refactor: rumik-oss (#1)

- refactor: rename TinyAya to RumikOSS (c22516b73eac945db6b1752c2ba6be3185bd6ece)
- feat: add chat template for the trained prompt layout (52f8c0d50db25c2e3c8a0c9196f8f74f261579d2)
- add: LICENSE, and record modifications in NOTICE.md (408f06115787f3c25d49bad9717f6ab6c3fcbbe5)
- update: inference.py to use the config-driven audio API (dcced1a8047fab03966f06244daa4837c3254237)
- docs: rewrite the model card inference section (1b3425a17559b6f9bba2525ebfa157a90c123ce0)
- chore: relax the transformers pin (8b30ada9cd47121798ca881ed60b9b6c43c0e19a)
- revert: drop the chat template (96c18115faf0beac8e87cd5aa1277173d0391d6c)
- fix: recover from a broken round robin instead of truncating (d583cae6ae0c0bf9d3018b6885b327deacf8fd7b)
- fix: drop the huggingface_hub upper bound (7fba0e6a9924404af90ce00484f13570a71d1cbc)
- fix: install soundfile for inference examples (85848bff82548e4bdbb7e7c30faa67e05da396df)


Co-authored-by: nullHawk <nullHawk@users.noreply.huggingface.co>

LICENSE

@@ -0,0 +1,430 @@
+rumik-oss 1 base
+Copyright 2026 Rumik AI
+
+rumik-oss 1 base is a derivative of CohereLabs/tiny-aya-fire and is released under
+the Creative Commons Attribution-NonCommercial 4.0 International license
+(CC-BY-NC 4.0), together with the Cohere Labs acceptable-use addendum that
+governs the upstream model.
+
+  License ......... https://cohere.com/cohere-labs-cc-by-nc-license
+  Acceptable use .. https://docs.cohere.com/docs/cohere-labs-acceptable-use-policy
+
+Use of this model is permitted for research and other non-commercial purposes
+only, and must comply with the Cohere Labs Acceptable Use Policy in addition
+to the terms below. Redistribution must retain this notice, the license link,
+upstream notices, and an indication of any modifications made (see NOTICE).
+
+The Mimi codec in codec/ is licensed separately by Kyutai under CC-BY-4.0
+(https://creativecommons.org/licenses/by/4.0/). That permission applies to the
+codec only, not to the tiny-aya-derived TTS weights.
+
+=======================================================================
+
+Attribution-NonCommercial 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+     Considerations for licensors: Our public licenses are
+     intended for use by those authorized to give the public
+     permission to use material in ways otherwise restricted by
+     copyright and certain other rights. Our licenses are
+     irrevocable. Licensors should read and understand the terms
+     and conditions of the license they choose before applying it.
+     Licensors should also secure all rights necessary before
+     applying our licenses so that the public can reuse the
+     material as expected. Licensors should clearly mark any
+     material not subject to the license. This includes other CC-
+     licensed material, or material used under an exception or
+     limitation to copyright. More considerations for licensors:
+    wiki.creativecommons.org/Considerations_for_licensors
+
+     Considerations for the public: By using one of our public
+     licenses, a licensor grants the public permission to use the
+     licensed material under specified terms and conditions. If
+     the licensor's permission is not necessary for any reason--for
+     example, because of any applicable exception or limitation to
+     copyright--then that use is not regulated by the license. Our
+     licenses grant only permissions under copyright and certain
+     other rights that a licensor has authority to grant. Use of
+     the licensed material may still be restricted for other
+     reasons, including because others have copyright or other
+     rights in the material. A licensor may make special requests,
+     such as asking that all changes be marked or described.
+     Although not required by our licenses, you are encouraged to
+     respect those requests where reasonable. More considerations
+     for the public:
+    wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution-NonCommercial 4.0 International Public
+License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution-NonCommercial 4.0 International Public License ("Public
+License"). To the extent this Public License may be interpreted as a
+contract, You are granted the Licensed Rights in consideration of Your
+acceptance of these terms and conditions, and the Licensor grants You
+such rights in consideration of benefits the Licensor receives from
+making the Licensed Material available under these terms and
+conditions.
+
+
+Section 1 -- Definitions.
+
+  a. Adapted Material means material subject to Copyright and Similar
+     Rights that is derived from or based upon the Licensed Material
+     and in which the Licensed Material is translated, altered,
+     arranged, transformed, or otherwise modified in a manner requiring
+     permission under the Copyright and Similar Rights held by the
+     Licensor. For purposes of this Public License, where the Licensed
+     Material is a musical work, performance, or sound recording,
+     Adapted Material is always produced where the Licensed Material is
+     synched in timed relation with a moving image.
+
+  b. Adapter's License means the license You apply to Your Copyright
+     and Similar Rights in Your contributions to Adapted Material in
+     accordance with the terms and conditions of this Public License.
+
+  c. Copyright and Similar Rights means copyright and/or similar rights
+     closely related to copyright including, without limitation,
+     performance, broadcast, sound recording, and Sui Generis Database
+     Rights, without regard to how the rights are labeled or
+     categorized. For purposes of this Public License, the rights
