viai957/omnivoxtral-data

OmniVoxtral Data

This repository is the working-data backup of the OmniVoxtral project, a multilingual speech language model for Indic languages and English. The base model is Mistral-7B-v0.3, extended with a Mimi-codec audio vocabulary and a depth transformer. The repository holds three kinds of material:

• Training-ready token shards (tokens_v2): 155,861 interleaved text and audio sequences of 20 s each, with JSON sidecars.
• The VAD-chunked audio they were built from (chunks_v2, chunks_v2_quarantine), along with legacy audio dumps.
• Pipeline artefacts and experiment outputs: codec and tokenizer benchmarks, the SentencePiece tokenizer, the pinned validation split, the URL and search lists, the autoresearch ledger, and generated audio samples.

Rights warning: read before use. Most of the audio and pseudo-transcripts here come from YouTube recordings. The project does not hold redistribution rights to those recordings, and the project's own sidecars tag them data_license_class: "research_only". A sibling card in the same project states that YouTube-derived audio "must stay private" (viai957/omnivoxtral-yt-indic-raw, private). This repository is currently public. See Licence and rights.

All figures were generated from the repository's own files, or from the logs of the runs that consumed them. Every number in this card is traced in evidence.json (next to this README in the card build directory). Anything that was never measured is marked not measured.

---

1. TL;DR product sheet

Item	Value
What it is	A backup of the OmniVoxtral data tree. The training-ready part is data/tokens_v2.tar, holding Mimi-codec and SentencePiece token sequences for speech-LM pre-training
Consumer model	OmniVoxtral: a temporal transformer (Mistral-7B-v0.3, every 2nd layer pruned, LoRA r=128) plus a depth transformer over 8 Mimi codebooks
Token format	1,890 int64 tokens per clip = 90 windows × stride 21 (1 text token + 20 audio tokens per 200 ms window). Vocabulary is 81,920 = 65,536 text + 8 × 2,048 audio
Size on Hub	48 files, 191.87 GB (191,868,379,816 bytes)
tokens_v2	155,861 clips. 298.11 h of real speech (865.90 h including zero padding). 294,577,290 tokens
Sources inside tokens_v2	YouTube 123,239 clips (13 Indic languages + English). IndicVoices (Hindi) 20,151. FLEURS (6 languages) 12,471
Languages in tokens_v2	14: asm, ben, eng, guj, hin, kan, mal, mar, nep, ori, pan, tam, tel, urd
Transcripts	100% machine-generated (transcript_method = whisper_large_v3), including the IndicVoices and FLEURS clips. No human transcripts are used
Raw chunked audio	chunks_v2: 133,517 VAD chunks, 245.52 h, from 217 source recordings. chunks_v2_quarantine: 39,637 chunks, 35.78 h, rejected by LID confidence below 0.7
Validation split	val_split_v2.json: 420 clips (30 per language × 14). Not source-disjoint: see §6.3
Processing compute	Not recorded in a parseable log, so not measured
Reference training on this data	2 runs × 20,000 steps on 4 GPUs, each about 7.5 h wall time. Final val (run 5krth3fs): total 14.589, temporal 7.604, depth 6.984
Headline data-quality results	Whisper script consistency (sampled): asm 100% wrong-script, mal 89%, tel 30%, pan 27%. ori: 165/165 transcripts empty. 2,779 sidecars flagged repetition_loop
Headline downstream result	A model trained on this data (trial 22) reached free-run LID accuracy of 3/30 = 0.10 (95% Wilson CI 0.035–0.256), which is chance level for 10 languages
Upload	.hf_backup.py, 2026-06-04/05. Last commit f16d88f183131794ceb6b206c1e7bf7722fbff37

---

2. Dataset details and mechanism

2.1 What the token sequences encode

Each .npy in tokens_v2 is one 20 s audio chunk. The chunk is encoded as a single stream that interleaves text and audio:

• Text. Whisper-large-v3 transcribes the chunk with word timestamps. Words are bucketed at 5 Hz and tokenized with the project's 65,536-entry SentencePiece Unigram model (data/tokenizer/omnivoxtral_sp.model).
• Audio. The 24 kHz audio is encoded by Kyutai Mimi (12.5 Hz frames, 8 residual codebooks of 2,048 entries). Codebook q is offset by 65,536 + 2,048·q, so audio and text IDs never collide.
• Interleave. Each 200 ms window contains 1 text token followed by 20 audio tokens (2.5 frames × 8 codebooks), so the stride is 21.
• 2 s text/audio offset. The tokenizer drops the first 2 s of Mimi tokens (mimi_num_quantizers × frame_rate × 2 = 200 tokens) and the last 2 s of text tokens (text_hz × 2 = 10 tokens). That is a shift of 10 windows, not 2. Text for time t therefore sits next to the audio for t + 2 s, so text is delayed relative to audio. This is why 100 windows become 90, or 1,890 tokens, per 20 s chunk. The code comment reads "delay text and audio tokens by two windows … we avoid corresponding audio tokens appearing before the text tokens by accident". That comment does not match what the code does: as implemented, the audio for a moment appears in the stream 2 s before its text (src/voxtral/tokenizer/model.py, same at c416eb4).
• Padding mask. Chunks shorter than 20 s are zero-padded. valid_token_mask marks the real-audio prefix, and the trainer excludes padded positions from the loss.

