tiny-aya-translate/stage2-tpu-v6e-spot

Stage 2, TPU v6e spot — superseded run artifact

⚠️ Not a release. Use v0.3 instead.

A single intermediate checkpoint auto-pushed by the trainer during an early TPU v6e spot run. No evaluation, no suite, no step guarantee.

The released model is tr-hi-s2st-v0.3.

Note this one carries peft_adapter/adapter_model.bin rather than safetensors, reflecting an earlier save path. Kept public for provenance.

Weights are LoRA derivatives of CohereLabs/tiny-aya-base (CC-BY-NC-4.0) and inherit its non-commercial terms.

Code

repo	what it does
model	Stage-2 training, evaluation harness and TPU launch tooling

Project

TinyAya Stage 2 — Turkish⇄Hindi speech-to-speech translation with a text inner-monologue: a LoRA-adapted Cohere2 backbone driving a frozen Moshi depth decoder over Mimi codes.

The v0.3 run covered 76,250 steps / 2.07 epochs on a Cloud TPU v6e-16 (best val composite 2.8199 @ step 76,000). Read honestly: the text inner-monologue learns to translate (free-run chrF++ ~25.7 / 25.1), while intelligible audio synthesis remains the frontier (ASR-chrF++ 3.7 / 9.6 against a 92.1 / 86.6 ground-truth-audio ceiling) — bounded by the frozen depth decoder, not by translation understanding.

• Results: v0.3 evaluation report
• Training run: W&B xzcb60bl · emergence report
• Blog: Adapting Moshi for Low-Resource Speech Translation

Compute for the v0.3 run was provided by Google's TPU Research Cloud (TRC).