JunyiZou commited on
Commit
5b86301
·
verified ·
1 Parent(s): 095dfd5

Describe Cayley-Math as a collaborative collection

Browse files
README.md CHANGED
@@ -20,9 +20,14 @@ configs:
20
 
21
  # Cayley-Math
22
 
23
- Cayley-Math contains the mathematics portion of Phase 1 for [MMLA Anholon Pretrain](https://huggingface.co/datasets/MMLA-ORG/MMLA-Anholon-Pretrain). The two private repositories carry the same mathematics archive, manifest, tokenizer, and attribution files. Use this repository when a run needs only the mathematics subset.
24
 
25
- ## Available Files
 
 
 
 
 
26
 
27
  | Use | Files | Coverage |
28
  | --- | --- | --- |
@@ -30,7 +35,7 @@ Cayley-Math contains the mathematics portion of Phase 1 for [MMLA Anholon Pretra
30
  | Inspect examples | [Parquet preview](https://huggingface.co/datasets/MMLA-ORG/Cayley-Math/tree/main/phase1/preview) | 4 excerpts from the training split |
31
  | Trace sources and terms | [Source catalog](phase1/sources.json) · [Attribution guide](phase1/attribution/README.md) | All included components |
32
 
33
- ## Dataset Summary
34
 
35
  | Split | Loss-bearing labels | Packed rows |
36
  | --- | ---: | ---: |
@@ -49,7 +54,8 @@ phase1/
49
  ├── preview/ # Parquet excerpts
50
  ├── sources.json # Component counts and provenance references
51
  └── upload-manifest.json # File sizes and hashes by repository path
52
- tokenizer/ # SentencePiece model and vocabulary
 
53
  examples/ # Synthetic illustration of a source record
54
  ```
55
 
@@ -71,7 +77,6 @@ preview = hf_hub_download(
71
  repo_id="MMLA-ORG/Cayley-Math",
72
  repo_type="dataset",
73
  filename="phase1/preview/math.parquet",
74
- token=True,
75
  endpoint="https://huggingface.co",
76
  )
77
  table = pq.read_table(preview)
@@ -83,7 +88,7 @@ Each shard contains three `int32` arrays with shape `[rows, 2048]`. The `input_i
83
 
84
  ## Loading the Training Pack
85
 
86
- Authenticate with a Hugging Face account that can read this repository. Set `HF_DATA_REVISION` to a commit SHA for reproducible access.
87
 
88
  ```python
89
  import os
@@ -97,7 +102,6 @@ root = Path(snapshot_download(
97
  repo_type="dataset",
98
  revision=os.environ.get("HF_DATA_REVISION", "main"),
99
  endpoint="https://huggingface.co",
100
- token=True,
101
  allow_patterns=["phase1/math/**", "phase1/upload-manifest.json", "tokenizer/**"],
102
  local_dir="cayley-phase1",
103
  ))
@@ -109,10 +113,10 @@ Use `math/manifest.json` to locate and verify the extracted arrays. Development
109
 
110
  ## Processing and Source Terms
111
 
112
- The mathematics view uses the same source groups and split assignments as the complete pack. At source boundaries, only mathematics rows enter the math-only arrays. Its 50 million training labels are already counted in Anholon's one billion.
113
 
114
  Source terms are recorded per document. Common Pile arXiv text and Wen-Wei Li excerpts retain their applicable Creative Commons terms. The Stacks Project passages retain GFDL 1.2 or later. See the [batch notes](phase1/README.md) and attribution and notices in `phase1/attribution/math-rights/`.
115
 
116
  ## Citation and Revision
117
 
118
- For research records, cite MMLA-ORG, *Cayley-Math*, Phase 1, and the Hugging Face revision used. See the [MMLA project](https://mmla-org.github.io/) for the model family.
 