+     specified in Section 2(b)(1)-(2) are not Copyright and Similar
+     Rights.
+  d. Effective Technological Measures means those measures that, in the
+     absence of proper authority, may not be circumvented under laws
+     fulfilling obligations under Article 11 of the WIPO Copyright
+     Treaty adopted on December 20, 1996, and/or similar international
+     agreements.
+
+  e. Exceptions and Limitations means fair use, fair dealing, and/or
+     any other exception or limitation to Copyright and Similar Rights
+     that applies to Your use of the Licensed Material.
+
+  f. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  g. Licensed Rights means the rights granted to You subject to the
+     terms and conditions of this Public License, which are limited to
+     all Copyright and Similar Rights that apply to Your use of the
+     Licensed Material and that the Licensor has authority to license.
+
+  h. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  i. NonCommercial means not primarily intended for or directed towards
+     commercial advantage or monetary compensation. For purposes of
+     this Public License, the exchange of the Licensed Material for
+     other material subject to Copyright and Similar Rights by digital
+     file-sharing or similar means is NonCommercial provided there is
+     no payment of monetary compensation in connection with the
+     exchange.
+
+  j. Share means to provide material to the public by any means or
+     process that requires permission under the Licensed Rights, such
+     as reproduction, public display, public performance, distribution,
+     dissemination, communication, or importation, and to make material
+     available to the public including in ways that members of the
+     public may access the material from a place and at a time
+     individually chosen by them.
+
+  k. Sui Generis Database Rights means rights other than copyright
+     resulting from Directive 96/9/EC of the European Parliament and of
+     the Council of 11 March 1996 on the legal protection of databases,
+     as amended and/or succeeded, as well as other essentially
+     equivalent rights anywhere in the world.
+
+  l. You means the individual or entity exercising the Licensed Rights
+     under this Public License. Your has a corresponding meaning.
+
+
+Section 2 -- Scope.
+
+  a. License grant.
+
+       1. Subject to the terms and conditions of this Public License,
+          the Licensor hereby grants You a worldwide, royalty-free,
+          non-sublicensable, non-exclusive, irrevocable license to
+          exercise the Licensed Rights in the Licensed Material to:
+
+            a. reproduce and Share the Licensed Material, in whole or
+               in part, for NonCommercial purposes only; and
+
+            b. produce, reproduce, and Share Adapted Material for
+               NonCommercial purposes only.
+
+       2. Exceptions and Limitations. For the avoidance of doubt, where
+          Exceptions and Limitations apply to Your use, this Public
+          License does not apply, and You do not need to comply with
+          its terms and conditions.
+
+       3. Term. The term of this Public License is specified in Section
+          6(a).
+
+       4. Media and formats; technical modifications allowed. The
+          Licensor authorizes You to exercise the Licensed Rights in
+          all media and formats whether now known or hereafter created,
+          and to make technical modifications necessary to do so. The
+          Licensor waives and/or agrees not to assert any right or
+          authority to forbid You from making technical modifications
+          necessary to exercise the Licensed Rights, including
+          technical modifications necessary to circumvent Effective
+          Technological Measures. For purposes of this Public License,
+          simply making modifications authorized by this Section 2(a)
+          (4) never produces Adapted Material.
+
+       5. Downstream recipients.
+
+            a. Offer from the Licensor -- Licensed Material. Every
+               recipient of the Licensed Material automatically
+               receives an offer from the Licensor to exercise the
+               Licensed Rights under the terms and conditions of this
+               Public License.
+
+            b. No downstream restrictions. You may not offer or impose
+               any additional or different terms or conditions on, or
+               apply any Effective Technological Measures to, the
+               Licensed Material if doing so restricts exercise of the
+               Licensed Rights by any recipient of the Licensed
+               Material.
+
+       6. No endorsement. Nothing in this Public License constitutes or
+          may be construed as permission to assert or imply that You
+          are, or that Your use of the Licensed Material is, connected
+          with, or sponsored, endorsed, or granted official status by,
+          the Licensor or others designated to receive attribution as
+          provided in Section 3(a)(1)(A)(i).
+
+  b. Other rights.
+
+       1. Moral rights, such as the right of integrity, are not
+          licensed under this Public License, nor are publicity,
+          privacy, and/or other similar personality rights; however, to
+          the extent possible, the Licensor waives and/or agrees not to
+          assert any such rights held by the Licensor to the limited
+          extent necessary to allow You to exercise the Licensed
+          Rights, but not otherwise.
+
+       2. Patent and trademark rights are not licensed under this
+          Public License.
+
+       3. To the extent possible, the Licensor waives any right to
