Cayley-Math / README.md
JunyiZou's picture
Update README.md
7944460 verified
|
Raw History Blame
7.18 kB
metadata
pretty_name: Cayley Math
language:
  - en
  - zh
license: apache-2.0
license_name: mixed-upstream-licenses
task_categories:
  - text-generation
tags:
  - pretraining
  - language-modeling
  - mathematics
configs:
  - config_name: anholon-phase1
    default: true
    data_files:
      - split: train
        path:
          - phase1/preview/full/train-initial.parquet
          - phase1/preview/full/train-continuation.parquet
          - phase1/preview/full/train-rest-14672b537052.parquet
          - phase1/preview/full/train-rest-7d766bf7f01e.parquet
          - phase1/preview/full/train-rest-8d08c3c8fd37.parquet
          - phase1/preview/full/train-rest-a3846b4799ee.parquet
          - phase1/preview/full/train-rest-bd420910b68f.parquet
          - phase1/preview/full/train-rest-bee1b70f5d99.parquet
          - phase1/preview/full/train-rest-d4222945b5bb.parquet
          - phase1/preview/full/train-rest-d93823f0124c.parquet
      - split: validation
        path: phase1/preview/full/validation.parquet
      - split: test
        path: phase1/preview/full/test.parquet
  - config_name: openai-math
    data_files:
      - split: train
        path: components/openai-math/data/train.parquet
      - split: validation
        path: components/openai-math/data/dev.parquet
      - split: test
        path: components/openai-math/data/test.parquet

Cayley-Math

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.

Collection Location Maintenance
Anholon Phase 1 mathematics mirror Phase 1 bundle Anholon data maintainers
OpenAI mathematics manuscripts OpenAI Math component Source extraction and OMP DeepSeek metadata
Contributed mathematics collections Component guide Each component's contributors

Phase 1 Files

Use Files Coverage
Train with the frozen release NumPy archives 1,071,000,000 training labels across three batches
Browse every packed mathematics window Data Studio 1,009,501 training rows plus validation and test
Trace sources and terms Source catalog · Attribution guide All included components

Phase 1 Summary

Split Loss-bearing labels Packed rows
Train, initial batch 50,000,000 43,833
Train, continuation 200,000,000 103,132
Train, full pretraining continuation 821,000,000 862,536
Development 1,024 1
Test 1,024 1

Sources include mathematical arXiv passages cached by Common Pile, excerpts from the Stacks Project, and textbooks by Wen-Wei Li. Each window has up to 2,048 token positions and uses the same 32,768-piece SentencePiece vocabulary as the full corpus.

Repository Structure

phase1/
├── math/                # Mathematics archive and packed manifest
├── attribution/         # Document-level sources, licenses, and notices
├── preview/            # Full text viewer under full/
├── sources.json        # Component counts and provenance references
└── upload-manifest.json # File sizes and hashes by repository path
components/              # Independently contributed collections
tokenizer/               # SentencePiece model and vocabulary for Phase 1
examples/                # Synthetic illustration of a source record

Browse the dataset

Select anholon-phase1 in Data Studio. The train split contains all 43,833 initial mathematics windows, 103,132 first continuation windows, and 862,536 full pretraining continuation windows. The original validation and test splits each contain one row. Select openai-math for the independently contributed collection.

Each Parquet row decodes one frozen window and records its batch, row number within that batch and split, supervised label count, source, license and attribution. Padding and control tokens are omitted in the displayed text. Consecutive windows can share a boundary token. The NumPy archives contain the original token IDs and label masks.

The Parquet files and their hashes are in phase1/preview/full. The Stacks Project rows have a compressed editable UTF-8 copy with the source notices and GFDL text.

Data Fields

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.

Loading the Training Pack

Set HF_DATA_REVISION to a commit SHA for reproducible access. The bundle uses the included Phase 1 tokenizer.

import os
import subprocess
from pathlib import Path

from huggingface_hub import snapshot_download

root = Path(snapshot_download(
    repo_id="MMLA-ORG/Cayley-Math",
    repo_type="dataset",
    revision=os.environ.get("HF_DATA_REVISION", "main"),
    endpoint="https://huggingface.co",
    allow_patterns=["phase1/math/**", "phase1/upload-manifest.json", "tokenizer/**"],
    local_dir="cayley-phase1",
))
math = root / "phase1" / "math"
subprocess.run(["tar", "--zstd", "-xf", str(math / "math.tar.zst"), "-C", str(math)], check=True)

Use math/manifest.json to locate and verify the extracted initial arrays. The continuation bundle has its own archive and manifest. The full pretraining math continuation adds 821M training labels with its own archives and attribution. Development and test are small monitoring splits.

Processing and Source Terms

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.

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 and attribution and notices in phase1/attribution/math-rights/.

The full pretraining continuation source terms cover MathOverflow, Cross Validated, arXiv, and OpenAI Math.

Citation and Revision

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 for the model family.