JunyiZou commited on
Commit
095dfd5
·
verified ·
1 Parent(s): 24a1655

Document Phase 1 downloads and upstream components

Browse files
README.md CHANGED
@@ -22,6 +22,14 @@ configs:
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
  ## Dataset Summary
26
 
27
  | Split | Loss-bearing labels | Packed rows |
@@ -38,7 +46,9 @@ Sources include mathematical arXiv passages cached by Common Pile, excerpts from
38
  phase1/
39
  ├── math/ # Mathematics archive and packed manifest
40
  ├── attribution/ # Document-level sources, licenses, and notices
41
- └── upload-manifest.json # Archive hashes and repository-relative paths
 
 
42
  tokenizer/ # SentencePiece model and vocabulary
43
  examples/ # Synthetic illustration of a source record
44
  ```
@@ -51,13 +61,27 @@ The Parquet preview contains **4 excerpts** decoded from frozen training windows
51
  | --- | --- | --- |
52
  | Mathematics | , we start with the notion of joint transitivity. ###### Definition 1.4. Let $(X,S_{1},\ldots,S_{k})$ be a $\mathbb{Z}^{k}$-system, $T_{1},\ldots,T_{d}\in\langl… | [Upstream](https://arxiv.org/abs/2404.07876) |
53
 
54
- The full NumPy training pack is described in the loading section.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
55
 
56
  ## Data Fields
57
 
58
  Each shard contains three `int32` arrays with shape `[rows, 2048]`. The `input_ids` array uses zero for padding. The `labels` array stores shifted targets and masks unused positions with `-100`. The `role_ids` array is zero throughout plain pretraining text.
59
 
60
- ## Loading the Data
61
 
62
  Authenticate with a Hugging Face account that can read this repository. Set `HF_DATA_REVISION` to a commit SHA for reproducible access.
63
 
@@ -83,10 +107,12 @@ subprocess.run(["tar", "--zstd", "-xf", str(math / "math.tar.zst"), "-C", str(ma
83
 
84
  Use `math/manifest.json` to locate and verify the extracted arrays. Development and test are small monitoring splits.
85
 
86
- ## Preparation and Source Terms
87
 
88
  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.
89
 
90
  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/`.
91
 
 
 
92
  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.
 
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
+ | --- | --- | --- |
29
+ | Train with the frozen release | [NumPy archives](https://huggingface.co/datasets/MMLA-ORG/Cayley-Math/tree/main/phase1/math) | 50,000,000 training labels |
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 |
 
46
  phase1/
47
  ├── math/ # Mathematics archive and packed manifest
48
  ├── attribution/ # Document-level sources, licenses, and notices
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
  ```
 
61
  | --- | --- | --- |
62
  | Mathematics | , we start with the notion of joint transitivity. ###### Definition 1.4. Let $(X,S_{1},\ldots,S_{k})$ be a $\mathbb{Z}^{k}$-system, $T_{1},\ldots,T_{d}\in\langl… | [Upstream](https://arxiv.org/abs/2404.07876) |
63
 
64
+ The preview uses four fixed row offsets per domain at 20%, 40%, 60%, and 80% of its packed training rows. Each excerpt includes up to 192 tokens. The `packed_row` field is a zero-based index into Anholon's full Phase 1 train split, including in the mathematics mirror.
65
+
66
+ ```python
67
+ import pyarrow.parquet as pq
68
+ from huggingface_hub import hf_hub_download
69
+
70
+ 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)
78
+ ```
79
 
80
  ## Data Fields
81
 
82
  Each shard contains three `int32` arrays with shape `[rows, 2048]`. The `input_ids` array uses zero for padding. The `labels` array stores shifted targets and masks unused positions with `-100`. The `role_ids` array is zero throughout plain pretraining text.
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
 
