tiny-aya-translate/tr-hi-s2st-v0.3

🗣️🔁 TinyAya — Turkish⇄Hindi Speech-to-Speech Translation (v0.3)

✅ Training complete — plateau + anneal (2026-07-20). The long-horizon run (v0.3-long-horizon-mh-r2) early-stopped on the WSD plateau at step 65,250 (best 2.9048 @ 62,750), then completed its 11,000-step linear LR→0 anneal leg to step 76,250 — improving validation on essentially every cycle of the descent to a final val composite 2.8199 (text ppl 1.489, text acc 96.6%). Best = step 76,000 ≈ final. The repo is public and carries the FULL checkpoint suite (~87 training points as branches and under checkpoints/). End-task release evals are complete — a data-efficiency / emergence study (the model learns text translation ~25 chrF++ but audio synthesis is the next frontier); see Evaluation.

Moshi-style speech-to-speech translation with a text inner-monologue for Turkish ⇄ Hindi: a LoRA-fine-tuned Cohere2 backbone fused with a frozen Moshi depth decoder, operating on Mimi audio codes in a parallel two-stream format. Text+audio (text_weight=0.2): the corpus ships word-level alignments for every sample (see Dataset), so the inner-monologue/text stream is supervised alongside audio — earlier versions trained audio-only due to a loader bug, disclosed below.

• Developed by: tiny-aya-translate
• Blog post: Adapting Moshi for Low-Resource Speech Translation (Cohere Labs Community — the v0.3 training/infrastructure specifics land in its current revision)
• Funded by: Google TPU Research Cloud (TRC)
• Model type: parallel two-stream S2ST (Cohere2 + LoRA → CB0; frozen Moshi depth decoder → CB1–7)
• Languages: Turkish (tr), Hindi (hi)
• Previous version: tr-hi-s2st-v0.2

⚡ Training run

Run: v0.3-long-horizon-mh-r2 (xzcb60bl) — TPU v6e-16 (4 hosts × 4 chips, multi-host data-parallel), global batch 32, ~1.45 s/step, zero spot preemptions, config configs/tpu/stage2_tpu_v6e16_full_v03_mh.yaml. Ended by designed early stop at step 65,250 (patience 10 val cycles); 261 validation cycles over the run.

Validation metrics (teacher-forced, fixed 3,200-sample val gate — not the end-task release evals, which are reported in Evaluation below):

metric	step 250	plateau best (62,750)	annealed (76,000 = best ≈ final)
val composite (0.4·text + 0.6·audio)	6.719	2.9048	2.8199
val text loss / perplexity	4.328 / 75.8	0.486 / 1.63	0.398 / 1.489
val audio loss	8.313	4.518	4.435
val text token accuracy	25.6%	94.4%	96.6%
val cb0 (semantic codebook) accuracy	10.4%	40.4%	41.5%
val cb1–7 accuracies	10.5 → 0.1%	21.0 … 5.9%	22.0 / 18.2 / 12.1 / 9.2 / 7.5 / 6.4 / 6.1%

Anneal leg. The early stop fired on the WSD plateau, skipping the scheduled decay — so the run was resumed from step 65,250 (weights + Adam state + RNG) with max_steps extended to 76,250, placing the entire remaining horizon in the pre-registered 11,000-step linear LR descent 1.716e-4 → 0 (the shape and length our round-3 probe validated against cosine and plateau). Everything else stayed byte-identical; the same W&B run continues across both legs. The descent improved validation on essentially every 250-step cycle — the WSD "harvest" phase working as the literature predicts.

Every deep codebook is alive and far above the 0.05% chance floor — the deep-codebook collapse that capped v0.2 (cb0 ~14%, cb1–7 <4%) is resolved (coarse→fine unmask curriculum + per-codebook loss weights). Chart-reading notes: train/audio_loss shows upward steps at the curriculum onsets (cb2–cb7 activate at steps 1,579/3,157/4,735/6,313/7,891/9,469 — the metric's definition grows; per-codebook CEs actually drop at each onset). Use train/audio_loss_full (unweighted all-codebook mean, logged natively) for the jump-free audio learning curve.

