---

license: apache-2.0

blue-scrub-150M-it-dpo

DPO-aligned version of blue-scrub-150M-it.

DPO training data

Dataset	Training pairs	Evaluation pairs	Total retained pairs
blue-scrubs-preferences-v2	2,670	302	2,972

• Training epochs: 5
• Training method: Direct Preference Optimization (DPO)
• Maximum sequence length: 1,024 tokens
• DPO beta: 0.1

The preference dataset originally contained 2,978 pairs. Six weak pairs were filtered, leaving 2,972 retained pairs.