JunyiZou commited on
Commit
0fc741c
·
verified ·
1 Parent(s): b580897

Document pretraining data and synthetic shard formats

Browse files
Files changed (1) hide show
  1. README.md +26 -6
README.md CHANGED
@@ -4,16 +4,36 @@ license: apache-2.0
4
 
5
  # Cayley-Math
6
 
7
- Private mirror of the mathematics subset of [MMLA-Anholon-Pretrain](https://huggingface.co/datasets/MMLA-ORG/MMLA-Anholon-Pretrain). The first-round mathematics target is 50 million tokens that contribute to the training loss.
8
 
9
  ## Status
10
 
11
- Mathematics sources are being prepared and reviewed. The repository contains documentation and a [synthetic mathematics example](examples/synthetic-math-record.md). The example is for format illustration and is excluded from the corpus and token counts. The target will be measured with the frozen 32,768-piece tokenizer and final supervision masks.
12
 
13
- ## Mirror scope
14
 
15
- Each synchronized mathematics file is copied from the same prepared batch used in the main pretraining repository. Its split, bytes, SHA-256, and attribution match the main copy. The shared manifest identifies both repository revisions. Only mathematics batches that pass data admission and private hosting review are synchronized.
16
 
17
- ## Source licenses
 
 
 
 
18
 
19
- The repository metadata is separate from the licenses of the source material. Each batch records its original source licenses, versions, and attribution.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4
 
5
  # Cayley-Math
6
 
7
+ Cayley-Math is the private mirror of the reviewed mathematics pool for [MMLA-Anholon-Pretrain](https://huggingface.co/datasets/MMLA-ORG/MMLA-Anholon-Pretrain). It supports from-scratch pretraining of a 1.481B Transformer–MMLA model. The first-round mathematics target is 50 million tokens that contribute to the training loss, within the main repository's one-billion-token first round.
8
 
9
  ## Status
10
 
11
+ Mathematics sources are being prepared and reviewed. Admitted corpus shards and measured supervision are pending. The [synthetic mathematics record](examples/synthetic-math-record.md) illustrates the format using original prose. It has no assigned split and contributes zero tokens to the corpus. The first-round count will use the frozen 32,768-piece tokenizer and the labels that remain after masking and packing.
12
 
13
+ ## Record format
14
 
15
+ The normalized input format is UTF-8 JSONL, with one record per line.
16
 
17
+ | Field | Meaning |
18
+ | --- | --- |
19
+ | `text` | Cleaned mathematical prose with retained formulas. |
20
+ | `group` | Stable work or source-family identifier used to keep related editions and passages in the same split. |
21
+ | `license` | Applicable source license for the passage, supported by the batch's terms and attribution evidence. |
22
 
23
+ The example's `example_only` value is a format placeholder. Corpus records carry source-specific license and attribution information.
24
+
25
+ ## Formula preservation and review
26
+
27
+ Preparation starts from fixed originals. Depending on the source, static projection preserves mathematical content from TeX, PreTeXt, or structured XML and MathML. A passage's provenance identifies the work and edition, original member and location, applicable terms, and projection evidence. Formula evidence records the retained UTF-8 byte ranges and their SHA-256 digests so the normalized formulas can be checked against the documented transformation. Attribution records the source link, credit, and relevant changes made during extraction.
28
+
29
+ Train, validation, and test are assigned by group before packing. Related editions and passages within a source family stay together. Cross-pool review checks for overlapping material and split conflicts. Exact duplicate checks, near-duplicate candidate review, and benchmark protection checks govern retained scope. Passages with unresolved attribution or uncertain source scope remain held for review.
30
+
31
+ ## Shards, manifests, and access
32
+
33
+ Each synchronized batch identifies the split, exact file paths, byte sizes, SHA-256 digests, source versions, and attribution. After data admission and private hosting review, the same mathematics shard bytes, split labels, and accompanying attribution are copied to both private repositories. The batch manifest and synchronization record bind both repository commits. The mirror is a second copy of the same mathematics data.
34
+
35
+ For reproducible access, use an authorized Hugging Face account and the full Cayley-Math dataset commit named in the synchronization record. Retrieve the batch manifest and the files it lists at that commit. Verify their sizes and SHA-256 digests against the manifest. The same values can be checked against the corresponding commit in MMLA-Anholon-Pretrain.
36
+
37
+ ## Scope and source licenses
38
+
39
+ Cayley-Math covers the pretraining mathematics pool. Math evaluation data and later-stage supervision have separate preparation and evaluation contracts. The repository-level license metadata is separate from the source licenses documented with each batch.