🎧 Listen: audio samples (click ▶ to play)

Inline audio demos generated on the TPU during training every 5,000 steps — 4 s, greedy, free-running audio (text stream teacher-forced). Final milestone, full trio:

Step 65,000 — source (Turkish):

Step 65,000 — ground-truth target (Hindi, synthetic TTS):

Step 65,000 — model-generated translation:

Hear it learn — the same fixed sample generated at every 5,000-step milestone (source/target links per row):

step	generated	source / target
5,000		src / tgt
10,000		src / tgt
15,000		src / tgt
20,000		src / tgt
25,000		src / tgt
30,000		src / tgt
35,000		src / tgt
40,000		src / tgt
45,000		src / tgt
50,000		src / tgt
55,000		src / tgt
60,000		src / tgt
65,000		src / tgt

The same clips are browsable with a step slider in the W&B run's audio/ media panels.

📦 Checkpoints — the full training trajectory

Every checkpoint of the run is published (~87 training points): log-spaced early steps {1, 2, 4, …, 512}, every 1,000 steps from 1,000 → 76,000 (covering both the plateau and anneal legs), the annealed final step-76250, and 🏆 best (step 76,000, val composite 2.8199). Each is a complete weights-only bundle (peft_adapter/ + projection / depth-decoder / embeddings / audio heads + metadata.json with full provenance), available two ways:

• Browse in the file tree — no branch dropdown needed: checkpoints/ (e.g. checkpoints/best, checkpoints/step-76250)
• Git revisions (Pythia/OLMo convention) for programmatic loading:

  model = AutoModel.from_pretrained("tiny-aya-translate/tr-hi-s2st-v0.3",
                                    revision="best")   # or "step-42000", …
  

Trajectory landmarks (val composite): step 6,000 → 3.941 · 24,000 → 3.069 · 48,000 → 2.949 · 62,750 (plateau best) → 2.9048 · anneal onset 65,250 → 76,000 → 2.8199. Curriculum onsets, the WSD plateau, and the anneal descent are all visible across the suite — built for training-dynamics and mech-interp study, not just the final weights.

📊 Evaluation — a data-efficiency study of S2ST emergence

End-task release evals are complete. Full report + reproducibility: docs/v0.3-eval-report.md.

The question v0.3 answers (from the blog, written on a 26K-sample pilot): how much training on the full 840K dataset before translation quality emerges, not just language identity? v0.3 is that full-corpus run (2.07 epochs · 76,250 steps · 6.59B tokens). The answer: capability emerges in a clear order — language identity (early) → text translation (strong) → audio synthesis (the remaining frontier).

Interactive training charts: W&B run xzcb60bl.

Disclosures: greedy decoding; frozen digest-verified subsets v03-val-500 (in-domain) + v03-fleurs-200 (acoustic shift only — texts 200/200 seen, not held-out); references are MT-synthetic; chrF++ primary (BLEU unreliable < ~5); MOS as Δ(gen − GT) only; judges hi vasista22/whisper-hindi-large-v2 · tr openai/whisper-large-v3; every model ASR score is shown beside its GT-audio topline (the codec+ASR ceiling).

End-task @best (step 76,000 · 500-row v03-val-500 · greedy):

metric	hi→tr	tr→hi	reads as
free-run text chrF++ (inner-monologue)	25.7	25.1	the model translates
generated-audio ASR-chrF++	3.7	9.6	speech not yet ASR-intelligible
GT-audio topline chrF++	92.1	86.6	ceiling intact → pipeline sound
BLASER-2.0 QE (ASR-free, 1–5)	2.53	2.49	real, weak speech-semantic signal
GEMBA adequacy (gemini-3.6-flash, 1–5)	1.08	1.10	transcript ≈ no meaning
DNSMOS Δ(gen − GT)	−1.34	−1.34	low perceptual quality
RTF / TTFA	0.95 / 76 ms		faster than real-time (A100)

What this means — honestly. The text inner-monologue learns the translation mapping data-efficiently (96.6% teacher-forced, ~25 free-run chrF++). The generated audio carries a genuine but weak translation signal — BLASER-2.0 QE 2.5/5 is higher than the ASR/GEMBA metrics imply, so the audio is acoustically degraded rather than semantically empty. What has not yet emerged at this data/compute budget is intelligible speech synthesis (ASR-chrF++ 3.7–9.6 vs an 86–92 ceiling; DNSMOS −1.34) — the bottleneck is audio-codebook generation, bounded by the frozen Moshi depth decoder, not the translation understanding. Release checkpoint = @best (76,000): @final is a statistical tie and LAWA gives no gain (paired bootstrap). On FLEURS (real human speech, acoustic shift) even the text stream collapses to ~8 chrF++ → v0.3 is distribution-bound to its synthetic-TTS training acoustics. This is a first full-corpus run framed as what it is: a data-efficiency / emergence result and a training-dynamics study artifact, with speech-synthesis fidelity as the concrete next frontier.

