Dolci Instruct SFT: 2.1 Million Multilingual Samples for Olmo 3 Training

2,152,112 instruction-tuning samples comprise this mixture released by Allen Institute for AI in 2026 for training the Olmo 3 7B Instruct model. It aggregates prompts from sources like OpenThoughts 3, featuring a 32K context length and a blend of expert, machine, and crowdsourced annotations.

Use Cases

• Supervised Fine-Tuning (SFT) of large language models using the 2,152,112 instruction-response pairs
• Long-context model development leveraging the 32K context length feature from the OpenThoughts 3 source
• Multilingual instruction following research across specific language tags such as Amharic (amh), Bengali (ben), and Arabic (arb)

Strengths

• 2,152,112 total samples
• Extended 32K context length
• Multilingual coverage across 15+ language codes including Amharic and Cebuano

Limitations

• Includes machine-generated annotations which may contain hallucinations
• Code prompts were downsampled by a factor of 16, potentially reducing density in programming domains

Provenance

Source

Allen Institute for AI (Ai2)

Collection Method

mixture of crowdsourced, machine-generated, and expert-generated annotations

Freshness

Last updated February 2026.

Geography

Global (multilingual)