+          collect royalties from You for the exercise of the Licensed
+          Rights, whether directly or through a collecting society
+          under any voluntary or waivable statutory or compulsory
+          licensing scheme. In all other cases the Licensor expressly
+          reserves any right to collect such royalties, including when
+          the Licensed Material is used other than for NonCommercial
+          purposes.
+
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+  a. Attribution.
+
+       1. If You Share the Licensed Material (including in modified
+          form), You must:
+
+            a. retain the following if it is supplied by the Licensor
+               with the Licensed Material:
+
+                 i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if
+                    designated);
+
+                ii. a copyright notice;
+
+               iii. a notice that refers to this Public License;
+
+                iv. a notice that refers to the disclaimer of
+                    warranties;
+
+                 v. a URI or hyperlink to the Licensed Material to the
+                    extent reasonably practicable;
+
+            b. indicate if You modified the Licensed Material and
+               retain an indication of any previous modifications; and
+
+            c. indicate the Licensed Material is licensed under this
+               Public License, and include the text of, or the URI or
+               hyperlink to, this Public License.
+
+       2. You may satisfy the conditions in Section 3(a)(1) in any
+          reasonable manner based on the medium, means, and context in
+          which You Share the Licensed Material. For example, it may be
+          reasonable to satisfy the conditions by providing a URI or
+          hyperlink to a resource that includes the required
+          information.
+
+       3. If requested by the Licensor, You must remove any of the
+          information required by Section 3(a)(1)(A) to the extent
+          reasonably practicable.
+
+       4. If You Share Adapted Material You produce, the Adapter's
+          License You apply must not prevent recipients of the Adapted
+          Material from complying with this Public License.
+
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+     to extract, reuse, reproduce, and Share all or a substantial
+     portion of the contents of the database for NonCommercial purposes
+     only;
+
+  b. if You include all or a substantial portion of the database
+     contents in a database in which You have Sui Generis Database
+     Rights, then the database in which You have Sui Generis Database
+     Rights (but not its individual contents) is Adapted Material; and
+
+  c. You must comply with the conditions in Section 3(a) if You Share
+     all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+  c. The disclaimer of warranties and limitation of liability provided
+     above shall be interpreted in a manner that, to the extent
+     possible, most closely approximates an absolute disclaimer and
+     waiver of all liability.
+
+
+Section 6 -- Term and Termination.
+
+  a. This Public License applies for the term of the Copyright and
+     Similar Rights licensed here. However, if You fail to comply with
+     this Public License, then Your rights under this Public License
+     terminate automatically.
+
+  b. Where Your right to use the Licensed Material has terminated under
+     Section 6(a), it reinstates:
+
+       1. automatically as of the date the violation is cured, provided
+          it is cured within 30 days of Your discovery of the
+          violation; or
+
+       2. upon express reinstatement by the Licensor.
+
+     For the avoidance of doubt, this Section 6(b) does not affect any
+     right the Licensor may have to seek remedies for Your violations
+     of this Public License.
+
+  c. For the avoidance of doubt, the Licensor may also offer the
+     Licensed Material under separate terms or conditions or stop
+     distributing the Licensed Material at any time; however, doing so
+     will not terminate this Public License.
+
+  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+     License.
+
+
+Section 7 -- Other Terms and Conditions.
+
+  a. The Licensor shall not be bound by any additional or different
+     terms or conditions communicated by You unless expressly agreed.
+
+  b. Any arrangements, understandings, or agreements regarding the
+     Licensed Material not stated herein are separate from and
+     independent of the terms and conditions of this Public License.
+
+
+Section 8 -- Interpretation.
+
+  a. For the avoidance of doubt, this Public License does not, and
+     shall not be interpreted to, reduce, limit, restrict, or impose
+     conditions on any use of the Licensed Material that could lawfully
+     be made without permission under this Public License.
+
+  b. To the extent possible, if any provision of this Public License is
+     deemed unenforceable, it shall be automatically reformed to the
+     minimum extent necessary to make it enforceable. If the provision
+     cannot be reformed, it shall be severed from this Public License
+     without affecting the enforceability of the remaining terms and
+     conditions.
+
+  c. No term or condition of this Public License will be waived and no
+     failure to comply consented to unless expressly agreed to by the
+     Licensor.
+
+  d. Nothing in this Public License constitutes or may be interpreted
+     as a limitation upon, or waiver of, any privileges and immunities
+     that apply to the Licensor or You, including from the legal
+     processes of any jurisdiction or authority.
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.
+