flowchart LR
  subgraph Acquisition
    S[search terms<br/>searches*.txt] --> U[YouTube URLs<br/>indic_urls*.txt]
    U --> D[yt-dlp + ffmpeg segment<br/>chunks_indic_yt .m4a]
    HF1[google/fleurs] --> F[FLEURS clips]
    HF2[ai4bharat/IndicVoices] --> I[IndicVoices clips<br/>chunks_indic*]
  end
  D --> V[Silero VAD v6<br/>min speech 300 ms, min silence 500 ms,<br/>pad 150 ms, max 20 s]
  V --> L{MMS-LID-2048<br/>conf >= 0.7?}
  L -- yes --> C2[chunks_v2<br/>133,517 wav + json + lang.json]
  L -- no --> Q[chunks_v2_quarantine<br/>39,637]
  C2 --> M[music flags per source<br/>harmonicity + tempo]
  C2 --> W[Whisper-large-v3 word timestamps<br/>batched faster-whisper]
  F --> W
  I --> W
  W --> SP[SentencePiece Unigram 65,536<br/>5 Hz text buckets]
  C2 --> MI[Mimi encoder<br/>12.5 Hz x 8 codebooks]
  F --> MI
  I --> MI
  SP --> X[interleave: 1 text + 20 audio / 200 ms<br/>2 s text/audio offset, zero-pad to 20 s]
  MI --> X
  X --> T[tokens_v2 .npy + .meta.json<br/>155,861 clips]
  T --> RL[repetition-loop flag<br/>max trigram fraction >= 0.30]
  T --> VS[val_split_v2.json<br/>30 clips x 14 langs]
  T --> TR[OmniVoxtral training<br/>temporal Mistral + depth transformer]

2.2 Repository layout (what is actually on the Hub)

Everything is packed as tar archives because the source tree had more than 1M small files. chunks_v2 and autoresearch_loop were split with split -b 18G. Reassemble them with cat chunks_v2.tar.part.* | tar x.

Path in repo	Bytes	Contents	Provenance / licence class
data/tokens_v2.tar	5,946,265,600	155,861 × (.npy + .meta.json), layout tokens_v2/yt/<iso3 dir>/<2-hex shard>/<id>.* (Nepali and Odia directories are npi/ and ory/, although the sidecar language field says nep/ori), tokens_v2/fleurs/<iso1>/<iso1>/<id>.* (e.g. fleurs/hi/hi/), tokens_v2/iv/hi/…/<id>.*. See _index/tokens_v2.jsonl for exact paths	mixed (see §3.1)
data/chunks_v2/chunks_v2.tar.part.000–004	85,235,998,720	133,517 × (.wav 24 kHz mono + .json + .lang.json)	YouTube, research-only
data/chunks_v2_quarantine.tar	12,478,914,560	39,637 rejected chunks, same schema	YouTube, research-only
data/chunks_indic_yt.tar	18,703,032,320	57,668 raw 20 s .m4a chunks (pre-VAD input)	YouTube, research-only
data/chunks_indic.tar, chunks_indic_extra.tar	13,160,857,600 / 8,954,941,440	95,000 / 65,000 .flac files (count from the project's sharding README)	IndicVoices audio with no text, CC-BY-4.0
data/chunks_indic_g{0..3}.tar, chunks_indic_g1_pa.tar, chunks_indic_ml_fix.tar, chunks_g{1..3}_batch2.tar	0.28–0.57 GB each	smaller per-language batches (24,000 files together, per the sharding README)	IndicVoices audio with no text, CC-BY-4.0
data/codec_benchmark_results.tar	194,560	Mimi round-trip benchmark (raw_results.csv, summary.md, ANALYSIS.md)	project-created
data/tokenizer.tar	12,072,960	omnivoxtral_sp.model/.vocab, training_corpus.txt (FLEURS train transcripts), comparison.md	project-created; corpus text from FLEURS (CC-BY-4.0)
data/tokenizer_benchmark_results.tar	358,400	Mistral BPE fertility per language	project-created
data/dataset_inventory.tar	1,157,120	3,500-sample, 8.8 h survey of FLEURS and IndicVoices across 22 languages (2026-03-13)	project-created
data/indic_test_samples.tar	141,588,480	test WAVs used by the codec benchmark	FLEURS / IndicVoices / synthetic placeholders
data/yt_g{0..3}.tar	20,480 each	empty shard-directory skeletons (no files)	n/a
data/.cache.tar	10,240	empty cache stub	n/a
data/_loose_files.tar + top-level copies	about 0.43 MB	searches.txt (538 lines), searches_indic.txt (139), indic_searches.txt (193), indic_urls.txt (1,845 URLs), indic_urls_new.txt (1,501), indic_urls_all.txt (3,219), source_music_flags.json (+4 per-rank parts), val_split_v2.json	project-created; URLs point to third-party content
experiments/autoresearch_loop/*.tar.part.000–002	43,716,782,080	autoresearch loop code, per-trial run directories, SQLite ledger, rejected checkpoints	project-created
experiments/outputs.tar	32,317,440	4 sets of 14 generated WAV samples (one per language) from the two reference runs (§5)	model outputs
autoresearch_snapshot.tar.gz	13,611,385	snapshot of autoresearch/ (memory files, results.tsv with 528 experiment rows)	project-created

There is no datasets loading script, and the Hub viewer cannot read these archives. To use the data, download the tars and read the .npy and .json files directly (§9).

