Light-R1 SFT: 79,000 Two-Stage Curriculum Samples for Reasoning Models

Qihoo360 released this 79,000-record text collection in March 2025 for two-stage Supervised Fine-Tuning (SFT) of reasoning models. It consists of a 76,000-sample initial stage and a 3,000-sample curriculum stage designed to train models like Light-R1-32B without relying on long Chain-of-Thought distillation.

Use Cases

• Fine-tuning 32B parameter models for AIME24 and AIME25 reasoning benchmarks using the two-stage JSON files
• Implementing a two-stage curriculum training pipeline using stage1-76k.json and stage2-3k.json
• Researching model performance gains from non-distilled reasoning data as described in Arxiv 2503.10460

Strengths

• 79,000 total training records
• Two-stage curriculum structure (76k and 3k splits)
• Apache 2.0 license for open redistribution

Limitations

• Small sample size for the second curriculum stage (3,000 records)
• Narrow focus on reasoning tasks may limit general conversational performance
• Metadata does not specify the exact source of the individual instructions

Provenance

Source

Qihoo360 (Technical Report Arxiv 2503.10460)

Freshness

Last updated March 17, 2025.