20
 
21
  # Cayley-Math
22
 
23
+ Cayley-Math is a collaboratively maintained collection of mathematical text for language-model research. Contributors can publish cleaned collections as separately versioned components. The first component is a frozen mirror of the mathematics subset prepared for Anholon Phase 1.
24
 
25
+ | Collection | Location | Maintenance |
26
+ | --- | --- | --- |
27
+ | Anholon Phase 1 mathematics mirror | [Phase 1 bundle](phase1/README.md) | Anholon data maintainers |
28
+ | Contributed mathematics collections | [Component guide](components/README.md) | Each component's contributors |
29
+
30
+ ## Phase 1 Files
31
 
32
  | Use | Files | Coverage |
33
  | --- | --- | --- |
 
35
  | Inspect examples | [Parquet preview](https://huggingface.co/datasets/MMLA-ORG/Cayley-Math/tree/main/phase1/preview) | 4 excerpts from the training split |
36
  | Trace sources and terms | [Source catalog](phase1/sources.json) · [Attribution guide](phase1/attribution/README.md) | All included components |
37
 
38
+ ## Phase 1 Summary
39
 
40
  | Split | Loss-bearing labels | Packed rows |
41
  | --- | ---: | ---: |
 
54
  ├── preview/ # Parquet excerpts
55
  ├── sources.json # Component counts and provenance references
56
  └── upload-manifest.json # File sizes and hashes by repository path
57
+ components/ # Independently contributed collections
58
+ tokenizer/ # SentencePiece model and vocabulary for Phase 1
59
  examples/ # Synthetic illustration of a source record
60
  ```
61
 
 
77
  repo_id="MMLA-ORG/Cayley-Math",
78
  repo_type="dataset",
79
  filename="phase1/preview/math.parquet",
 
80
  endpoint="https://huggingface.co",
81
  )
82
  table = pq.read_table(preview)
 
88
 
89
  ## Loading the Training Pack
90
 
91
+ Set `HF_DATA_REVISION` to a commit SHA for reproducible access. The bundle uses the included Phase 1 tokenizer.
92
 
93
  ```python
94
  import os
 
102
  repo_type="dataset",
103
  revision=os.environ.get("HF_DATA_REVISION", "main"),
104
  endpoint="https://huggingface.co",
 
105
  allow_patterns=["phase1/math/**", "phase1/upload-manifest.json", "tokenizer/**"],
106
  local_dir="cayley-phase1",
107
  ))
 
113
 
114
  ## Processing and Source Terms
115
 
116
+ The Phase 1 mirror preserves the original mathematics group assignments, splits and training labels. When building a new training mix from Cayley-Math, record the selected commit, component paths and file digests. Check overlap with any existing Anholon mathematics data before combining the components.
117
 
118
  Source terms are recorded per document. Common Pile arXiv text and Wen-Wei Li excerpts retain their applicable Creative Commons terms. The Stacks Project passages retain GFDL 1.2 or later. See the [batch notes](phase1/README.md) and attribution and notices in `phase1/attribution/math-rights/`.
119
 
120
  ## Citation and Revision
121
 
122
+ For research records, cite MMLA-ORG, *Cayley-Math*, the component and its Hugging Face revision. Cite the component's upstream works as listed in its attribution. See the [MMLA project](https://mmla-org.github.io/) for the model family.
components/README.md ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Cayley-Math Contributions
2
+
3
+ Place each contributed collection under `components/<name>/`. The component can be cited and downloaded at a pinned Hugging Face commit.
4
+
5
+ ```text
6
+ components/<name>/
7
+ ├── README.md
8
+ ├── sources.json
9
+ └── data/
10
+ ```
11
+
12
+ Use the README to describe the contributors, source material, cleaning, record fields and splits. Record upstream references, versions, licenses and attribution in `sources.json`. Parquet works for cleaned text and structured examples. Use the format appropriate to the component and document how to load it.
13
+
14
+ The [Phase 1 bundle](../phase1/README.md) is the original packed mathematics mirror. Contributions have their own paths and can use a different schema or tokenizer. A training run can select individual files by dataset commit and record those paths and digests in its data manifest.
phase1/README.md CHANGED
@@ -1,19 +1,10 @@
1
- # P1 packed pretraining dataset
2
 
3
- This private batch contains the frozen 1B-target P1 pretraining dataset for Anholon. The original 2,048-position NumPy arrays, their manifest, and the 32,768-piece tokenizer are preserved byte for byte in the Anholon repository. Training targets are 550M English, 300M Chinese, 100M code, and 50M mathematics. Development and test each have 3,072 targets in the full pack.
4
 
5
- The `math/` files are identical in both repositories. This view contains 50M training targets plus 1,024 targets each for development and test. At boundaries where an original shard straddles source pools, the view contains just the math rows in a separately hashed array. All other math arrays retain the original bytes.
6
 
