ODA Scored Data: 25 Million SFT Samples with 30 Quality Metrics

OpenDataArena released this collection of 25 million instruction-following samples across 63 source datasets in March 2026. Each record is annotated with 30 distinct quality metrics to support fine-grained data selection for Supervised Fine-Tuning (SFT).

Use Cases

• Curriculum learning by sorting samples based on the PPL (Perplexity) metric
• Data filtering to improve model alignment by excluding records with low Deita_Quality scores
• Fine-grained data selection for mixture optimization using the IFD (Instruction-Following Difficulty) feature

Strengths

• 25 million instruction-following samples
• 30 distinct scoring metrics per sample
• Consolidates 63 separate instruction-following datasets

Limitations

• Scoring metrics are derived from specific models or algorithms which may contain inherent systematic biases
• Large storage and memory requirements for processing 25 million records

Provenance

Source

OpenDataArena

Collection Method

Aggregated from 63 existing datasets and scored using 30 different metrics.

Freshness

Last updated March 2026.

Geography

United States