BerenMillidge commited on
Commit
aae9101
1 Parent(s): 6e2848c

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +10 -18
README.md CHANGED
@@ -6,22 +6,26 @@ license: odc-by
6
 
7
  <!-- Provide a quick summary of the dataset. -->
8
 
9
- Zyda2 is a 5 trillion token language modeling dataset created by collecting open and high quality datasets and combining them and performing a uniform filtering and deduplication step. We find that Zyda performs extremely well in ablations and is at least comparable and potentially better to the best openly available datasets available, due to our meticulous post-processing pipeline. We think the best use of Zyda is either as a standalone dataset for language model training up to the 1T scale, or in combination with Fineweb or Dolma for multi-trillion token training.
 
 
10
 
11
  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.
12
 
13
- Models trained on Zyda2 significantly outperform identical models trained on the Pile, RefinedWeb, FineWeb, FineWeb-Edu, and DCLM.
 
14
 
15
- 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).
16
 
 
17
 
18
  ## How to download
19
 
20
- // TODO
21
 
22
  ## Breakdown by component
23
 
24
- // TODO
25
 
26
  ### Dataset Description
27
 
@@ -47,7 +51,7 @@ Dataset fields:
47
 
48
  ### Source Data
49
 
50
- Zyda2 is comprised of four high quality open-source datasets.
51
 
52
  Zyda1: https://huggingface.co/datasets/Zyphra/Zyda
53
 
@@ -60,18 +64,6 @@ FineWeb-Edu-2 https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu
60
 
61
  // Pie chart of composition -- YURY!
62
 
63
- #### Data Collection and Processing
64
-
65
- <!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->
66
-
67
- Zyda was created using a two stage post-processing pipeline consisting of *filtering* and *deduplication*.
68
-
69
- For the filtering stage, we utilized a set of hand-crafted and tuned filters derived from a number of sources such as C4, RedPajama, and Gopher, in addition to our own filters.
70
-
71
- For the deduplication stage, we used minhash approximate deduplication. We deduplicated on 13-grams and used a minhash signature size of 128 and filtered out documents above a Jaccard similarity of 0.4.
72
-
73
- For full details on our data processing, see the [Zyda technical report](https://arxiv.org/abs/2406.01981) and our [dataset processing code](https://github.com/Zyphra/Zyda_processing).
74
-
75
 
76
  #### Personal and Sensitive Information
77
 
 
6
 
7
  <!-- Provide a quick summary of the dataset. -->
8
 
9
+ 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.
10
+
11
+ 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.
12
 
13
  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.
14
 
15
+ 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).
16
+
17
 
18
+ // TODO Ablation scores key plots
19
 
20
+ For more information, please see our technical blog (-/TODO LINK)
21
 
22
  ## How to download
23
 
24
+ // TODO YURY
25
 
26
  ## Breakdown by component
27
 
28
+ // TODO YURY
29
 
30
  ### Dataset Description
31
 
 
51
 
52
  ### Source Data
53
 
54
+ Zyda2 is comprised of four high quality open-source datasets:
55
 
56
  Zyda1: https://huggingface.co/datasets/Zyphra/Zyda
57
 
 
64
 
65
  // Pie chart of composition -- YURY!
66
 
 
 
 
 
 
 
 
 
 
 
 
 
67
 
68
  #### Personal and Sensitive Information
69