See it learn — mel-spectrograms of the generated audio across training (companion to Hear it learn above; acoustic structure develops even before it is ASR-intelligible):

Intended use & limitations

• Intended: research on speech-to-speech translation, training-dynamics study over the checkpoint trajectory, and TR↔HI S2ST prototyping. Non-commercial only (CC-BY-NC-4.0, inherited from the base model).
• Not intended: production/commercial use, surveillance, or speaker impersonation. Training speech is synthetic multi-voice TTS (kokoro / XTTS-v2 / chatterbox) — no real-speaker cloning data — and output voices are those synthetic voices.
• Limitations: Turkish↔Hindi only; translation references are MT-synthetic (quality ceilings reflect that); Mimi operates at 12.5 Hz frames (80 ms granularity); the training-time demos use a 4 s generation window; the end-task release-eval numbers are reported in Evaluation above.

License & attribution

• Weights (this repo): CC-BY-NC-4.0 — derivative of CohereLabs/tiny-aya-base (CC-BY-NC-4.0). Depth-decoder and Mimi components derive from kyutai's Moshi (CC-BY-4.0; attribution hereby given).
• Training/eval code: Apache-2.0 — the GitHub repository.
• Some evaluation tools referenced by the harness (BLASER-2.0/SONAR, CometKiwi) are CC-BY-NC and are used for evaluation only; nothing from them ships in the weights.

Dataset (corrected from v0.2)

v0.3 trains on tiny-aya-translate/tr-hi-mimi-encoded — the project's synthetic pipeline: parallel text from FLORES, OPUS-100, and machine-translated conversational datasets, rendered with multi-voice TTS (kokoro / XTTS-v2 / chatterbox) into ~1.24M Mimi-encoded clips. After filtering ~5% of rows with missing .pt files: 1,178,302 train / 62,036 val. The corpus ships word-level text alignments for every sample ({stem}.{src,tgt}.alignments.json, 840,426 pairs, 100% coverage) → trained text+audio. Note for reimplementers: the alignment files live at the dataset root (not encoded/) under names that differ from the split manifests' src_align_path/tgt_align_path fields — v0.1–v0.2 missed them entirely because of this (silently zero text loss); our loader maps the names (src/data/dataset.py::_resolve_alignment).

Recipe (capacity-sweep winner)

Beyond the data-source fix, v0.3 carries codebase corrections and a recipe chosen by a two-stage capacity sweep on the full corpus (not the small-data anti-overfit tuning):

• Parallel-stream collator fix — v0.2's pre-fix collator dropped the model audio stream, so model_audio_embed received zero gradient. Restored in v0.3.
• Capacity sweep — Stage 1 (structural grid) chose +MLP target modules (q,k,v,o + gate,up,down + embed_tokens); Stage 2 (Bayesian lr × rank) chose lora_r=32, alpha=64, rsLoRA, lr_lora=1.716e-4. In the data-rich regime more LoRA capacity → lower loss (opposite of the small-data overfit regime). The final re-validation (below) then flipped exclude_top 2 → 0.
• Deep-codebook learning — per-codebook loss weighting; the frozen depth decoder's I/O layers train while its blocks stay frozen.
• Pipeline validated — an overfit gate (32-example train==val) memorizes all 8 codebooks to 89–98%. Note: an earlier per-codebook accuracy metric scored CB1–7 against the undelayed target and read a false ~0%; fixed — CB1–7 were always learning.

Long-horizon run config: configs/tpu/stage2_tpu_v6e16_full_v03_mh.yaml — 110,463 steps ≈ 3 real epochs at global batch 32 (2 rows/chip × 16 chips, multi-host data-parallel), WSD schedule (linear warmup 1100 → peak plateau → 11,000-step linear anneal to 0; a stop-anytime anneal template covers early stops).

¹ Batch-semantics correction (2026-07-12 audit): earlier configs (and the sweep table above) reported batch × accum × chips as "global batch 256" (batch-semantics). A live on-mesh audit proved the real optimizer batch is loader batch × accum — 32 for the reval arms and for this run. All v0.3 numbers in this card use the corrected semantics; the nominal-256 label is retained only where it names historical runs.