NOTICE.md

@@ -8,6 +8,33 @@ the model for multilingual, named-speaker speech synthesis.
 - 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](https://huggingface.co/rumik-ai/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

README.md

@@ -1,13 +1,39 @@
 ---
 pipeline_tag: text-to-speech
+library_name: transformers
 base_model: CohereLabs/tiny-aya-fire
-license: other
-license_name: cohere-cc-by-nc-4.0-with-acceptable-use-addendum
-license_link: https://cohere.com/cohere-labs-cc-by-nc-license
+base_model_relation: finetune
+license: cc-by-nc-4.0
+language:
+  - hi
+  - bn
+  - ta
+  - te
+  - mr
+  - gu
+  - kn
+  - ml
+  - pa
+  - or
+  - as
+  - ur
+  - ne
+  - sa
+  - mai
+  - mni
+  - brx
+  - doi
+  - kok
+  - sat
+  - ks
+  - en
 tags:
- - tts
- - indic
- - multilingual
+  - tts
+  - text-to-speech
+  - indic
+  - multilingual
+  - mimi
+  - speech
 ---
 
 <p align="center">
@@ -38,17 +64,73 @@ this base release has not undergone description-conditioned delivery or inline v
 
 ## inference
 
+rumik-oss 1 base generates **audio tokens**. decoding them to a waveform is a
+separate step that uses the [mimi](https://huggingface.co/kyutai/mimi) codec,
+bundled in this repository under `codec/`.
+
+### installation
+
 ```bash
-hf download rumik-ai/rumik-oss-1-base --local-dir rumik-oss-1-base
-cd rumik-oss-1-base
 pip install -r requirements.txt
-python inference.py --speaker Ira --text "Hello, how are you today?" --output speech.wav
 ```
 
-the `--speaker` argument accepts `Ira`, `Aisha`, `Siya`, or `Zoya`. pass the spoken text through `--text`; the inference script constructs the speaker-prefixed sequence:
+requires an nvidia gpu with cuda support. the one-shot example below also
+needs `soundfile` to write a wav.
+
+### one-shot example
+
+text in, `speech.wav` out:
+
+```python
+"""rumik-oss 1 base: text -> speech.wav"""
+import soundfile as sf
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, MimiModel
+
+REPO = "rumik-ai/rumik-oss-1-base"
+SPEAKER = "Ira"
+TEXT = "Hello, how are you today?"
+
+tokenizer = AutoTokenizer.from_pretrained(REPO, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(
+    REPO, trust_remote_code=True, dtype=torch.bfloat16
+).eval().cuda()
+mimi = MimiModel.from_pretrained(REPO, subfolder="codec").eval().cuda()
+
+# [BOS] <text> Ira: text <audio>
+# the tokenizer adds [BOS] itself, so it is not written here
+prompt = f'<text>{SPEAKER}: {TEXT}<audio>'
+inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+
+# 1. text -> audio tokens
+ids = model.generate_audio(**inputs, max_new_tokens=2048,
+                           temperature=0.8, top_k=30, do_sample=True)
+audio_tokens = ids[0].tolist()[inputs.input_ids.shape[1]:]
+
+# 2. audio tokens -> codec frames -> waveform
+codes = model.audio_tokens_to_codes(audio_tokens)
+with torch.inference_mode():
+    wav = mimi.decode(codes.to(mimi.device)).audio_values[0, 0]
+
+sf.write("speech.wav", wav.float().cpu().numpy(), 24000)
+print(f"{len(audio_tokens)} tokens -> {len(audio_tokens) / 100:.2f}s -> speech.wav")
+```
+
+`SPEAKER` accepts `Ira`, `Aisha`, `Siya`, or `Zoya`. the model was trained on
+a flat token layout, not a chat template:
 
 ```text
-Ira: Hello, how are you today?<audio>
+[BOS] <text> {speaker}: spoken text <audio> ... </audio>
+```
+
+this checkpoint has no description-conditioned delivery or inline
+vocalization controls; use the
+[post-trained model](https://huggingface.co/rumik-ai/rumik-oss-1) for those.
+
+### command line
+
+```bash
+python inference.py --speaker Ira --text "Hello, how are you today?" --output speech.wav
 ```
 
 ## further training
@@ -63,6 +145,9 @@ permitted non-commercial uses include speech-synthesis research, benchmarking, t
 
 the bundled [mimi codec](https://huggingface.co/kyutai/mimi) is separately licensed under [cc-by-4.0](https://creativecommons.org/licenses/by/4.0/). its commercial permissions do not extend to the tiny aya-derived tts weights.
 
