README.md

metadata

dataset_info:
  features:
    - name: text
      dtype: string

Pre-1900 Training Corpus

Chunked and resharded pre-1900 English text corpus, ready for language model training.

Format

• 266 parquet shards (265 train + 1 validation)
• 12.8M documents (chunks of ≤8,000 characters)
• ~22B tokens estimated
• Text-only — single text column per row
• Row groups divisible by 8 for even DDP distribution across GPUs
• Last shard (shard_00265) is the validation split

Processing Pipeline

Built from the full pre-1900 filtered corpus through:

1. OCR cleanup — removal of OCR artifacts, boilerplate, and unicode normalization
2. Quality filtering — token frequency prior-based filtering
3. Anachronism detection — three-tier post-1900 physics filter
4. Document chunking — long documents split at paragraph/sentence boundaries (max 8K chars, min 200 chars)
5. Token balancing — sort-by-length + round-robin distribution across shards for even token counts

Usage

from datasets import load_dataset
ds = load_dataset("mhla/pre1900-training")

Related

• mhla/pre1900-corpus — full documents with metadata (title, year, source, OCR scores)
• mhla/gpt1900-d26-8btok — GPT-1900 model trained on this data