license: other
license_name: mixed-source-licenses
OPD prefixes
Prompt-only subsets of NVIDIA Nemotron-Post-Training-Dataset-v1 and NVIDIA SWE-Zero-openhands-trajectories, plus codeparrot/APPS and GSM8K training questions. Nemotron tool-call entries preserve native tools in metadata and tokenize through the model native tool template. These are first-decision contexts, not a complete tool interaction environment. Each source retains its own license, not a choice of dual licenses: NVIDIA sources declare CC-BY-4.0; APPS and GSM8K declare MIT. Attribution: NVIDIA, APPS dataset authors Dan Hendrycks and collaborators, codeparrot, OpenAI/GSM8K authors and upstream contributors. Source cards: https://huggingface.co/datasets/nvidia/Nemotron-Post-Training-Dataset-v1 ; https://huggingface.co/datasets/nvidia/SWE-Zero-openhands-trajectories ; https://huggingface.co/datasets/codeparrot/apps ; https://github.com/openai/grade-school-math . Attribution: NVIDIA and the upstream dataset contributors; source revisions are in manifest.json. Preserve upstream repository license obligations recorded in provenance.
Cached DiffusionGemma tokenizer IDs; original native system/user messages; no assistant targets, solutions or tool outputs. Train/dev task groups are disjoint and lexically screened. This is not universal contamination clearance. SWE entries are static task starts, not an executable repository environment. Broader benchmark adapters and full 8-H100 performance admission remain pending.