Dolci Think SFT Olmo Hybrid: 1M-10M Text Records for Instruction Tuning

1M to 10M text records for supervised fine-tuning (SFT) of the OLMo language model, released by the Allen Institute for AI (allenai) in early 2026. The collection is distributed in optimized-parquet format for high-performance processing using libraries like Polars and Dask.

Use Cases

• Supervised fine-tuning of OLMo-based models using the text modality
• Large-scale data processing benchmarks using the Polars and Dask libraries
• Research into hybrid data distributions for instruction-following tasks

Strengths

• 1M to 10M record scale
• Optimized-parquet format for high-speed data loading
• ODC-By license allows for research and educational applications

Limitations

• Lack of explicit column documentation or schema details
• Geographic bias toward US-based data sources as indicated by metadata tags
• Unknown internal label distribution or quality of the hybrid sources

Provenance

Source

Allen Institute for AI (allenai)

Freshness

Last updated March 2026.

Geography

United States