+the full license text is in [`LICENSE`](LICENSE), and [`NOTICE.md`](NOTICE.md)
+records the modifications made to tiny aya fire, as cc-by-nc 4.0 requires.
+
 ## acknowledgments
 
 rumik-oss 1 base builds on [tiny aya fire](https://huggingface.co/CohereLabs/tiny-aya-fire), [mimi](https://huggingface.co/kyutai/mimi), [llama-mimi](https://arxiv.org/abs/2509.14882), and [torchtitan](https://github.com/pytorch/torchtitan). we thank their authors for making this work available.

config.json

@@ -1,21 +1,54 @@
 {
-  "vocab_size": 277395,
-  "max_position_embeddings": 500000,
-  "hidden_size": 2048,
-  "logit_scale": 1.0,
-  "intermediate_size": 11008,
-  "num_hidden_layers": 36,
-  "num_attention_heads": 16,
-  "num_key_value_heads": 4,
+  "_sliding_window_pattern": 4,
+  "add_cross_attention": false,
+  "architectures": [
+    "RumikOSSForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "audio_end_token_id": 277394,
+  "audio_start_token_id": 277393,
+  "auto_map": {
+    "AutoConfig": "configuration_rumik_oss.RumikOSSConfig",
+    "AutoModelForCausalLM": "modeling_rumik_oss.RumikOSSForCausalLM"
+  },
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 2,
+  "chunk_size_feed_forward": 0,
+  "codebook_size": 2048,
+  "cross_attention_hidden_size": null,
+  "decoder_start_token_id": null,
+  "diversity_penalty": 0.0,
+  "do_sample": false,
+  "dtype": "bfloat16",
+  "early_stopping": false,
+  "encoder_no_repeat_ngram_size": 0,
+  "eos_token_id": 3,
+  "exponential_decay_length_penalty": null,
+  "finetuning_task": null,
+  "first_unit_id": 261008,
+  "forced_bos_token_id": null,
+  "forced_eos_token_id": null,
+  "frame_rate_hz": 12.5,
+  "head_dim": 128,
   "hidden_act": "silu",
+  "hidden_size": 2048,
+  "id2label": {
+    "0": "LABEL_0",
+    "1": "LABEL_1"
+  },
   "initializer_range": 0.02,
+  "intermediate_size": 11008,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "last_unit_id": 277391,
   "layer_norm_eps": 1e-05,
-  "use_cache": true,
-  "rope_theta": 50000,
-  "rope_scaling": null,
-  "attention_bias": false,
-  "attention_dropout": 0.0,
-  "sliding_window": 4096,
+  "layer_switch": 4,
   "layer_types": [
     "sliding_attention",
     "sliding_attention",
@@ -54,94 +87,64 @@
     "sliding_attention",
     "full_attention"
   ],
-  "head_dim": 128,
-  "return_dict": true,
-  "output_hidden_states": false,
-  "torchscript": false,
-  "dtype": "bfloat16",
-  "pruned_heads": {},
-  "tie_word_embeddings": true,
-  "chunk_size_feed_forward": 0,
-  "is_encoder_decoder": false,
-  "is_decoder": false,
-  "cross_attention_hidden_size": null,
-  "add_cross_attention": false,
-  "tie_encoder_decoder": false,
-  "architectures": [
-    "TinyAyaForCausalLM"
-  ],
-  "finetuning_task": null,
-  "id2label": {
-    "0": "LABEL_0",
-    "1": "LABEL_1"
-  },
-  "label2id": {
-    "LABEL_0": 0,
-    "LABEL_1": 1
-  },
-  "task_specific_params": null,
-  "problem_type": null,
-  "tokenizer_class": null,
-  "prefix": null,
-  "bos_token_id": 2,
-  "pad_token_id": 0,
-  "eos_token_id": 3,
-  "sep_token_id": null,
-  "decoder_start_token_id": null,
+  "length_penalty": 1.0,
+  "logit_scale": 1.0,
   "max_length": 20,
+  "max_position_embeddings": 500000,
   "min_length": 0,
-  "do_sample": false,
-  "early_stopping": false,
-  "num_beams": 1,
-  "temperature": 1.0,
-  "top_k": 50,
-  "top_p": 1.0,
-  "typical_p": 1.0,
-  "repetition_penalty": 1.0,
-  "length_penalty": 1.0,
+  "model_type": "rumik_oss",
   "no_repeat_ngram_size": 0,
-  "encoder_no_repeat_ngram_size": 0,
-  "bad_words_ids": null,
-  "num_return_sequences": 1,
-  "output_scores": false,
-  "return_dict_in_generate": false,
-  "forced_bos_token_id": null,
-  "forced_eos_token_id": null,
-  "remove_invalid_values": false,
-  "exponential_decay_length_penalty": null,
-  "suppress_tokens": null,
-  "begin_suppress_tokens": null,
+  "num_attention_heads": 16,
   "num_beam_groups": 1,
-  "diversity_penalty": 0.0,
-  "_sliding_window_pattern": 4,
-  "cache_implementation": "hybrid",
-  "layer_switch": 4,
-  "model_type": "tinyaya_mimi",
+  "num_beams": 1,
+  "num_hidden_layers": 36,
+  "num_key_value_heads": 4,
+  "num_quantizers": 8,
+  "num_return_sequences": 1,
   "order_of_interleaved_layers": "local_attn_first",
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "output_scores": false,
+  "pad_token_id": 0,
   "position_embedding_type": "rope_gptj",
+  "prefix": null,
+  "problem_type": null,
+  "pruned_heads": {},
+  "remove_invalid_values": false,
+  "repetition_penalty": 1.0,
+  "return_dict": true,
+  "return_dict_in_generate": false,
+  "rope_scaling": null,
+  "rope_theta": 50000,
   "rotary_pct": 1.0,
+  "sep_token_id": null,
+  "sliding_window": 4096,
   "sliding_window_pattern": 4,
-  "use_embedding_sharing": true,
-  "use_gated_activation": true,
-  "use_parallel_block": true,
-  "use_parallel_embedding": false,
-  "use_qk_norm": false,
-  "tf_legacy_loss": false,
-  "use_bfloat16": false,
-  "output_attentions": false,
-  "num_quantizers": 8,
-  "codebook_size": 2048,
-  "audio_start_token_id": 277393,
-  "audio_end_token_id": 277394,
   "speakers": [
     "Ira",
     "Aisha",
     "Siya",
     "Zoya"
   ],
-  "auto_map": {
-    "AutoConfig": "configuration_tinyaya.TinyAyaConfig",
-    "AutoModelForCausalLM": "modeling_tinyaya.TinyAyaForCausalLM"
-  },
-  "torch_dtype": "bfloat16"
+  "suppress_tokens": null,
+  "task_specific_params": null,
+  "temperature": 1.0,
+  "text_start_token_id": 277392,
+  "tf_legacy_loss": false,
+  "tie_encoder_decoder": false,
+  "tie_word_embeddings": true,
+  "tokenizer_class": null,
+  "top_k": 50,
+  "top_p": 1.0,
+  "torch_dtype": "bfloat16",
+  "torchscript": false,
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "use_embedding_sharing": true,
+  "use_gated_activation": true,
+  "use_parallel_block": true,
+  "use_parallel_embedding": false,
+  "use_qk_norm": false,
+  "vocab_size": 277395
 }

configuration_rumik_oss.py

@@ -0,0 +1,55 @@
+"""Configuration for rumik-oss 1."""
+
+from transformers import Cohere2Config
+
+
+class RumikOSSConfig(Cohere2Config):
+    """rumik-oss 1: tiny aya fire extended with flattened Mimi codec tokens.
+
+    Beyond the base Cohere2 fields, this records where the audio vocabulary
+    lives so that callers can map token ids to codec (code, quantizer) pairs
+    arithmetically, without loading the tokenizer or scanning its vocabulary.
+
+    The unit tokens are laid out code-major, quantizer-minor::
+
+        <0_0> <0_1> ... <0_7> <1_0> ... <2047_7>
+
+    so for any id in ``[first_unit_id, last_unit_id]``::
+
+        code      = (token_id - first_unit_id) // num_quantizers
+        quantizer = (token_id - first_unit_id) %  num_quantizers
+    """
+
+    model_type = "rumik_oss"
+
+    def __init__(
+        self,
+        num_quantizers=8,
+        codebook_size=2048,
+        audio_start_token_id=None,
+        audio_end_token_id=None,
+        text_start_token_id=None,
+        first_unit_id=None,
+        last_unit_id=None,
+        frame_rate_hz=12.5,
+        speakers=None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.num_quantizers = int(num_quantizers)
+        self.codebook_size = int(codebook_size)
+        self.audio_start_token_id = audio_start_token_id
+        self.audio_end_token_id = audio_end_token_id
+        self.text_start_token_id = text_start_token_id
+        self.first_unit_id = first_unit_id
+        # Derivable, but stored so a hand-edited config stays self-consistent.
+        if last_unit_id is None and first_unit_id is not None:
+            last_unit_id = int(first_unit_id) + self.codebook_size * self.num_quantizers - 1
+        self.last_unit_id = last_unit_id
+        self.frame_rate_hz = float(frame_rate_hz)
+        self.speakers = speakers or ["Ira", "Aisha", "Siya", "Zoya"]
+
+    @property
+    def audio_tokens_per_second(self) -> float:
+        """Tokens the model emits per second of audio (8 x 12.5 = 100)."""
+        return self.frame_rate_hz * self.num_quantizers

configuration_tinyaya.py

@@ -1,21 +0,0 @@
-from transformers import Cohere2Config
-
-
-class TinyAyaConfig(Cohere2Config):
-    model_type = "tinyaya_mimi"
-
-    def __init__(
-        self,
-        num_quantizers=8,
-        codebook_size=2048,
-        audio_start_token_id=None,
-        audio_end_token_id=None,
-        speakers=None,
-        **kwargs,
-    ):
-        super().__init__(**kwargs)
-        self.num_quantizers = int(num_quantizers)
-        self.codebook_size = int(codebook_size)
-        self.audio_start_token_id = audio_start_token_id
-        self.audio_end_token_id = audio_end_token_id
-        self.speakers = speakers or ["Ira", "Aisha", "Siya", "Zoya"]

inference.py

@@ -1,7 +1,6 @@
 """Speaker-conditioned inference for rumik-oss 1 base."""
 