---

3. Sources, schema, splits, statistics

3.1 Sources and licences

Source	Where	Upstream licence	Sidecar tag	Notes
YouTube recordings (podcasts, interviews, talks) found with project search terms	chunks_indic_yt, chunks_v2, chunks_v2_quarantine, tokens_v2/yt	copyright of the original rights holders; no redistribution licence held	research_only	217 source recordings back chunks_v2. source_url is null in the sidecars, so per-clip provenance cannot be recovered from this repo alone
FLEURS (google/fleurs)	tokens_v2/fleurs	CC-BY-4.0	apache (a code default in the sidecar writer, not the upstream licence)	12,471 clips in 6 languages
IndicVoices (ai4bharat/IndicVoices)	tokens_v2/iv (Hindi), chunks_indic*, chunks_g*_batch2	CC-BY-4.0	cc-by-4.0	20,151 Hindi clips in tokens. The upstream human transcripts were not used; text was re-generated with Whisper

3.2 Schema

tokens_v2/<group>/<lang>/<shard>/<id>.npy holds int64[1890]. Position k with k mod 21 = 0 is a text token (ID < 65,536). The other 20 positions are audio tokens (codebook q in [65,536 + 2,048q, 65,536 + 2,048(q+1))).

<id>.meta.json (schema_version 2). An abridged real example from tokens_v2/yt/hin/00/00045e3004d24a8ab24fc07126704adc.meta.json:

{
  "schema_version": 2, "preprocessing_version": "v2.0",
  "source": "youtube", "source_id": "00045e3004d24a8ab24fc07126704adc", "source_url": null,
  "language": "hin", "language_confidence": 0.9846662282943726, "language_method": "mms_lid_2048",
  "transcript": " पहला पास साल ऐसा रहा लोगों ने धूमधाम से पूजा किया, लोगों ने धूमधाम से किर्तन किया.",
  "transcript_method": "whisper_large_v3", "transcript_avg_logprob": -0.05110206234073916,
  "duration_s": 20.0, "sample_rate": 24000, "num_channels": 1, "stream_layout": "single",
  "snr_db": null, "speech_ratio": null, "clip_ratio": null, "music_likely": null,
  "tokenizer_config": {"sp_model": "omnivoxtral_sp.model", "text_hz": 5, "mimi_num_quantizers": 8, "stride": 21},
  "token_count": 1890, "valid_token_mask": [true, "... 1890 booleans ..."],
  "data_license_class": "research_only", "quarantine_reason": null, "task": "transcribe"
}

duration_s is always the padded length (20.0). To get the real speech length, use duration_s × mean(valid_token_mask).

chunks_v2/yt/unknown/<shard>/<id>.json holds source, source_id (<recording-uuid>_<segment>), chunk_index, start_s_in_source, end_s_in_source, duration_s, num_channels, stream_role, and language (always null here). <id>.lang.json holds language, confidence, method (mms_lid_2048), secondary, agreement, and chunk_duration_s.

val_split_v2.json holds seed (1729), per_lang_target (30), min_lid_confidence (0.7), per_language, per_language_per_source, missing_tier1 (san, sin, snd, bod, doi, kok, mni, sat, kas), and files (420 paths).

3.3 Splits

Split	Clips	Definition
train	155,441	every tokens_v2 file not in val_split_v2.json
val	420	30 per language × 14 languages, sampled at clip level with seed 1729 and LID confidence ≥ 0.7. Mix per language: hin = 10 FLEURS + 10 IV + 10 YT; ben/kan/mal/tam/tel = 15 FLEURS + 15 YT; all others = 30 YT
test	none	no held-out test split exists

3.4 Statistics (tokens_v2)

Figure 1. Left: hours of real speech per language (sum of duration_s × mean(valid_token_mask)). Right: clip counts. Source: every *.meta.json inside data/tokens_v2.tar, indexed by streaming the archive from this repo (_tools/stream_index.py, _tools/stats_tokens.py → _stats/tokens_v2_stats.json).

Group	Lang	Clips	Speech h	Padded h	Median LID conf	Empty transcripts
yt	tam	27,732	56.74	154.07	0.9997	5
yt	hin	25,696	43.93	142.76	0.9785	0
yt	ben	20,322	28.17	112.90	0.9999	0
yt	asm	9,642	20.06	53.57	0.9998	4
yt	kan	6,828	14.38	37.93	0.9991	1
yt	urd	6,053	12.94	33.63	0.9473	0
yt	tel	6,694	12.30	37.19	0.9966	0
yt	mar	5,857	11.12	32.54	0.9995	0
yt	pan	4,846	8.18	26.92	0.9932	0
yt	guj	3,453	8.02	19.18	0.9997	0
yt	mal	2,731	5.02	15.17	0.9998	0
yt	eng	2,838	4.66	15.77	0.9969	0
yt	nep	382	0.71	2.12	0.9997	0
yt	ori	165	0.46	0.92	0.9881	165
iv	hin	20,151	29.64	111.95	(dataset label)	0
fleurs	mal	2,931	9.40	16.28	(dataset label)	0
fleurs	tam	2,367	8.41	13.15	(dataset label)	0
fleurs	kan	2,283	8.07	12.68	(dataset label)	0
fleurs	tel	2,302	7.75	12.79	(dataset label)	0
fleurs	hin	2,120	6.51	11.78	(dataset label)	0
fleurs	ben	468	1.64	2.60	(dataset label)	0
total	14 langs	155,861	298.11	865.90

The real speech per 20 s window has a mean of 6.62 s for YT (median 4.89), 5.30 s for IV (median 3.78) and 12.07 s for FLEURS (median 11.56). Across the corpus, about 66% of token positions are padding.