 
107
 
108
  Use `math/manifest.json` to locate and verify the extracted arrays. Development and test are small monitoring splits.
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.
phase1/attribution/README.md ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # Phase 1 Attribution
2
+
3
+ The records here link packed training sources to upstream works, source identifiers and license notices. The catalog records the scope and frozen input digest of each component. [Source catalog](../sources.json)
4
+
5
+ | Component | Attribution records |
6
+ | --- | --- |
7
+ | Mathematics | [math-attribution.jsonl.gz](math-rights/math-attribution.jsonl.gz) · [stacks-CONTRIBUTORS](math-rights/stacks-CONTRIBUTORS) · [stacks-COPYING](math-rights/stacks-COPYING) · [stacks-introduction.tex](math-rights/stacks-introduction.tex) |
8
+
9
+ Source-specific terms accompany each component. The batch notes describe the text transformations and preserve the Stacks Project license and attribution notice. [Batch notes](../README.md)
phase1/sources.json ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "phase": "phase1",
3
+ "dataset": "MMLA-ORG/Cayley-Math",
4
+ "parent_packed_manifest_sha256": "8affc5c5feea945f246ef170254e1577f68ffbf0268cd9ba2dbce0c1355b8be0",
5
+ "components": [
6
+ {
7
+ "id": "math",
8
+ "name": "Mathematics",
9
+ "upstream": [
10
+ {
11
+ "name": "Common Pile arXiv cached text",
12
+ "url": "https://huggingface.co/datasets/common-pile/arxiv_papers_filtered/tree/033cf7f53f9b348deec868c1a5a48484f3ee9e52",
13
+ "files": [
14
+ "arxiv-papers-0006.json.gz",
15
+ "arxiv-papers-0007.json.gz"
16
+ ],
17
+ "license": "CC BY 4.0 or CC0, per work"
18
+ },
19
+ {
20
+ "name": "The Stacks Project",
21
+ "url": "https://github.com/stacks/stacks-project/tree/a04446e57ec1fbc252a871afcec7752fb2807b14",
22
+ "license": "GFDL 1.2 or later"
23
+ },
24
+ {
25
+ "name": "Wen-Wei Li, Methods of Algebra, Volume 1",
26
+ "url": "https://github.com/wenweili/AlJabr-1/tree/1e2d0bf4e5b71194b85958ca6b16a6f396ad25be",
27
+ "license": "CC BY 4.0"
28
+ },
29
+ {
30
+ "name": "Wen-Wei Li, Introduction to Modular Forms",
31
+ "url": "https://github.com/wenweili/Modulform/tree/5e462b0254321ea37f67583562f49575f75f34b7",
32
+ "license": "CC BY 4.0"
33
+ }
34
+ ],
35
+ "split_labels": {
36
+ "train": 50000000,
37
+ "dev": 1024,
38
+ "test": 1024
39
+ },
40
+ "packed_train_rows": 43833,
41
+ "normalized_input_sha256": "c958ca8a1cde9b0d115fdb0a81fb89595790f5034492cd03f3fb093a2530a0c3",
42
+ "source_terms_receipt_sha256": "137a091fbd8666a7950d5b959d526e9307667ad8394e7d2eb42055d7521ed3e1",
43
+ "attribution_files": [
44
+ "phase1/attribution/math-rights/math-attribution.jsonl.gz",
45
+ "phase1/attribution/math-rights/stacks-CONTRIBUTORS",
46
+ "phase1/attribution/math-rights/stacks-COPYING",
47
+ "phase1/attribution/math-rights/stacks-introduction.tex"
48
+ ],
49
+ "preview_file": "phase1/preview/math.parquet"
50
+ }
51
+ ]
52
+ }
phase1/upload-manifest.json CHANGED
@@ -39,12 +39,20 @@
39
  "sha256": "5ed89ba3ac20243fecf1a8945a7d21e41a9a39e7ff3dfc1a584889360a92a646"
40
  },
41
  "README.md": {
42
- "bytes": 4012,
43
- "sha256": "69883c02fbd29f930c9127ef93e5f68d19c429ffc40e501a1a428f4ee87e2a20"
44
  },
45
  "phase1/preview/math.parquet": {
46
  "bytes": 5952,
47
  "sha256": "cc20bc77d6dc1ad1c0f6ff03913c68e477d21fc9e2241bf1c40dccd8c22c687b"
 
 
 
 
 
 
 
 
48
  }
49
  }
50
  }
 
39
  "sha256": "5ed89ba3ac20243fecf1a8945a7d21e41a9a39e7ff3dfc1a584889360a92a646"
40
  },
41
  "README.md": {
42
+ "bytes": 5208,
43
+ "sha256": "f0f86a8d1532306640155eca8ea1314a4a7b650ad4daaddc302f8ae6586f70d4"
44
  },
45
  "phase1/preview/math.parquet": {
46
  "bytes": 5952,
47
  "sha256": "cc20bc77d6dc1ad1c0f6ff03913c68e477d21fc9e2241bf1c40dccd8c22c687b"
48
+ },
49
+ "phase1/sources.json": {
50
+ "bytes": 1978,
51
+ "sha256": "8e076769bc786057f7fabee103a07b5af9a3d167169bddc04739f6c419ec810c"
52
+ },
53
+ "phase1/attribution/README.md": {
54
+ "bytes": 734,
55
+ "sha256": "6da2f626eb78254934931fbfeb365fe90c6dffa322bfdf148043ed3eaac41f92"
56
  }
57
  }
58
  }