v0.3 — TR↔HI S2ST training-code freeze (long-horizon run in flight)

TinyAya Stage 2 — v0.3 training-code freeze

Code freeze for the v0.3 Turkish↔Hindi speech-to-speech translation training
program: a Cohere2 backbone (LoRA) fused with a frozen Moshi depth decoder over
Mimi codes, trained dual-backend (GPU/TPU) with the multi-host TPU v6e-16 SPMD
pipeline.

Status: the v0.3 long-horizon training run (110,463 steps) launched from
this code and is currently in flight — live metrics:
W&B run v0.3-long-horizon-mh-r2.
Model weights are not published yet — the Hugging Face checkpoint suite
(tiny-aya-translate/tr-hi-s2st-v0.3) goes public after the run completes
and the release evals pass. This tag freezes the training/eval code that
produces it.

What's in v0.3

• Recipe (locked, sweep-validated): LoRA r=32 rsLoRA (+MLP, embed_tokens),
  WSD schedule with stop-anytime anneal, global batch 32, grad-clip 1.0,
  per-codebook loss multipliers + progressive coarse→fine codebook unmasking
  (the deep-codebook-collapse fix; onset→codebook mapping is test-pinned).
• Multi-host TPU pipeline: minibatch data-parallel input
  (per_chip_batch × DistributedSampler × shard_to_device) on v6e-16 SPMD,
  rendezvous-synchronized launch, spot-preemption self-heal, keep-all
  checkpointing to GCS with log-spaced early saves + async HF-hub publishing.
• Evals program (src/evaluation/, scripts/eval_release.py): 8-stage
  resumable release evaluator — AR generation, Whisper ASR judges
  (chrF++/BLEU/WER with GT-audio topline), DNSMOS/Distill-MOS deltas,
  BLASER-2.0 + COMET, Gemini GEMBA adequacy, RTF/TTFA latency — plus frozen
  digest-verified eval subsets (v03-val-500, v03-fleurs-200; FLEURS is
  acoustic-shift only, 200/200 text overlap audited and disclosed).
• Release instrumentation: exact token axes (tokens_seen = 9 tokens/frame),
  curriculum-independent train/audio_loss_full, per-codebook loss/entropy/
  accuracy tracking, MFU estimate, checkpoint index, provenance stamping.

Development trail

The full v0.3 implementation history (sweeps, probes, dress rehearsals,
hardening drills, findings) is in PR #10.

🤖 Generated with Claude Code