Figure 2. Left: real speech seconds per window by source (log y). Middle: MMS-LID-2048 confidence for the 123,239 YouTube clips (minimum 0.70002; the 0.7 quarantine gate is visible). Right: Whisper transcript_avg_logprob for the 120,786 YT clips with a non-zero value (median −0.1506; the remaining 2,453 store 0.0 or nothing, because of a faster-whisper adapter bug documented in the project notes). Source: _index/tokens_v2.jsonl → _stats/tokens_v2_distributions.json.

3.5 Statistics (chunks_v2 and quarantine)

Figure 3. Left: VAD chunk durations from the .json sidecars (accepted vs quarantined, log y). Right: top-25 MMS-LID top-1 labels over accepted chunks (log y). Source: chunks_v2 sidecars and chunks_v2_quarantine.tar (streamed from this repo) → _stats/chunks_stats.json.

	chunks_v2 (accepted)	chunks_v2_quarantine
Chunks	133,517	39,637
Audio hours	245.52	35.78
Mean / median duration (s)	6.62 / 4.90	3.25 / 1.89
Min / max duration (s)	0.2 / 20.0	0.1 / 20.0
Median LID confidence	0.9984	0.4026
Chunks with LID confidence < 0.7	0	39,637 (all)
Distinct LID top-1 labels	197	702
Source recordings	217	217

The accepted chunks are spread across 197 distinct LID labels. Only 14 languages were tokenized. Accepted chunks in other labels were not tokenized, for example san 2,971, npi 2,900 (of which 382 tokenized as nep), ory 1,259 (165 tokenized as ori), sin 531, snd 462 and ara 317. Of the 133,517 accepted chunks, 123,239 appear in tokens_v2/yt.

The sidecars come from the local chunks_v2 tree, which produced the uploaded tar. A partial re-index of the uploaded tar found byte-identical sidecars: 34,843 of 34,843 compared (26% of the archive) were identical and none were missing locally. The quarantine numbers come from the uploaded tar itself.

3.6 Language mixture used in training

Figure 4. Natural share of training clips vs the temperature-sampled share π_l ∝ n_l^{1/τ} with τ = 3.3 (the VoxtralDataset default at the reference-run commit c416eb4). Source: _stats/sampling_weights.json, computed from the clip counts in Figure 1 minus val_split_v2.json.

Temperature sampling compresses Hindi from 30.8% to 12.8% of draws and raises Odia from 0.09% to 2.2%. Odia's 135 training clips all have empty transcripts.

3.7 Filtering and decontamination

• VAD. Silero VAD v6 with min speech 300 ms, min silence 500 ms, speech pad 150 ms and max segment 20 s. Mono mixdown applies unless the stereo L/R correlation is below 0.5 (scripts/vad_chunker.py).
• LID gate. MMS-LID-2048 runs on every chunk. A chunk is quarantined if its confidence is below 0.7 or the primary and secondary predictions disagree (scripts/detect_language.py).
• Music. Per-source harmonicity plus tempo heuristic over 217 sources, of which 49 are flagged music_likely (source_music_flags.json). The flag was intended to gate Demucs. It is not applied as a training filter, and music_likely is null in every sampled token sidecar.
• Repetition loops. 2,779 token sidecars carry quarantine_reason = "repetition_loop" (max trigram fraction ≥ 0.30). The loader at the reference-run commit c416eb4 has no quarantine filter. A later commit (9015984) adds one, so whether the §5 runs excluded these clips cannot be verified.
• SNR / speech-ratio / clipping. The snr_db, speech_ratio and clip_ratio fields exist but are null in all 155,861 sidecars (full scan of _index/tokens_v2.jsonl), so they were not measured.
• Decontamination against public benchmarks. Not performed. FLEURS train and test clips may both appear in tokens_v2/fleurs, because the split of origin is not stored in the sidecars. The tokenizer was trained on FLEURS train and benchmarked on FLEURS test.

---

4. Processing procedure and the objective this data feeds

4.1 Processing stages (as run)

Stage	Tool / setting	Output
Index	project search terms → YouTube search (min duration 30 min)	indic_urls*.txt
Scrape	yt-dlp + ffmpeg segment (stream copy, no transcode)	chunks_indic_yt (57,668 × 20 s .m4a)
VAD chunk	Silero VAD v6 (settings in §3.7), resample to 24 kHz	chunks_v2
LID	MMS-LID-2048, gate at 0.7	.lang.json, chunks_v2_quarantine
ASR + alignment	Whisper-large-v3 word timestamps. YouTube used batched faster-whisper (8 same-language clips concatenated per call); Odia is unsupported by faster-whisper, so its transcripts are empty	transcript
Text tokenizer	SentencePiece Unigram, vocab 65,536, character_coverage 0.9999, trained on FLEURS-train transcripts (training_corpus.txt, 8.56 MB); 23 language tags (IDs 4–26) + 6 control tokens	omnivoxtral_sp.model
Audio tokenizer	Kyutai Mimi, 8 of its RVQ codebooks, 12.5 Hz	audio tokens
Interleave	5 Hz text, stride 21, 2 s text/audio offset (first 2 s audio + last 2 s text dropped), zero-pad to 20 s, valid_token_mask	tokens_v2
Post-hoc QC	repetition-loop flag (trigram ≥ 0.30), pinned val split	sidecar fields, val_split_v2.json

Index timestamps: the newest tokens_v2 member is dated 2026-05-06 09:49 UTC, and the newest chunks_v2 member 2026-05-05 18:01 UTC. Processing compute and wall time were not measured. The project notes give throughput for batched faster-whisper (113 files/min/rank in a smoke test), but there is no end-to-end log.

4.2 Training objective that consumes these tokens

The objective below is transcribed from src/voxtral/trainer/omni_trainer.py at commit c416eb4, which is the commit the §5 runs logged.

Let $x_{1..N}$ be the 1,890-token sequence and $m_t \in {0,1}$ the valid_token_mask. The temporal transformer predicts $x_{t+1}$. Position weights are indexed by target position $k$ (full_weights[1:]) and repeat a stride-21 pattern

$${\mathcal{L}}_{\text{temporal}}=\frac{\sum _t{m}_t{\tilde{\omega }}_t(-\log p_{\theta }(x_{t+1}\mid x_{\le t}))}{\sum _t{m}_t}\backslash mathcal\; L\_\{\backslash text\{temporal\}\}=\backslash frac\{\backslash sum\_t\; m\_t\backslash ,\backslash tilde\backslash omega\_t\backslash ,\backslash bigl(-\backslash log\; p\_\backslash theta(x\_\{t+1\}\backslash mid\; x\_\{\backslash le\; t\})\backslash bigr)\}\{\backslash sum\_t\; m\_t\}$$

At each audio frame $f$, the depth transformer predicts the 8 codebooks autoregressively, using focal cross-entropy with $\gamma = 2$, a per-codebook weight $\alpha_0 = 100$, $\alpha_{1..7} = 1$, and RVQ dropout applied only during training, with $d_0 = 1$ (q0 is never dropped, drop_mask[0] = True) and $d_{1..7} \sim \text{Bernoulli}(0.7)$:

$${\mathrm{\ell }}_q=\frac{\sum _f{m}_f(1-p_{f,q}{)}^{\gamma }(-\log p_{f,q})}{\sum _f{m}_f},{\mathcal{L}}_{\text{depth}}=\frac{\sum _q{\alpha }_q{d}_q{\mathrm{\ell }}_q}{\sum _q{\alpha }_q{d}_q},\mathcal{L}={\mathcal{L}}_{\text{temporal}}+{\mathcal{L}}_{\text{depth}}\backslash ell\_q=\backslash frac\{\backslash sum\_f\; m\_f\backslash ,(1-p\_\{f,q\})^\{\backslash gamma\}\backslash ,(-\backslash log\; p\_\{f,q\})\}\{\backslash sum\_f\; m\_f\},\backslash qquad\; \backslash mathcal\; L\_\{\backslash text\{depth\}\}=\backslash frac\{\backslash sum\_q\; \backslash alpha\_q\; d\_q\backslash ,\backslash ell\_q\}\{\backslash sum\_q\; \backslash alpha\_q\; d\_q\},\backslash qquad\; \backslash mathcal\; L=\backslash mathcal\; L\_\{\backslash text\{temporal\}\}+\backslash mathcal\; L\_\{\backslash text\{depth\}\}$$

Language mixing is $\pi_l = n_l^{1/\tau} / \sum_{l'} n_{l'}^{1/\tau}$ with $\tau = 3.3$.

No preference optimisation, RL, reward model or PRM was applied to this data. Those methods do not apply to this repository.

4.3 Reference-run hyperparameters (from W&B config.yaml)

Hyperparameter	5krth3fs	dq7kimwg
Base	mistralai/Mistral-7B-v0.3, prune_layers=2 (temporal 4,550,955,008 params)	same
Adapter	LoRA r=128 (no DoRA)	same
Depth transformer	4 layers, 16 heads, dim 1024 (88,351,744 params)	same
loss_weights [text, sem, ac]	[100, 1, 1]	[10, 1, 1]
Depth	focal γ=2, mask rate 0.15, q-dropout 0.3, label smoothing 0	same
Optimiser	AdamW β=(0.9, 0.95), ε=1e-9, weight decay 0	same
LR	peak 3e-4, warm-up 500, decay to 0 at 20,000 (as logged)	peak 1e-4, same schedule
Batch	1 per rank × 4 ranks (DDP), grad-accum 1	same
Grad clip	1.0	same
Steps	20,000	20,000
Data	./data/tokens_v2 + pinned val_split_v2.json; val every 500 steps on 50 batches	same
Precision / memory	bf16, gradient checkpointing, no torch.compile	same
Seed	42	42
Hardware	4 GPUs. W&B records "NVIDIA A10G"; a later audit of this host (2026-10-25) found 4× NVIDIA L4	same
Wall time	26,936 s (7.48 h)	26,922 s (7.48 h)
Code	c416eb4bd6ed7e465d3c6f8bfe7160a592326f62 (W&B git record)	same

---

5. Training curves (reference runs on this data)

Figure 5. The two 20k-step OmniVoxtral runs on data/tokens_v2. Thin lines: train losses averaged over 50 steps. Thick lines: val every 500 steps. The bottom panels show the logged LR, depth-transformer token accuracy (q0 and q1–7), and text-token accuracy. Sources: logs/wandb/run-20260506_201931-5krth3fs/files/output.log and logs/wandb/run-20260507_040612-dq7kimwg/files/output.log, parsed by _tools/plot_curves.py → _stats/reference_runs.json. The total and temporal losses are weighted by loss_weights and are not comparable across the two runs. The depth loss uses identical settings in both runs and is comparable.