 import argparse
-import re
 import wave
 from pathlib import Path
 
@@ -14,34 +13,10 @@ from transformers import (
     MimiModel,
 )
 
-SPEAKERS = ("Ira", "Aisha", "Siya", "Zoya")
-AUDIO_TOKEN = re.compile(r"^<(\d+)_(\d+)>$")
-
-
-def decode_frames(tokens, mapping, quantizers=8):
-    frames, frame = [], []
-    for token in tokens:
-        item = mapping.get(token)
-        if item is None:
-            frame = []
-            continue
-        code, quantizer = item
-        if quantizer == len(frame):
-            frame.append(code)
-            if len(frame) == quantizers:
-                frames.append(frame)
-                frame = []
-        else:
-            frame = [code] if quantizer == 0 else []
-    if not frames:
-        raise RuntimeError("No complete audio frames were generated.")
-    return frames
-
-
 def main():
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument("--model", default=str(Path(__file__).resolve().parent))
-    parser.add_argument("--speaker", choices=SPEAKERS, default="Ira")
+    parser.add_argument("--speaker", default="Ira")
     parser.add_argument("--text", required=True)
     parser.add_argument("--output", default="speech.wav")
     parser.add_argument("--device", default="cuda")
@@ -61,37 +36,30 @@ def main():
         root = Path(snapshot_download(args.model))
     torch.manual_seed(args.seed)
     dtype = torch.bfloat16 if args.device.startswith("cuda") else torch.float32
-    tokenizer = AutoTokenizer.from_pretrained(root)
+    tokenizer = AutoTokenizer.from_pretrained(root, trust_remote_code=True)
     model = (
         AutoModelForCausalLM.from_pretrained(
-            root, trust_remote_code=True, torch_dtype=dtype, attn_implementation="sdpa"
+            root, trust_remote_code=True, dtype=dtype, attn_implementation="sdpa"
         )
         .eval()
         .to(args.device)
     )
     codec = MimiModel.from_pretrained(root / "codec").eval().to(args.device)
     sample_rate = AutoFeatureExtractor.from_pretrained(root / "codec").sampling_rate
-    mapping = {
-        token_id: (int(match[1]), int(match[2]))
-        for token, token_id in tokenizer.get_vocab().items()
-        if (match := AUDIO_TOKEN.fullmatch(token))
-    }
-    eos = tokenizer.convert_tokens_to_ids("</audio>")
+    speakers = tuple(model.config.speakers)
+    if args.speaker not in speakers:
+        parser.error(f"--speaker must be one of: {', '.join(speakers)}")
     inputs = tokenizer(
         f"<text>{args.speaker}: {args.text}<audio>", return_tensors="pt"
     ).to(args.device)
     with torch.inference_mode():
         output = model.generate_audio(
             **inputs,
-            allowed_ids=torch.tensor(sorted([*mapping, eos]), device=args.device),
             max_new_tokens=args.max_new_tokens,
             temperature=args.temperature,
             top_k=args.top_k,
         )[0, inputs.input_ids.shape[1] :].tolist()
-        if eos in output:
-            output = output[: output.index(eos)]
-        frames = decode_frames(output, mapping)
-        codes = torch.tensor(frames, device=args.device).T.unsqueeze(0)
+        codes = model.audio_tokens_to_codes(output).to(args.device)
         audio = codec.decode(codes).audio_values[0, 0].float().cpu().clamp(-1, 1)
     with wave.open(args.output, "wb") as stream:
         stream.setnchannels(1)

modeling_tinyaya.py → modeling_rumik_oss.py

@@ -1,14 +1,24 @@
+"""rumik-oss 1: text -> flattened Mimi codec tokens.
+
+This model produces *audio tokens*, not waveforms. Decoding them to audio is
+the caller's job and needs the Mimi codec (shipped in ``codec/``)::
+
+    ids    = model.generate_audio(**inputs)                  # prompt + audio
+    tokens = ids[0].tolist()[inputs.input_ids.shape[1]:]     # drop the prompt
+    codes  = model.audio_tokens_to_codes(tokens)             # [1, 8, T] for Mimi
+    wav    = mimi.decode(codes).audio_values                 # your call
+"""
 from __future__ import annotations
 
 import torch
 from torch import nn
 from transformers.models.cohere2.modeling_cohere2 import Cohere2ForCausalLM
 
-from .configuration_tinyaya import TinyAyaConfig
+from .configuration_rumik_oss import RumikOSSConfig
 
