Card: full release metadata + code cross-links
Browse files
README.md
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- tr
|
| 4 |
+
- hi
|
| 5 |
+
license: cc-by-nc-4.0
|
| 6 |
+
base_model: CohereLabs/tiny-aya-base
|
| 7 |
+
base_model_relation: adapter
|
| 8 |
+
tags:
|
| 9 |
+
- superseded
|
| 10 |
+
- checkpoint-artifact
|
| 11 |
+
- tpu
|
| 12 |
+
- tinyaya
|
| 13 |
+
pipeline_tag: audio-to-audio
|
| 14 |
+
pretty_name: Stage 2 TPU v6e spot (superseded run artifact)
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# Stage 2, TPU v6e spot — superseded run artifact
|
| 18 |
+
|
| 19 |
+
> ### ⚠️ Not a release. Use v0.3 instead.
|
| 20 |
+
> A **single intermediate checkpoint** auto-pushed by the trainer during an early
|
| 21 |
+
> TPU v6e spot run. No evaluation, no suite, no step guarantee.
|
| 22 |
+
>
|
| 23 |
+
> The released model is
|
| 24 |
+
> **[`tr-hi-s2st-v0.3`](https://huggingface.co/tiny-aya-translate/tr-hi-s2st-v0.3)**.
|
| 25 |
+
|
| 26 |
+
Note this one carries `peft_adapter/adapter_model.bin` rather than
|
| 27 |
+
`safetensors`, reflecting an earlier save path. Kept public for provenance.
|
| 28 |
+
|
| 29 |
+
Weights are LoRA derivatives of `CohereLabs/tiny-aya-base` (CC-BY-NC-4.0) and
|
| 30 |
+
inherit its **non-commercial** terms.
|
| 31 |
+
|
| 32 |
+
## Code
|
| 33 |
+
|
| 34 |
+
| repo | what it does |
|
| 35 |
+
|---|---|
|
| 36 |
+
| [`model`](https://github.com/tiny-aya-simultaneous-translation/model) | Stage-2 training, evaluation harness and TPU launch tooling |
|
| 37 |
+
|
| 38 |
+
## Project
|
| 39 |
+
|
| 40 |
+
**TinyAya Stage 2** — Turkish⇄Hindi speech-to-speech translation with a text
|
| 41 |
+
inner-monologue: a LoRA-adapted Cohere2 backbone driving a **frozen** Moshi depth
|
| 42 |
+
decoder over Mimi codes.
|
| 43 |
+
|
| 44 |
+
The v0.3 run covered **76,250 steps / 2.07 epochs** on a Cloud TPU v6e-16
|
| 45 |
+
(best val composite **2.8199** @ step 76,000). Read honestly: the text
|
| 46 |
+
inner-monologue **learns to translate** (free-run chrF++ ~25.7 / 25.1), while
|
| 47 |
+
**intelligible audio synthesis remains the frontier** (ASR-chrF++ 3.7 / 9.6
|
| 48 |
+
against a 92.1 / 86.6 ground-truth-audio ceiling) — bounded by the frozen depth
|
| 49 |
+
decoder, not by translation understanding.
|
| 50 |
+
|
| 51 |
+
- **Results:** [v0.3 evaluation report](https://github.com/tiny-aya-simultaneous-translation/model/blob/main/docs/v0.3-eval-report.md)
|
| 52 |
+
- **Training run:** [W&B `xzcb60bl`](https://wandb.ai/cataluna84/tinyaya-stage2-tpu/runs/xzcb60bl) · [emergence report](https://wandb.ai/cataluna84/tinyaya-stage2-tpu/reports/TinyAya-v0.3-Emergence-and-Data-Efficiency--VmlldzoxNzU1OTU1NQ==)
|
| 53 |
+
- **Blog:** [Adapting Moshi for Low-Resource Speech Translation](https://labscommunity.cohere.com/blog/2026/adapting-moshi-low-resource-speech-translation/)
|
| 54 |
+
|
| 55 |
+
Compute for the v0.3 run was provided by **Google's TPU Research Cloud (TRC)**.
|