| rumik-oss 1 |
| Copyright 2026 Rumik AI |
|
|
| This model is a derivative work. CC-BY-NC 4.0 requires that redistribution |
| indicate what was changed relative to the upstream material; this file records |
| those changes. |
|
|
| |
| Lineage: |
|
|
| CohereLabs/tiny-aya-fire vocab 261,008 |
| https://huggingface.co/CohereLabs/tiny-aya-fire |
| Licensed CC-BY-NC 4.0 with the Cohere Labs acceptable-use addendum. |
|
|
| -> rumik-ai/rumik-oss-1-base vocab 277,395 |
| https://huggingface.co/rumik-ai/rumik-oss-1-base |
| Named-speaker text-to-speech training. |
|
|
| -> rumik-ai/rumik-oss-1 vocab 277,404 (this model) |
| Delivery-control fine-tuning and reinforcement learning. |
| |
|
|
| Modifications made by Rumik AI, relative to tiny aya fire: |
|
|
| 1. Vocabulary extended from 261,008 to 277,404 tokens: |
| - 16,384 Mimi codec unit tokens <code>_<quantizer>, ids 261008-277391 |
| - 3 structure delimiters <text>, <audio>, </audio>, ids 277392-277394 |
| - 9 delivery and vocalization tags, ids 277395-277403 |
| The first two groups were added for rumik-oss 1 base; the tags were |
| added for rumik-oss 1. 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 post-trained for text-to-speech on multilingual Indic and |
| English speech data, followed by a reinforcement-learning stage. |
|
|
| 4. Model class and configuration were reimplemented as RumikOSSForCausalLM / |
| RumikOSSConfig (model_type "rumik_oss"), subclassing the transformers |
| Cohere2 implementation. |
|
|
| 5. The inherited chat template was removed. This is not a chat model and |
| does not take a `messages` list; it uses the flat prompt layout |
| documented in README.md. |
|
|
| |
| Also bundled: |
|
|
| Mimi codec (codec/) - Kyutai |
| https://huggingface.co/kyutai/mimi |
| Licensed CC-BY-4.0. Redistributed unmodified. |
| |
|
|
