Hugging Face's logo

Datasets:
ArrowLabs
/
WebMathUltra

Tasks:
Text Generation
Question Answering
Modalities:
Text
Formats:
parquet
Languages:
English
Size:
100K - 1M
Tags:
finemath
fineweb
edu
pretrain
pretraining
web
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
MoreThought commited on
Commit
d856cb4
·
verified ·
1 Parent(s): d83832b

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +54 -16
README.md CHANGED
@@ -1,26 +1,64 @@
1
  ---
2
- license: other
3
  task_categories:
4
  - text-generation
 
5
  language:
6
  - en
7
  tags:
8
- - math
9
- - filtered
10
- - finemath
11
- - arithmark-containment
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12
  ---
 
13
 
14
- # WebMath-Ultra
15
 
16
- FineMath rows that survived:
17
- - whole-word drop of `post`, `posted`, `algebra` (case-insensitive)
18
- - any row containing `#` or `]`
19
- - any row containing emojis
20
- - **containment similarity >= 0.5** against at least one
21
- row of `AxiomicLabs/Arithmark-3.0`
22
 
23
- Containment(doc, ref) = |tokens(doc) ∩ tokens(ref)| / |tokens(ref)|, max over refs.
24
- C++ inverted-index kernel (`g++ -O3 -std=c++17`), streamed from the Hub with
25
- no local cache, 2 scoring threads, 5 parquet shards
26
- (494,060 rows).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ license: apache-2.0
3
  task_categories:
4
  - text-generation
5
+ - question-answering
6
  language:
7
  - en
8
  tags:
9
+ - DCLM
10
+ - fineweb
11
+ - edu
12
+ - pretrain
13
+ - pretraining
14
+ - web
15
+ - reason
16
+ - reasoning
17
+ - dense
18
+ - fine
19
+ - fine data
20
+ - ultra
21
+ - pro
22
+ - DCLM Pro
23
+ - DCLM baseline
24
+ - thought
25
+ - thoughts
26
+ - think
27
+ - thinking
28
+ - post train
29
+ - posttrain
30
+ size_categories:
31
+ - 100K<n<1M
32
  ---
33
+ # DCLM Ultra
34
 
35
+ **WebMathUltra** is a heavily filtered and deduplicated, improved version of **finemath-3plus**.
36
 
37
+ It achieves **superior performance** on math benchmarks including arithmark 3.0, and achieving better math loss.
 
 
 
 
 
38
 
39
+ **3M model trained on only finemath-3plus vs WebMathUltra, same architecture, same training budget.**
40
+
41
+ | Metric | finemath 3+ | WebMathUltra |
42
+ |---|---:|---:|
43
+ | Math loss | 3.27 | **3.25** |
44
+ | arithmark 3.0 | 54.95% | **55.55%** |
45
+
46
+ ## Dataset Details
47
+
48
+ - **Curated by:** AInion
49
+ - **Funded by:** ArrowAI
50
+ - **Shared by:** MoreThought
51
+ - **Language:** english
52
+ - **License:** apache 2.0
53
+
54
+ ### Dataset Sources
55
+
56
+ - **Main Source:** https://huggingface.co/datasets/mlfoundations/dclm-baseline-1.0
57
+ - **Ultra Source:** https://huggingface.co/datasets/mlfoundations/dclm-baseline-1.0-parquet
58
+ - **Ultra Repo:** https://huggingface.co/datasets/gair-prox/DCLM-pro
59
+
60
+ ## Uses
61
+
62
+ - PreTraining SLMs on clean, heavily filtered math.
63
+ - Continued PreTraining SLMs on clean, heavily filtered math.
64
+ - PostTraining LLMs on clean, heavily filtered math.