minnesotanlp/Finch-8B
minnesotanlp/Finch-8B is an 8 billion parameter causal language model from the Finch family, built on Qwen3-8B and developed by minnesotanlp. It is uniquely fine-tuned using Evolution Fine-Tuning (EFT) to act as a mutation operator within evolutionary search frameworks. This model excels at learning how to evolve solutions across diverse optimization tasks, demonstrating strong performance in discovery and competitive programming problems.
Loading preview...
What is minnesotanlp/Finch-8B?
minnesotanlp/Finch-8B is an 8 billion parameter large language model, part of the Finch family, developed by minnesotanlp. Unlike other models in its family that use Qwen3.5, Finch-8B is built upon the Qwen3-8B architecture. Its core innovation lies in Evolution Fine-Tuning (EFT), a mid-training "practice phase" that teaches the model how to evolve solutions rather than just providing them. This process involves training on evolutionary search trajectories from the Finch Collection across 355 optimization tasks.
Key Capabilities & Differentiators
- Evolutionary Search Mutation Operator: Finch-8B is specifically designed to function as a powerful mutation operator within evolutionary search scaffolds like OpenEvolve, improving the discovery process.
- