Discrete Audio Pretrain
Collection
Pretraining Discrete Audio Models
•
10 items
•
Updated
Yodas2-Mimi Pretraining Dataset
Dataset Summary
Based on Yodas2, this dataset contains interleaved (at the utterance level) text-audio documents derived from Yodas2-Mimi tokenized. It is designed for pretraining audio language models that can process both text and audio in a unified format.
The audio is represented as unicode strings (converted from Mimi audio codec tokens), making it compatible with standard language model training pipelines.
Tokenizer
Since text column contains special tokens, you should use this tokenizer: marin-mimi-bpe-8cb-16k-tokenizer.
Data Format
Each document in the dataset contains:
id: Unique identifier for the document (suffixed with_type1or_type2)split: Source shard/subshard identifier (format:{shard_id}/{subshard_id})text: The interleaved text-audio content as a single string with special tokens included already
Two Document Types
The dataset provides two types of interleaved formats for each source document:
Type 1: Text → Audio (text comes first)
<|begin_of_text|><|text_start|>text1<|text_end|><|audio_start|>audio1<|audio_end|>...<|end_of_text|>
Type 2: Audio → Text (audio comes first)
<|begin_of_text|><|audio_start|>audio1<|audio_end|><|text_start|>text1<|text_end|>...<|end_of_text|>
This dual format enables training models for both text-to-speech and speech-to-text tasks simultaneously.
Special Tokens
The following special tokens are used to structure the documents:
<|begin_of_text|>: Marks the start of a document<|end_of_text|>: Marks the end of a document<|text_start|>: Marks the start of a text segment<|text_end|>: Marks the end of a text segment<|audio_start|>: Marks the start of an audio segment<|audio_end|>: Marks the end of an audio segment
Audio Encoding
Audio is encoded using the Mimi audio codec with the following configuration:
- Number of codebooks: 8
- Codebook size: 2048
- Unicode offset:
0xe000(private use area)
Audio tokens are converted to Unicode characters, allowing them to be processed as text by language models. Each audio code is mapped to a unique Unicode character starting from the private use area.
Dataset Statistics
- Documents (Rows): 4,258,874
- Total tokens: 295B (tokenized by
marin-mimi-bpe-8cb-16k-tokenizer)
Dataset Statistics by Language
| Language | Rows | Shards | Subshards | Num Tokens |
|---|---|---|---|---|
| aa | 86 | 1 | 1 | 219,532 |
| ab | 264 | 1 | 1 | 626,448 |
| af | 174 | 1 | 1 | 938,570 |
| ak | 32 | 1 | 1 | 222,516 |
| am | 664 | 1 | 1 | 7,609,996 |
| ar | 20,838 | 1 | 2 | 313,273,752 |
| as | 240 | 1 | 1 | 475,060 |
| ay | 48 | 1 | 1 | 64,148 |
