Tars weights and BPE vocabulary were trained from random initialization.
Public training sources retained by the research project:
- HuggingFaceTB/smollm-corpus, cosmopedia-v2; declared license ODC-By.
  https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus
- HuggingFaceTB/smol-smoltalk; declared license Apache-2.0.
  https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk
- Project-authored checked math, reasoning, short code and tool examples.
Public-source synthetic text may reflect other language models used by its
original data authors. No pretrained checkpoint or tokenizer was imported.

Dependencies are installed from their publishers under their own licenses:
Python (PSF), uv (MIT/Apache-2.0), MLX (MIT), NumPy (BSD and bundled notices),
Hugging Face tokenizers (Apache-2.0). Their installation includes notices.
This preview does not grant a separate redistribution license for Tars.
