Clarify sharding restrictions and data-parallel support

Restrict Transformers to data-parallel replicas

fix: align EOS token ID with tokenizer and training contract

Update README.md

Support assistant reasoning content and clarify single-turn usage

Add Hugging Face Transformers inference support

Document recommended Violetto sampling settings

Set default top_p to 0.95

Use release blog wording for intended use

Add Violetto model card and evaluation table

Set default sampling temperature to 0.6

Enforce canonical math system and text-only chat

Add files using upload-large-folder tool

initial commit