Recipe re-validation: 6-arm text+audio sweep (v03-5k-reval-ta, 2026-07-09)

Before the long-horizon run, the recipe was re-validated as text+audio on the full 1.24 M-pair corpus — 6 arms × 5,000 steps (≈1 epoch) at a nominal global batch 256 (batch-semantics note¹ — real 32), one v6e-8 per arm. Winner: arm D, lora_exclude_top: 0 — adapters on all 36 layers. The previously frozen champion (exclude_top=2) placed last at every composite weighting; the ranking E ≺ D ≺ C ≺ F ≺ B ≺ A is unanimous across text/audio weightings {0.2/0.8, 0.4/0.6, 0.5/0.5}, and D is the winner after the pre-registered cb0-accuracy gate (E and C fall

1 pt below best cb0). Headline science: top-layer adapters are the text lever — exclude_top=0 buys ~0.5 text CE at zero audio cost.

arm	delta	val text loss	val audio loss	composite (0.4/0.6)	W&B
D (winner)	exclude_top=0	1.181	4.985	3.464	0noyz5tr
E	dropout .10/wd .05	1.140	4.989	3.450 (cb0 gate ⚠)	7rb9pc85
C	r16, lr 2.4e-4	1.155	5.009	3.467 (cb0 gate ⚠)	rag7amc2
F	depth_unfreeze=2	1.476	4.953	3.562	jqozgc36
B	r64	1.566	4.958	3.601	2jtqcnla
A	frozen champion	1.692	4.960	3.653	powp1a50

All per-arm best_by_val checkpoints: gs://tinyaya-stage2-eu/checkpoints/stage2-reval-5k-ta/arm_{A..F}/best_by_val.

Training infrastructure: replicated strategy + XLA architecture changes

Parallelism = replicated (SPMD data-parallel), multi-host. The composite is 5.24B params total but only ~192M trainable (LoRA r=32 on all 36 layers incl. embed_tokens, projection, depth-decoder I/O), so the whole model is replicated on every TPU chip and only the data is sharded: the long-horizon run trains on a v6e-16 (4 hosts × 4 chips) where each host's DistributedSampler draws a disjoint corpus shard and the minibatch input pipeline assembles the global batch of 32 (2 rows/chip) across the 16-chip mesh — verified bit-exact by a gradient-identity probe; inter-host all-reduce costs ≤3% of the 1.8 s step. There is no tensor/FSDP sharding of weights in the released checkpoints — a checkpoint is a plain single-replica state and loads on one GPU without any resharding. (The trainer auto-selects replicated whenever trainable params < 500M; see src/backend/tpu_backend.py::_resolve_strategy.)

Architecture / lowering changes made to train this on TPU (all verified numerics-identical to stock; needed because XLA compiles static graphs and has no stride-0 broadcast views):

Change	Why	Inference impact
MoshiFlexibleLinear.forward rewritten as equal-batch bmm (src/model/depth_decoder.py::_patch_flexible_linear_bmm)	stock broadcast-batched matmul materialises the per-codebook weight per token on XLA (5.5 GiB/FFN call → OOM)	none on GPU (identical math); apply the patch if running inference on XLA
Identity-gather skip in the same patch (index_select(weight, arange(C)) → read weight directly)	the training path always selects ALL codebook rows; XLA copies the full weight per call otherwise	none (identical math)
Full-attention forcing under use_scan_layers (composite.py::_force_full_attention_for_scan)	Cohere2 interleaves sliding/full attention (sliding_window_pattern=4); scan_layers needs 36 homogeneous layers. Sliding window 4096 ≫ max seq 300 ⇒ identical	none — attention pattern is a config read at load; released config unchanged
LoRA adapters on ALL 36 layers, top-2 frozen (lora_setup.py::apply_lora(scan_homogeneous=True)) instead of exclude_top=2 omitting them	scan stacks per-layer param pytrees and requires identical keys	checkpoint-structural: peft_adapter/ contains 36 layers of adapters; the top-2 are zero (lora_B never trained) ⇒ mathematically identical to exclusion. Load with the shipped adapter_config.json, not a hand-written one
Scan-safe dropout (scan_utils.py::_ScanSafeDropout)	native_dropout's bool-mask meta vs bf16 XLA lowering breaks scan's stacked activation buffers	none — train-time only, eval-mode is a no-op
Per-micro-batch graph break (train.micro_mark_step) + depth_chunk_size	XLA buffer-assignment fragmentation (81 GiB "used" over 14 GiB real) when 8 grad-accum micros trace into one program	none — pure scheduling

Note for checkpoint consumers: only the bolded row changes what is in the checkpoint (extra zero adapters on the top layers). Everything else is training-time lowering. Runs trained without use_scan_layers (e.g. an unscanned v6e-16 run) keep the classic 34-layer adapter layout; metadata.json records which applies.