7
- Use `tar --zstd -xf` to extract the packed archives into the directory containing the corresponding `manifest.json`. Verify the archive hashes in `upload-manifest.json`, and verify the extracted array hashes against the relevant packed manifest. The upload manifest applies to the files listed for the current repository.
8
-
9
- ## Sources and attribution
10
-
11
- Each pool retains its upstream terms. The associated sidecars identify the upstream work, attribution, license and packed document coordinate. The datasets were formed by selecting and normalizing source text, tokenizing it, and taking causal windows. A quota may retain only part of a document.
12
-
13
- - Wikimedia content retains the source article, revision and history links and the applicable CC BY-SA 4.0 terms. Known unresolved exceptions from the frozen selection remain excluded. Existing article-level attribution records describe the extent of external-import checks.
14
- - PMC text retains article-specific CC BY 4.0 or CC0 terms. The selection uses the article prose projections and omits separately scoped content marked by the extractor. General third-party notices are retained where applicable.
15
- - Code retains each file's license and the upstream notices. See `attribution/code-rights/README.md` in Anholon.
16
- - The mathematics pool includes arXiv text cached by [Common Pile](https://huggingface.co/datasets/common-pile/arxiv_papers_filtered), excerpts from Wen-Wei Li's books, and a modified selection from the Stacks Project.
17
 
18
  ### P1 mathematics excerpts from the Stacks Project
19
 
 
1
+ # Anholon Phase 1 Mathematics Mirror
2
 
3
+ This component contains 50 million loss-bearing training tokens in 43,833 packed training windows. The development and test splits each have 1,024 labels. The mirror preserves the frozen mathematics windows and uses the bundled 32,768-piece SentencePiece tokenizer.
4
 
5
+ Extract `math/math.tar.zst` into the directory containing `math/manifest.json`. The [file manifest](upload-manifest.json) records the repository-relative archive hashes, and the math manifest records the extracted array hashes.
6
 
7
+ The [source catalog](sources.json) and [attribution guide](attribution/README.md) identify the upstream works and their licenses. The preparation selected and normalized text, encoded the selected passages with the Phase 1 tokenizer, and packed the token sequences into causal windows.
 
 
 
 
 
 
 
 
 
8
 
9
  ### P1 mathematics excerpts from the Stacks Project
10
 
phase1/upload-manifest.json CHANGED
@@ -3,8 +3,8 @@
3
  "original_dataset_revision": "eebaf76b125a9972ab7d9303b879a03192084844",
4
  "files": {
5
  "phase1/README.md": {
6
- "bytes": 2944,
7
- "sha256": "879331c27c9c9fa7a5cd3c3ec065e44e5be3b32bd77e5401520d28b928f7d955"
8
  },
9
  "phase1/attribution/math-rights/math-attribution.jsonl.gz": {
10
  "bytes": 1351793,
@@ -39,8 +39,8 @@
39
  "sha256": "5ed89ba3ac20243fecf1a8945a7d21e41a9a39e7ff3dfc1a584889360a92a646"
40
  },
41
  "README.md": {
42
- "bytes": 5208,
43
- "sha256": "f0f86a8d1532306640155eca8ea1314a4a7b650ad4daaddc302f8ae6586f70d4"
44
  },
45
  "phase1/preview/math.parquet": {
46
  "bytes": 5952,
 
3
  "original_dataset_revision": "eebaf76b125a9972ab7d9303b879a03192084844",
4
  "files": {
5
  "phase1/README.md": {
6
+ "bytes": 1550,
7
+ "sha256": "622b3c454abe9b3ed20edbcd5e411f6d770323509a08a1c9a40e742a0c806683"
8
  },
9
  "phase1/attribution/math-rights/math-attribution.jsonl.gz": {
10
  "bytes": 1351793,
 
39
  "sha256": "5ed89ba3ac20243fecf1a8945a7d21e41a9a39e7ff3dfc1a584889360a92a646"
40
  },
41
  "README.md": {
42
+ "bytes": 5590,
43
+ "sha256": "5b1bfda83e351adaff742c45d71255b410d594daf1b7077968147db6642a8e9f"
44
  },
45
  "phase1/preview/math.parquet": {
46
  "bytes": 5952,