Metric	5krth3fs	dq7kimwg
val total @500 → @20k	14.813 → 14.589	14.996 → 10.609
best val total (step)	14.209 (5,000)	10.596 (15,500)
val temporal @20k (best, step)	7.604 (7.360, 16,000)	5.682 (5.674, 14,500)
val depth @20k (comparable)	6.984	4.927
train depth-acc q1–7 @20k (max)	0.307 (0.346)	0.328 (0.367)
train depth-acc q0 @20k	0.182	0.242
train text-token accuracy (max over run)	0.000 (0.001)	0.000 (0.001)
50-step windows with grad_norm=inf	400/400	400/400

There is also a model-side hill-climb over the project's experiment ledger, which ships in autoresearch_snapshot.tar.gz:

Figure 6. total_loss for the 522 numeric rows of autoresearch/memory/results.tsv (495 keep, 21 discard, 3 crash, 3 oom_risk), with the running best of kept rows. The rows mix fake-data smoke tests, 1-layer to 7B models, the earlier v1 token set and changing loss definitions, so this is a chronology of logged numbers, not a controlled hill-climb on this dataset. Source: autoresearch_snapshot.tar.gz → _stats/autoresearch_ledger.json.

---

6. Evaluation

6.1 Codec fidelity on Indic speech (Mimi round trip)

Figure 7. Mimi encode/decode with 4 and 8 codebooks on 10 clips per language for the 13 real languages. PESQ is wideband at 16 kHz. Source: data/codec_benchmark_results/raw_results.csv (2026-03-13). Nine further languages in the same file are synthetic sine placeholders and are excluded.

Metric (mean over 13 langs, n=10 each)	4 codebooks	8 codebooks (used)
PESQ	1.44	1.85 (range 1.41 ori – 2.16 guj)
F0 correlation	0.892	0.943 (min 0.831 nep)
Energy correlation	0.963	0.983

The benchmark's own pass threshold (PESQ ≥ 3.0) was met by 0/22 languages. The project kept Mimi at 8 codebooks after recalibrating against Moshi's reported English PESQ of 2.45, as cited in ANALYSIS.md. No subjective listening test (MOS) was run, so it is not measured. Confidence intervals were not computed (n=10 per cell).

6.2 Text-tokenizer fertility

Figure 8. Mean tokens per whitespace word on up to 50 FLEURS-test sentences per language. The SentencePiece model was trained only on FLEURS-train transcripts. Source: data/tokenizer/comparison.md.

Mean fertility drops from 10.79 (Mistral BPE, 32,768 vocab) to 1.91 (OmniVoxtral SP, 65,536). The largest gains are for Malayalam (22.20 → 2.49) and Odia (18.19 → 2.05).

6.3 Transcript and split quality

Figure 9. Fraction of sampled YouTube transcripts whose characters are mostly outside the expected script (phase-0 audit, 2026-05-07; n per bar). Source: autoresearch_loop/runs/phase0_audit.json (inside experiments/autoresearch_loop).

Check	Result	Source
Wrong-script transcripts (sampled)	asm 33/33 (1.00), mal 59/66 (0.89), tel 20/66 (0.30), pan 9/33 (0.27), guj 4/33 (0.12), ben 4/66 (0.06), hin 3/99 (0.03); eng, kan, mar, tam, urd 0	phase0_audit.json
Empty transcripts	ori 165/165; tam 5, asm 4, kan 1	tokens_v2 sidecars
Repetition-loop transcripts	2,779 flagged (1.8% of clips)	tokens_v2 sidecars
transcript_avg_logprob 0.0 or absent (adapter bug)	2,453 of 123,239 YT clips; all IV and FLEURS clips	tokens_v2 sidecars
Val split source-disjointness	Fails. The 325 YT val clips come from 134 source recordings. All 134 also contribute training clips, and 94,875 of 122,914 YT training clips share a recording with some val clip	val_split_v2.json × chunks_v2 source_id
Human-transcript WER of the pseudo-labels	not measured	none
Speaker / gender / dialect balance	not measured (no speaker labels)	none

6.4 Downstream: a model trained on this data

Model (all trained on tokens_v2)	Eval	n	Result
5krth3fs, step 20k	val losses (in-recording, see 6.3)	50 batches	total 14.589, temporal 7.604, depth 6.984
dq7kimwg, step 20k	same	50 batches	total 10.609, temporal 5.682, depth 4.927
trial 22 (PURE-NLL: loss weights [1,1,1], γ=0, LR 1e-4), step 2,000	free-run generation, Whisper LID on generated audio, 10 languages × 3 seeds, T=0.8, top-k 250	30	LID accuracy 0.10 (3/30; 95% Wilson CI 0.035–0.256; the only correct language was English, 3/3). Silent rate 0.40. Non-empty ASR transcript rate 0.90
5krth3fs / dq7kimwg generated samples	none	none	audio is in experiments/outputs.tar; not scored

No standard speech benchmark (FLEURS ASR WER, TTS MOS, Full-Duplex-Bench, and so on) was run on models trained from this data, so none is reported.

