hanzceo's picture
Penny-1B-Base-L1: 4,076,511 documents (text)
84ea80a verified
|
Raw History Blame Contribute Delete
5.04 kB
metadata
license: other
license_name: mixed
license_link: LICENSE
language:
  - en
task_categories:
  - text-generation
size_categories:
  - 1M<n<10M
pretty_name: Penny-1B-Base-L1 pretraining data
tags:
  - penny
  - pretraining
  - synthetic
  - reading-comprehension
  - knowledge-light
configs:
  - config_name: default
    default: true
    data_files:
      - split: train
        path: data/train/*.parquet
      - split: validation
        path: data/validation/*.parquet
  - config_name: warmup
    data_files:
      - split: train
        path: data/train/warmup-*.parquet
      - split: validation
        path: data/validation/warmup-*.parquet
  - config_name: web
    data_files:
      - split: train
        path: data/train/web-*.parquet
      - split: validation
        path: data/validation/web-*.parquet
  - config_name: worlds
    data_files:
      - split: train
        path: data/train/worlds-*.parquet
      - split: validation
        path: data/validation/worlds-*.parquet
  - config_name: kg
    data_files:
      - split: train
        path: data/train/kg-*.parquet
      - split: validation
        path: data/validation/kg-*.parquet
  - config_name: code
    data_files:
      - split: train
        path: data/train/code-*.parquet
      - split: validation
        path: data/validation/code-*.parquet
  - config_name: math_solutions
    data_files:
      - split: train
        path: data/train/math_solutions-*.parquet
      - split: validation
        path: data/validation/math_solutions-*.parquet

Penny-1B-Base-L1 pretraining data

The text of every document in the pretraining stage Penny-1B-Base-L1 of Penny, 2.99B tokens of training in all. Each row is one whole document, exactly as the stage's build kept it, before tokenization.

Penny is meant to be knowledge-light: it should learn from the material it is given (a knowledge graph, documentation) and reason over it, rather than recall facts. So about half of this stage is generated documents that define something new (rules, a knowledge graph, a programming language, every name invented for that document) and then use it. Nothing in them can be memorized, so a model has to read them. The rest is English web text, for language, and worked math solutions.

Sources

source what it is license unique train tokens tokens in the stage train docs validation docs files
warmup procedural warm-up: short algorithmic tasks (bracket matching, sorting, set operations, stack machines, string transforms), trained first apache-2.0 (generated) 0.010B 0.010B (0.3%) 6,759 45 2
web English web text, openbmb/Ultra-FineWeb-L3 (multi_style) apache-2.0 (upstream) 1.390B 1.390B (46.5%) 2,983,202 1,830 33
worlds generated: rule systems, counterfactual worlds, invented units, operators and languages, each defined in the document, then statements of what follows (easy ones first) apache-2.0 (generated) 0.600B 0.600B (20.1%) 383,582 338 5
kg generated: knowledge graphs of invented people, organizations and places, then plain statements of what follows from them (one-hop lookups first, then chains, counts and comparisons) apache-2.0 (generated) 0.450B 0.450B (15.1%) 241,510 286 4
code generated: invented programming languages and libraries (reference first, then programs with their outputs and programs written from a description) apache-2.0 (generated) 0.450B 0.450B (15.1%) 205,934 227 4
math_solutions worked math solutions, nvidia/OpenMathInstruct-2 cc-by-4.0 (upstream) 0.090B 0.090B (3.0%) 252,058 740 2

"Unique train tokens" are what the build tokenized (generated sources: including the end-of-window padding of packed 32K windows); "tokens in the stage" is what training reads from that source. Tokenizer: tiny-random/deepseek-v4.1-flash-1017 (the DeepSeek-V4.1 tokenizer); documents are separated by token id 1 in training.

Use

from datasets import load_dataset

ds = load_dataset("LocalWisdom/Penny-1B-Base-L1-data", split="train")              # every source
kg = load_dataset("LocalWisdom/Penny-1B-Base-L1-data", "kg", split="validation")   # one source

Columns: text (the document), source (one of the names above). Files are zstd parquet of at most ~95 MB.

How it was built

scripts/prepare_mix.py --stage Penny-1B-Base-L1 --text ... from the Penny code: web and math rows in the spec's file order up to each source's token budget (validation first, then train), generated documents from fixed seeds (penny.synth, no language model involved). The build is deterministic, so the same command rebuilds this dataset.

License

Mixed, per source (table above): the generated sources (warmup, worlds, kg, code) are Apache-2.0; web keeps Ultra-FineWeb's Apache-2.0; math_solutions keeps OpenMathInstruct-2's CC-BY-4.0 (attribution: NVIDIA, OpenMathInstruct-2). Check the upstream cards for their terms.