| az | 416 | 1 | 1 | 3,597,864 |
| ba | 40 | 1 | 1 | 392,942 |
| be | 2,126 | 1 | 1 | 35,385,082 |
| bg | 1,668 | 1 | 1 | 43,204,440 |
| bh | 4 | 1 | 1 | 9,696 |
| bi | 10 | 1 | 1 | 197,866 |
| bm | 8 | 1 | 1 | 1,654 |
| bn | 11,060 | 1 | 1 | 54,150,790 |
| bo | 14 | 1 | 1 | 690,586 |
| br | 36 | 1 | 1 | 564,178 |
| bs | 252 | 1 | 1 | 7,175,348 |
| ca | 2,662 | 1 | 1 | 57,126,328 |
| co | 12 | 1 | 1 | 32,626 |
| cr | 6 | 1 | 1 | 4,288 |
| cs | 4,746 | 1 | 1 | 128,435,506 |
| cy | 260 | 1 | 1 | 3,996,964 |
| da | 1,338 | 1 | 1 | 21,472,962 |
| de | 132,832 | 4 | 15 | 10,989,993,042 |
| dz | 18 | 1 | 1 | 48,842 |
| ee | 10 | 1 | 1 | 31,186 |
| el | 3,220 | 1 | 1 | 100,290,700 |
| en | 1,791,316 | 38 | 183 | 131,123,750,680 |
| eo | 396 | 1 | 1 | 12,922,032 |
| es | 476,532 | 10 | 49 | 35,383,236,180 |
| et | 282 | 1 | 1 | 11,377,258 |
| eu | 954 | 1 | 1 | 15,428,750 |
| fa | 914 | 1 | 1 | 43,572,308 |
| ff | 12 | 1 | 1 | 29,884 |
| fi | 3,220 | 1 | 1 | 164,835,624 |
| fj | 14 | 1 | 1 | 13,054 |
| fo | 28 | 1 | 1 | 141,920 |
| fr | 202,596 | 5 | 19 | 16,624,847,906 |
| fy | 16 | 1 | 1 | 64,850 |
| ga | 102 | 1 | 1 | 1,195,830 |
| gd | 16 | 1 | 1 | 15,858 |
| gl | 432 | 1 | 1 | 8,044,654 |
| gn | 40 | 1 | 1 | 308,564 |
| gu | 1,178 | 1 | 1 | 2,757,342 |
| ha | 346 | 1 | 1 | 120,984 |
| hi | 36,462 | 2 | 3 | 413,234,530 |
| ho | 10 | 1 | 1 | 41,884 |
| hr | 738 | 1 | 1 | 20,202,718 |
| ht | 228 | 1 | 1 | 969,466 |
| hu | 2,558 | 1 | 1 | 156,280,022 |
| hy | 284 | 1 | 1 | 2,621,494 |
| ia | 32 | 1 | 1 | 63,498 |
| id | 153,938 | 3 | 15 | 8,083,093,336 |
| ie | 28 | 1 | 1 | 309,702 |
| ig | 16 | 1 | 1 | 126,742 |
| ik | 6 | 1 | 1 | 3,374 |
| is | 306 | 1 | 1 | 3,981,186 |
| it | 115,366 | 3 | 11 | 8,292,491,254 |
| iu | 12 | 1 | 1 | 8,024 |
| iw | 3,478 | 1 | 1 | 131,223,906 |
| ja | 65,482 | 2 | 5 | 3,108,465,094 |
| jv | 106 | 1 | 1 | 906,828 |
| ka | 694 | 1 | 1 | 21,052,458 |
| ki | 2 | 1 | 1 | 420 |
| kk | 540 | 1 | 1 | 5,890,808 |
| kl | 8 | 1 | 1 | 3,812 |
| km | 816 | 1 | 1 | 4,602,562 |
| kn | 574 | 1 | 1 | 3,944,608 |
| ko | 239,366 | 5 | 23 | 15,873,307,780 |
| ks | 18 | 1 | 1 | 26,118 |
| ku | 102 | 1 | 1 | 1,621,536 |
| ky | 1,824 | 1 | 1 | 11,731,944 |
| la | 162 | 1 | 1 | 5,851,778 |
| lb | 16 | 1 | 1 | 85,592 |
| lg | 4 | 1 | 1 | 2,866 |
| ln | 64 | 1 | 1 | 155,462 |
| lo | 28 | 1 | 1 | 202,632 |
| lt | 428 | 1 | 1 | 13,809,912 |
| lv | 108 | 1 | 1 | 1,965,880 |
| mg | 38 | 1 | 1 | 151,252 |
| mi | 132 | 1 | 1 | 846,874 |
| mk | 196 | 1 | 1 | 4,300,234 |
| ml | 1,496 | 1 | 1 | 19,551,968 |
| mn | 120 | 1 | 1 | 1,776,112 |
| mr | 2,570 | 1 | 1 | 6,299,296 |
| ms | 1,160 | 1 | 1 | 21,647,004 |
| my | 174 | 1 | 1 | 2,025,598 |
