OpenDataArena/Qwen3-8B-ODA-Math-460k

Qwen3-8B-ODA-Math-460k: Specialized for Mathematical Reasoning

This model is a supervised fine-tuned (SFT) version of Qwen3-8B-Base, developed by OpenDataArena, with a focus on enhancing mathematical reasoning. It leverages the unique ODA-Math-460k dataset, a meticulously curated collection of approximately 460,000 math problems.

Key Differentiators & Training Data

The ODA-Math-460k dataset is distinguished by its rigorous curation pipeline:

• Data Collection: Aggregated from top-performing math datasets identified by the OpenDataArena leaderboard.
• Deduplication & Decontamination: Exact deduplication and benchmark decontamination to prevent evaluation leakage.
• Question Filtering: Multi-stage LLM-based filtering for domain specificity, validity, and problem type, excluding proofs and multiple-choice questions to focus on free-form problems.
• Data Selection: Problems are selected to be challenging for smaller models but solvable by stronger reasoning models, using a two-stage filtering process with Qwen3-8B and Qwen3-30B-A3B.
• Distillation & Verification: Solutions are distilled using AM-Thinking-v1 as a teacher and verified by Compass-Verifier-7B, ensuring only correct problem-response pairs are included.

Performance

Evaluations show that Qwen3-8B-ODA-Math-460k achieves consistent gains over base checkpoints, with notable improvements on competition-style benchmarks. It demonstrates strong performance across various math datasets, including GSM8K, Math500, Omni-Math, and Olympiad-style problems, achieving an average score of 68.8 across a suite of benchmarks.

Ideal Use Cases

• Mathematical Problem Solving: Excels in solving complex, free-form mathematical problems.
• Educational Tools: Can be integrated into platforms requiring robust mathematical reasoning.
• Competitive Math Preparation: Particularly strong in handling competition-style math challenges.