 
-class TinyAyaForCausalLM(Cohere2ForCausalLM):
-    config_class = TinyAyaConfig
+class RumikOSSForCausalLM(Cohere2ForCausalLM):
+    config_class = RumikOSSConfig
 
     def __init__(self, config):
         super().__init__(config)
@@ -20,8 +30,67 @@ class TinyAyaForCausalLM(Cohere2ForCausalLM):
             nn.Linear(max(64, hidden // 4), 1),
         )
 
+    # ---- audio vocabulary -------------------------------------------------
+    def audio_token_ids(self, device=None) -> torch.Tensor:
+        """Every id the model may legally emit inside an <audio> span.
+
+        Derived from the config, so callers no longer have to scan the
+        tokenizer vocabulary for ``<code>_<quantizer>`` tokens.
+        """
+        c = self.config
+        if c.first_unit_id is None:
+            raise ValueError(
+                "config.first_unit_id is unset; pass allowed_ids explicitly or "
+                "add first_unit_id/last_unit_id to config.json.")
+        ids = torch.arange(int(c.first_unit_id), int(c.last_unit_id) + 1, dtype=torch.long)
+        end = torch.tensor([int(c.audio_end_token_id)], dtype=torch.long)
+        return torch.cat((ids, end)).to(device if device is not None else self.device)
+
+    def audio_tokens_to_codes(self, token_ids) -> torch.Tensor:
+        """Generated ids -> ``[1, num_quantizers, num_frames]`` codec codes.
+
+        Shaped for ``MimiModel.decode``. Purely arithmetic -- no tokenizer, no
+        codec, no audio. Pass only the generated tail, not the prompt.
+
+        A token that breaks the round robin does not end the audio: the
+        partial frame is dropped and the next complete frame is picked up
+        again, so one bad token costs one frame rather than the rest of the
+        clip. ``</audio>`` ends it.
+        """
+        c = self.config
+        first, last, Q = int(c.first_unit_id), int(c.last_unit_id), int(c.num_quantizers)
+        end_id = int(c.audio_end_token_id)
+        if torch.is_tensor(token_ids):
+            token_ids = token_ids.flatten().tolist()
+
+        frames: list[list[int]] = []
+        frame: list[int] = []
+        for tid in (int(t) for t in token_ids):
+            if tid == end_id:
+                break
+            if not first <= tid <= last:
+                frame = []                    # stray token: resync
+                continue
+            code, q = divmod(tid - first, Q)
+            if q == len(frame):
+                frame.append(code)
+                if len(frame) == Q:
+                    frames.append(frame)
+                    frame = []
+            else:                             # off the round robin
+                frame = [code] if q == 0 else []
+        if not frames:
+            raise ValueError(
+                f"no complete codec frame in {len(token_ids)} tokens "
+                f"(need at least {Q})")
+        return torch.tensor(frames, dtype=torch.long).T.unsqueeze(0)
+
+    # ---- generation -------------------------------------------------------
     @staticmethod
-    def _sample(scores, do_sample, temperature, top_k):
+    def _constrained_sample(scores, do_sample, temperature, top_k):
+        # NB: deliberately not named `_sample`. `GenerationMixin._sample` is the
+        # method `generate()` dispatches to, and shadowing it makes every call
+        # to `generate()` fail with a TypeError on `logits_processor`.
         if not do_sample:
             return scores.argmax(dim=-1, keepdim=True)
         scores = scores / max(float(temperature), 1e-5)
@@ -36,13 +105,19 @@ class TinyAyaForCausalLM(Cohere2ForCausalLM):
         self,
         input_ids,
         attention_mask,
-        allowed_ids,
+        allowed_ids=None,
         max_new_tokens=2048,
         min_new_tokens=8,
         temperature=0.8,
         top_k=30,
         do_sample=True,
     ):
+        """Autoregressively emit audio tokens, ending at ``</audio>``.
+
+        ``allowed_ids`` defaults to the audio vocabulary from the config.
+        """
+        if allowed_ids is None:
+            allowed_ids = self.audio_token_ids(device=input_ids.device)
         out = self(
             input_ids=input_ids,
             attention_mask=attention_mask,
@@ -67,7 +142,7 @@ class TinyAyaForCausalLM(Cohere2ForCausalLM):
             restricted.index_copy_(1, allowed_ids, scores.index_select(1, allowed_ids))
             if step < int(min_new_tokens):
                 restricted[:, audio_end_id] = torch.finfo(scores.dtype).min
-            token = self._sample(restricted, do_sample, temperature, top_k)
+            token = self._constrained_sample(restricted, do_sample, temperature, top_k)
             emitted = torch.cat((emitted, token), dim=1)
             if bool((token == audio_end_id).all()):
                 break

requirements.txt

@@ -1,5 +1,6 @@
 torch>=2.9,<3
-transformers==4.57.6
-huggingface_hub>=0.36,<1
+transformers>=4.57,<6
+huggingface_hub>=0.36
 accelerate>=1.12,<2
 safetensors>=0.6,<1
+soundfile>=0.12