DCLM Crossover Source: A Subset for Format-Aware Synthetic Augmentation

251,661 documents were selected from a scan of 255,841 source documents, totaling 196,694,035 words with an average of 781 words per document. This subset of the DCLM-Baseline dataset was created by author essobi for synthetic augmentation with format-aware prompt routing, and was last updated on April 13, 2026. Selection criteria included picking every third shard, applying a word count filter of 50-8000 words, and skipping prompts that duplicate a document's native format.

Use Cases

• Training language models on format-diverse text based on the described format detection and filtering.
• Generating synthetic training data for instruction-following models based on the format-aware prompt routing methodology.
• Benchmarking text generation quality across different document formats based on the described selection process.
• Studying the impact of document length on model performance based on the 50-8000 word count filter applied.

Strengths

• Large scale with 251,661 selected documents and nearly 200 million total words.
• Explicit length filtering (50-8000 words) and format-aware selection to likely improve data consistency.
• Per-site capping was applied, with zero instances reported, suggesting controlled source representation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for specific batch processing needs.
• The license is unknown, which restricts clarity on permissible use cases.

Provenance

Source

Subset of DCLM-Baseline dataset, hosted on Hugging Face by author essobi.

Collection Method

Selected from 255,841 scanned source docs by picking every third shard, applying word count and format filters, and implementing a per-site cap.

Time Range

null

Freshness

Last updated 2026-04-13 14:23:02; freshness should be verified.

Geography

null