File size: 3,717 Bytes
6e2848c
 
 
 
 
 
 
 
aae9101
 
 
6e2848c
 
 
aae9101
 
6e2848c
aae9101
6e2848c
aae9101
6e2848c
 
 
aae9101
6e2848c
 
 
aae9101
6e2848c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
aae9101
6e2848c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
---
license: odc-by
---

# Zyda2-5T

<!-- Provide a quick summary of the dataset. -->

Zyda2 is a 5 trillion token language modeling dataset created by collecting open and high quality datasets and combining them and cross-deduplication and model-based quality filtering. Zyda2 comprises diverse sources of web data, highly educational content, math, code, and scientific papers.

To construct Zyda2, we took the best open-source datasets available: Zyda, FineWeb, DCLM, Dolma. Models trained on Zyda2 significantly outperform identical models trained on the Pile, RefinedWeb, FineWeb, FineWeb-Edu, and DCLM. Due to our post-processing deduplication, filtering, and weighting pipeline, Zyda2 outperforms all its constituent datasets in resulting model quality.

An early version of Zyda2 was used as the primary dataset for phase 1 pretraining of our Zamba2 series [of](Zyphra/Zamba2-2.7B) [models](Zyphra/Zamba2-1.2B) which perform extremely strongly on a per-token basis and are often state-of-the-art for their size, testifying to the strength of Zyda2 as a pretraining dataset.

According to our evaluations, Zyda2 is the most performant per-token open dataset available. Zyda2 excels at educational and natural language reasoning content. For code performance, we reccomend mixing it with a pure code dataset such as [Starcoder](https://huggingface.co/bigcode/starcoder). 


// TODO Ablation scores key plots

For more information, please see our technical blog (-/TODO LINK)

## How to download

// TODO YURY

## Breakdown by component

// TODO YURY

### Dataset Description

<!-- Provide a longer summary of what this dataset is. -->

- **Curated by:** Zyphra
- **Language(s) (NLP):** Primarily English
- **License:** Open Data Commons License


## Dataset Structure

<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->

// TODO IS THIS CORRECT YURY?


Dataset fields:
- `text`: contains actual text for training
- `source`: component the text is coming from
- `filtering_features`: precomputed values of different features that were used for filtering (converted to json string)
- `source_other`: metadata from the source dataset (converted to json string)

### Source Data

Zyda2 is comprised of four high quality open-source datasets:

Zyda1: https://huggingface.co/datasets/Zyphra/Zyda

Dolma-1.7-cc https://huggingface.co/datasets/allenai/dolma

DCLM-baseline: https://huggingface.co/datasets/mlfoundations/dclm-baseline-1.0

FineWeb-Edu-2 https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu


// Pie chart of composition -- YURY!


#### Personal and Sensitive Information

As a language modelling dataset, it likely contains PII which has not been filtered out of the component datasets and which may have been missed by our own filters.

## Bias, Risks, and Limitations

As a dataset comprised of open web scrapes, it is likely that it contains biased and toxic content. 

## Licensing Information

We are releasing this dataset under the terms of [ODC-BY](https://opendatacommons.org/licenses/by/1-0/). By using this dataset, you are also bound by any license agreements and terms of use of the original data sources.

## Citation

If you use our dataset to train a model, please cite us at:

```
@misc{tokpanov2024zyda,
      title={Zyda: A 1.3T Dataset for Open Language Modeling}, 
      author={Yury Tokpanov and Beren Millidge and Paolo Glorioso and Jonathan Pilault and Adam Ibrahim and James Whittington and Quentin Anthony},
      year={2024},
      eprint={2406.01981},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
```