# Source terms and attribution

This draft adapts the weights of [OpenBMB MiniCPM5-2B-DSpark](https://huggingface.co/openbmb/MiniCPM5-2B-DSpark) for the [MiniCPM5-2B-SFT](https://huggingface.co/openbmb/MiniCPM5-2B-SFT) target. Both upstream model cards state Apache-2.0. A copy of the Apache-2.0 text is included as `UPSTREAM_APACHE-2.0.txt`; upstream attribution and notices must be retained.

The v4/v5 training sets include [PerfectBlend deduplicated](https://huggingface.co/datasets/tepirale/open-perfectblend-deduplicate-mlabonne), [NVIDIA OpenCodeInstruct](https://huggingface.co/datasets/nvidia/OpenCodeInstruct), [Salesforce APIGen-MT-5k](https://huggingface.co/datasets/Salesforce/APIGen-MT-5k), and [NVIDIA Nemotron-Agentic-v1](https://huggingface.co/datasets/nvidia/Nemotron-Agentic-v1). The APIGen dataset is marked CC BY-NC 4.0. Its card says some GPT-4-generated data should not be used to develop models competing with OpenAI. Per-row generator provenance for the selected samples was not available in the experiment records. This release is intended for noncommercial research and makes no commercial-use claim. Users are responsible for reviewing all source terms before redistribution or downstream use.

No training conversations or generated responses are distributed with these weights. This summary is provenance information, not a legal clearance determination.