Behavioral text patterns strongly influence how language models reason, plan, and simulate decision-making. This dataset packages behavior-focused text into shardable Parquet subsets so you can stream, weight, and combine domain slices for pretraining or controlled evaluations.
What Sets It Apart
- Focused domains: explicit subsets for reasoning, planning, data science, games, format rewrites and general/other content let you upweight behavioral genres during pretraining, rather than relying on undifferentiated web crawls.
- Engineering-friendly format: released as Parquet shards with predictable shard prefixes and streaming support, so pipelines using datasets, polars, dask or mIcroissant-style iterators can integrate it with minimal conversion overhead.
- Modular provenance guidance: subset-level metadata may include source-specific cleaning, deduplication, quality scores or synthetic generation flags, enabling targeted selection and risk assessment prior to model training.
Who It's For and Trade-offs
Great fit if you need to emphasize human-like reasoning/planning or behavioral scenarios in a pretraining mix, want shardable Parquet data for large-scale streaming pipelines, or need separable subsets for ablation studies. Look elsewhere if you require fully documented per-example provenance, multimodal signals (audio/video), or a curated benchmark with fixed evaluation splits — this release targets pretraining-scale text corpora rather than evaluation-only benchmarks.