Pipeline validation (memorization gate, 2026-07-09)

Before the long-horizon run, the exact shipping stack (scan + all-36-layer adapter layout + FlexibleLinear bmm + text+audio objective) passed a 32-example memorization gate (train==val, regularization stripped, 800 steps) with an independent checkpoint-reload inference examination. W&B: v03-overfit-ta-scan.

check	result
CB0 teacher-forced accuracy	99.6% (train-val) / 99.5% (independent reload+eval)
CB1–7 TF accuracy	97.9 → 88.6% monotone — the frozen Moshi depth-decoder ceiling (only its I/O layers train); at parity with the pre-scan stack, i.e. no regression from the XLA changes
Text TF accuracy	99.8% (CE 0.187); decoded predictions character-identical to targets in both TR→HI and HI→TR
Per-component losses	all → ~0 (audio 0.021, text 0.187; all 8 per-CB losses collapsed)
Checkpoint→eval parity	per-CB within 0.1–0.6 pt (CB0–3); CB4–7 1.3–1.7 pt (metric weighting + fp32-CPU vs bf16-TPU precision)
Greedy AR reproduction	CB0 100.0%; all-CB match numerically identical to TF accuracy — the AR path reproduces the training-time forward

Disclosure: this gate caught an off-by-one in the evaluation harness's autoregressive loop (predictions shifted one frame and conditioned on a placeholder token). The model and training were never affected, but AR/ASR-BLEU numbers reported for earlier versions (v0.2 included) used the broken decoding and understate AR quality. Fixed in scripts/eval_checkpoint.py; all v0.3 release numbers use the corrected loop.

🚀 Release design: the checkpoint suite you will get

This repo (tiny-aya-translate/tr-hi-s2st-v0.3, now public) follows the Pythia/OLMo one-branch-per-checkpoint convention, weights-only (optimizer/scheduler/RNG stay in archival storage):

• The full suite is live — see Checkpoints above for the complete ladder and loading examples. Ops disclosure: during the private training phase, private-repo storage limits (~50 GB) meant only a 12-point interim ladder could be hosted; the flip to public (no such cap) enabled the full ~87-point publication, backfilled from the keep-all GCS archive.
• samples/step_NNNNNN/ on main: source / ground-truth-target / generated WAVs from the inline audio demo that runs on the TPU every 5000 steps — you can listen to the model improve across training.
• logs/train_host0_latest.log: rolling training-log snapshot.
• Every checkpoint's metadata.json carries provenance (git SHA of the exact deployed code, dataset digest rows/pt/al/md5, seed, global batch) and a byte-exact file manifest.

Full telemetry is on W&B — the completed run v0.3-long-horizon-mh-r2 (also via the release dashboard and the emergence report) carries losses, per-codebook prediction entropy + active-code fraction (the codebook-collapse instrument), perplexities, tokens-seen axes, MFU estimate, per-chip HBM for all 16 chips, and the audio demos. Post-hoc, each published checkpoint gains teacher-forced text chrF/BLEU backfilled at its own step (eval/*, via scripts/eval_translation_proxy.py).

Status checklist

Item	Status
Data source repointed to tr-hi-mimi-encoded	✅
Capacity sweep → recipe frozen (r=32/+MLP/rsLoRA)	✅
Pipeline validated (all 8 codebooks memorize)	✅
Long-horizon training run (plateau leg)	✅ completed 2026-07-19 (early stop @65,250; plateau best 2.9048 @62,750)
WSD anneal leg (65,250 → 76,250, linear LR→0)	✅ completed 2026-07-20 — best val composite 2.8199 @76,000
Repo public + full checkpoint suite (~87 points)	✅ (branches + checkpoints/ tree)
Audio samples + training log on main	✅ (13 milestones, playable above)
Release evals (ASR-chrF++ / MOS / BLASER / GEMBA / RTF)	✅ complete 2026-07-22 — data-efficiency study; @best (76,000) released; report docs/v0.3-eval-report.md

🙏 Acknowledgements

Trained on Cloud TPU v6e-16 provided by Google's TPU Research Cloud (TRC).

Citation

@misc{tinyaya_tr_hi_s2st_v0_3,
  title  = {TinyAya: Turkish-Hindi Speech-to-Speech Translation (v0.3)},
  author = {tiny-aya-translate},
  year   = {2026},
  note   = {Cohere2 + frozen Moshi depth decoder, LoRA (r=32, +MLP, rsLoRA); text+audio S2ST on the synthetic FLORES/OPUS/conversational corpus; Google TRC TPU v6e},
  url    = {https://huggingface.co/tiny-aya-translate/tr-hi-s2st-v0.3}
}