| na | 10 | 1 | 1 | 30,850 |
| nd | 2 | 1 | 1 | 896 |
| ne | 554 | 1 | 1 | 3,252,512 |
| nl | 51,700 | 2 | 5 | 2,283,828,818 |
| no | 2,500 | 1 | 1 | 99,862,490 |
| nv | 6 | 1 | 1 | 2,418 |
| oc | 12 | 1 | 1 | 131,114 |
| om | 206 | 1 | 1 | 189,614 |
| or | 540 | 1 | 1 | 805,718 |
| pa | 636 | 1 | 1 | 1,529,608 |
| pl | 19,056 | 1 | 1 | 612,614,280 |
| ps | 172 | 1 | 1 | 1,326,312 |
| pt | 220,746 | 5 | 21 | 15,769,904,990 |
| qu | 6 | 1 | 1 | 572,198 |
| rm | 20 | 1 | 1 | 153,982 |
| rn | 6 | 1 | 1 | 3,306 |
| ro | 3,770 | 1 | 1 | 84,549,356 |
| ru | 405,758 | 9 | 41 | 31,069,562,816 |
| rw | 94 | 1 | 1 | 557,810 |
| sa | 42 | 1 | 1 | 942,280 |
| sc | 6 | 1 | 1 | 13,676 |
| sd | 40 | 1 | 1 | 58,524 |
| sg | 2 | 1 | 1 | 406 |
| sh | 6 | 1 | 1 | 189,934 |
| si | 616 | 1 | 1 | 5,219,200 |
| sk | 1,426 | 1 | 1 | 24,986,906 |
| sl | 748 | 1 | 1 | 17,687,176 |
| sm | 4 | 1 | 1 | 78,288 |
| sn | 6 | 1 | 1 | 10,480 |
| so | 426 | 1 | 1 | 4,872,624 |
| sq | 302 | 1 | 1 | 2,800,904 |
| sr | 540 | 1 | 1 | 12,476,862 |
| st | 6 | 1 | 1 | 2,656 |
| su | 36 | 1 | 1 | 30,028 |
| sv | 2,638 | 1 | 1 | 97,736,862 |
| sw | 350 | 1 | 1 | 1,520,946 |
| ta | 4,954 | 1 | 1 | 68,388,738 |
| te | 2,510 | 1 | 1 | 8,288,766 |
| tg | 90 | 1 | 1 | 283,358 |
| th | 16,372 | 2 | 2 | 515,106,000 |
| ti | 24 | 1 | 1 | 379,410 |
| tk | 20 | 1 | 1 | 922,818 |
| tn | 10 | 1 | 1 | 96,390 |
| to | 2 | 1 | 1 | 424 |
| tr | 64,938 | 2 | 7 | 3,847,050,740 |
| ts | 4 | 1 | 1 | 97,266 |
| tt | 10 | 1 | 1 | 33,672 |
| ug | 36 | 1 | 1 | 137,706 |
| uk | 20,554 | 2 | 2 | 687,471,758 |
| ur | 5,328 | 1 | 1 | 25,751,098 |
| uz | 1,200 | 1 | 1 | 7,591,098 |
| ve | 8 | 1 | 1 | 2,676 |
| vi | 133,954 | 3 | 14 | 8,394,394,682 |
| vo | 4 | 1 | 1 | 161,142 |
| wo | 54 | 1 | 1 | 122,826 |
| xh | 14 | 1 | 1 | 119,632 |
| yi | 14 | 1 | 1 | 73,686 |
| yo | 28 | 1 | 1 | 243,384 |
| zh | 4,330 | 1 | 1 | 239,288,480 |
| zu | 260 | 1 | 1 | 2,904,654 |
| TOTAL | 4,258,874 | 230 | 549 | 295,274,191,398 |
Usage
Loading the Dataset
from datasets import load_dataset
# Load all data
dataset = load_dataset("potsawee/yodas2-mm-pretrain")
# Load specific language
dataset = load_dataset("potsawee/yodas2-mm-pretrain", "en")
# Stream the dataset (recommended for large datasets)
dataset = load_dataset("potsawee/yodas2-mm-pretrain", "en", streaming=True)
Example Document
from datasets import load_dataset
dataset = load_dataset("potsawee/yodas2-mm-pretrain", "en", split="train", streaming=True)
sample = next(iter(dataset))
print(f"ID: {sample['id']}")
print(f"Split: {sample['split']}")
print(f"Text length: {len(sample['text'])} characters")
print(f"Preview: {sample['text'][:200]}...")
- Downloads last month
- 185