---

7. Discussion: negative results and lessons

1. The pseudo-labels are the weakest link. All text in tokens_v2 is Whisper-large-v3 output, including the IndicVoices clips, whose human transcripts were available upstream and were not used. For Assamese, the sampled transcripts were 100% in the wrong script (33/33 sampled; examples include English, Gurmukhi and Devanagari text), Malayalam was 89%, and Odia transcripts are empty. The text stream for these languages teaches the wrong target. Mask or re-transcribe them before you rely on text loss.
2. Two-thirds of the sequence is padding. Fixed 20 s windows around VAD chunks with median length 4.9 s leave about 66% of positions masked. Packing several chunks per window would roughly triple the useful tokens per step.
3. The val split measures memorisation of recordings, not generalisation. All 217 recordings appear in train, and every YT val clip shares its recording with training clips. Lower val loss on this split should not be read as transfer to new speakers or channels. Rebuild the split by source_id prefix.
4. Concentration. Only 217 source recordings back 123,239 YT clips. Tamil and Hindi together are 50% of clips (natural share), and nep and ori have 382 and 165 clips. Temperature sampling (τ = 3.3) up-weights the tails, but it cannot add speakers.
5. Loss went down while the model learned no language identity. Across more than 500 logged trials on this corpus family, teacher-forced loss fell while free-run LID stayed at chance (§6.4). Project diagnostics traced this to a model-side defect, not a data defect: the SentencePiece language-tag rows (IDs 10–26) inherited all-zero Mistral embedding rows and were never initialised. The lesson for users of this data is that val loss alone is not evidence of a usable model. Gate on a generation-side metric.
6. grad_norm=inf in every logging window of both runs. This happens under both loss_weights=[100,1,1] (5krth3fs) and [10,1,1] (dq7kimwg, 400/400 windows), so the weights are not the cause. A later root-cause analysis (2026-05-23, _safe_grad_norm docstring in omni_trainer.py at 9015984) found that the depth-transformer backward injects a gradient of about 5e33 into the temporal embedding. The fp32 sum of squares then overflows to inf. The isfinite skip guard active at run time would then skip optimizer.step(). The curves shown may therefore reflect far fewer optimiser updates than the step count suggests. The eventual fix is overflow-proof per-parameter fp64 clipping (9015984). Text-token accuracy never left 0.000–0.001 in either run.
7. The codec is a ceiling. Mimi at 8 codebooks reaches mean PESQ 1.85 on Indic read speech. Any model trained on these tokens inherits that reconstruction ceiling.

---

8. Intended use, limitations, risks

Intended use. This is an internal research backup for reproducing OmniVoxtral speech-LM experiments: inspecting the token format, reproducing the reference runs, and auditing the pipeline.

Out of scope. Do not use it for:

• redistributing the YouTube-derived audio or transcripts;
• commercial training;
• evaluating ASR quality, since the labels are machine-generated;
• speaker identification or any use that profiles individuals.

Limitations.

• Machine transcripts only.
• Weak language labels: LID comes from audio, and the chunks_v2 sidecars carry language: null.
• No speaker, gender or dialect metadata.
• source_url is null, so YouTube provenance is not recoverable per clip.
• No SNR or music filtering applied.
• The val split is not source-disjoint.
• Nine Tier-1 languages are missing from tokens: san, sin, snd, bod, doi, kok, mni, sat, kas.

Risks.

• Copyright. The YouTube audio belongs to third parties. Hosting it publicly redistributes it.
• Privacy. The recordings contain identifiable voices of people who did not consent to inclusion.
• Bias. The corpus is skewed toward the content and speakers of 217 recordings.
• Mislabelled licences. FLEURS sidecars carry apache, while the upstream licence is CC-BY-4.0.

---

9. Reproducibility

# download (use a disk with >200 GB free; never / or /tmp)
huggingface-cli download viai957/omnivoxtral-data --repo-type dataset --local-dir ./omnivoxtral-data

# reassemble split archives
cat data/chunks_v2/chunks_v2.tar.part.* | tar -x
cat experiments/autoresearch_loop/autoresearch_loop.tar.part.* | tar -x
tar -xf data/tokens_v2.tar

# read one clip
python - <<'EOF'
import json, numpy as np
p = "tokens_v2/yt/hin/00/00045e3004d24a8ab24fc07126704adc"
x = np.load(p + ".npy"); m = json.load(open(p + ".meta.json"))
text = x[0::21]; audio = x.reshape(90, 21)[:, 1:]        # 90 windows x (1 text + 20 audio)
speech_s = m["duration_s"] * np.mean(m["valid_token_mask"])
print(x.shape, m["language"], round(speech_s, 2), m["transcript"][:60])
EOF

