minnesotanlp/Finch-4B

Finch-4B: Evolution Fine-Tuned for Discovery

Finch-4B is a 4.5 billion parameter model, part of the Finch family developed by minnesotanlp, specifically designed for Evolution Fine-Tuning (EFT). Built upon Qwen3.5-4B, this model learns to act as a powerful mutation operator within evolutionary search algorithms, effectively discovering solutions across a wide range of optimization tasks.

Key Capabilities

• Evolutionary Search Optimization: Finch-4B is trained on "improved" transitions from the Finch Collection across 355 training tasks, enabling it to learn how to evolve solutions (e.g., which parts to mutate, what to retain, when to backtrack).
• Outperforms Base Models: It achieves up to +10.24% improvement across 22 held-out tasks spanning 5 domains, with per-task gains reaching +290% compared to its base model.
• Size Efficiency: Finch-4B demonstrates strong performance, matching models roughly twice its size on tasks like the Erdős minimum-overlap problem.
• Competitive Programming: Finch models show significant gains in NP-hard competitive programming tasks, composing strategies learned across diverse domains.

Good For

• Automated Discovery Systems: Ideal for integration into evolutionary search scaffolds like OpenEvolve, where it can propose improved candidate programs.
• Complex Problem Solving: Particularly effective for optimization problems requiring iterative refinement and discovery, such as circle packing or competitive programming challenges.
• Research in AI Evolution: Useful for exploring how LLMs can learn and apply evolutionary strategies to solve novel problems.