraw_to_standardized.py- Script for converting raw data to standardized formatsample_raw.json- 3 example trajectories in raw formatsample_std.json- Standardized data formatsample_sft.json- OpenHands SFT formatTesting
All existing tests pass with the new dataset:
test_dataset_structure.pytest_raw_schemas.pytest_standardized_schemas.pytest_std_to_sft_conversion.pyFixes #164