# reference run 5krth3fs, reconstructed from its logged config.yaml (OmniVoxtral repo near c416eb4; env-var config via pydantic-settings)
DATA_PATH=./data/tokens_v2 MISTRAL_PRETRAINED_PATH=mistralai/Mistral-7B-v0.3 PRUNE_LAYERS=2 \
LORA_RANK=128 LR=3e-4 LOSS_WEIGHTS='[100,1,1]' MAX_STEPS=20000 WARMUP_STEPS=500 WEIGHT_DECAY=0 \
DEPTH_NUM_LAYERS=4 DEPTH_FOCAL_GAMMA=2.0 DEPTH_MASK_RATE=0.15 DEPTH_Q_DROPOUT=0.3 GRADIENT_CHECKPOINTING=true SEED=42 \
torchrun --nproc_per_node=4 scripts/train_omni.py

Artefact	Identifier
Hub revision documented here	f16d88f183131794ceb6b206c1e7bf7722fbff37 (2026-06-05)
Upload script	.hf_backup.py (OmniVoxtral repo root) and .hf_backup.log
Code that trained the reference runs	c416eb4bd6ed7e465d3c6f8bfe7160a592326f62, as recorded by W&B. The working tree was not clean: scripts/retokenize_v2_batched.py, which produced the YT tokens, is first committed in 9015984. The non-finite-gradient step skip was first committed in 56ae3c6 (2026-05-13), but it was an uncommitted working-tree edit when the May 6–7 runs were launched. Treat c416eb4 as approximate
Code snapshot at upload time	90159847a4625420465fb99ee82a6719e95f3b82 (2026-06-05, adds the quarantine-aware loader and quarantine_repetitions.py)
Pipeline scripts	scripts/vad_chunker.py, scripts/detect_language.py, scripts/detect_music.py, scripts/retokenize_v2_batched.py, scripts/quarantine_repetitions.py, scripts/build_val_split.py, scripts/train_tokenizer.py, scripts/benchmark_codec.py
Card statistics	_tools/stream_index.py, _tools/stats_tokens.py, _tools/stats_chunks.py, _tools/plot_*.py; outputs in _stats/*.json

The token shards can be regenerated only partially. The YouTube source audio has no stored URLs, and the IndicVoices and FLEURS text would differ if re-transcribed with a different Whisper build.

---

10. Related repositories, licence, citation

Sibling repositories (viai957).

• viai957/omnivoxtral-checkpoints (private): checkpoints from runs on this data.
• viai957/omnivoxtral-yt-indic-raw (private): WebDataset re-shard of the YouTube audio (chunks_v2, chunks_indic_yt) with a parquet manifest.
• viai957/omnivoxtral-indic-artifacts (private): project-created artefacts only (codec and tokenizer benchmarks, tokenizer, synthetic duplex).
• viai957/indic-duplex-speech-artifacts (public).

Licence and rights

This repository has no single licence, so the YAML uses license: other.

• YouTube-derived material (chunks_indic_yt, chunks_v2, chunks_v2_quarantine, tokens_v2/yt, indic_urls*.txt). Copyright remains with the original rights holders, and the project holds no licence to redistribute it. It is research-only. The project's own guidance for this material is that it stays private.
• IndicVoices-derived material (tokens_v2/iv, chunks_indic*, chunks_g*_batch2). CC-BY-4.0, © AI4Bharat. Attribution required.
• FLEURS-derived material (tokens_v2/fleurs, data/tokenizer/training_corpus.txt). CC-BY-4.0, © Google. Attribution required.
• Project-created material (benchmarks, tokenizer model, scripts, ledgers, split files). Released by the project. The upstream OmniVoxtral/Voxtral code is MIT.
• Mimi codec (Kyutai), used to produce the audio tokens: see the Kyutai model licence.

Citation

@misc{omnivoxtral_data_2026,
  title        = {OmniVoxtral Data: interleaved Mimi/SentencePiece token corpus for Indic speech-language modelling (working backup)},
  author       = {{viai957}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/viai957/omnivoxtral-data}},
  note         = {Revision f16d88f183131794ceb6b206c1e7bf7722fbff37}
}
@article{conneau2022fleurs,
  title   = {FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech},
  author  = {Conneau, Alexis and Ma, Min and Khanuja, Simran and Zhang, Yu and Axelrod, Vera and Dalmia, Siddharth and Riesa, Jason and Rivera, Clara and Bapna, Ankur},
  journal = {arXiv preprint arXiv:2205.12446}, year = {2022}
}
@article{javed2024indicvoices,
  title   = {IndicVoices: Towards building an Inclusive Multilingual Speech Dataset for Indian Languages},
  author  = {Javed, Tahir and others},
  journal = {arXiv preprint arXiv:2403.01926}, year = {2024}
}
@article{defossez2024moshi,
  title   = {Moshi: a speech-text foundation model for real-time dialogue},
  author  = {D{\'e}fossez, Alexandre and others},
  journal = {arXiv preprint arXiv:2410.00037}, year = {2024}
}
@article{radford2022whisper,
  title   = {Robust Speech Recognition via Large-Scale Weak Supervision},
  author  = {Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg and McLeavey, Christine and Sutskever, Ilya},
  journal = {arXiv preprint arXiv:2212.04356}, year = {2022}
}
@article{pratap2023mms,
  title   = {Scaling Speech Technology to 1,000+ Languages},
  author  = {Pratap, Vineel and others},
  journal = {arXiv preprint arXiv:2305.13516}, year = {2023}
}