NOTICE.md

Attribution and modifications

rumik-oss 1 base is a speech-adapted derivative of CohereLabs/tiny-aya-fire. Rumik AI added audio-token embeddings and a learned stop predictor and trained the model for multilingual, named-speaker speech synthesis.

• Tiny Aya Fire: https://huggingface.co/CohereLabs/tiny-aya-fire
• License: https://cohere.com/cohere-labs-cc-by-nc-license
• Acceptable-use policy: https://docs.cohere.com/docs/cohere-labs-acceptable-use-policy

CC-BY-NC 4.0 requires that redistribution indicate what was changed relative to the upstream material. The modifications are:

1. Vocabulary extended from 261,008 to 277,395 tokens:

   • 16,384 Mimi codec unit tokens <code>_<quantizer>, ids 261008-277391
   • 3 structure delimiters <text>, <audio>, </audio>, ids 277392-277394

   Input and output embedding matrices were resized accordingly.

2. A stop-prediction head (stop_predictor) was added on top of the final hidden state to terminate audio generation.

3. The model was trained for multilingual, named-speaker text-to-speech, progressing from English through Indic-language continuation to named-speaker adaptation.

4. Model class and configuration are provided as RumikOSSForCausalLM / RumikOSSConfig (model_type rumik_oss), subclassing the transformers Cohere2 implementation. This is not a chat model and ships no chat template; it uses the flat prompt layout documented in README.md.

The delivery-control fine-tuning and reinforcement learning that follow this checkpoint are applied in rumik-oss 1.

Bundled codec

The unmodified Mimi codec in codec/ is by Kyutai, distributed under CC BY 4.0.

• Mimi: https://huggingface.co/kyutai/mimi
• License: https://creativecommons.org/licenses/by/4.0/

The codec's license does not replace the license governing the TTS model.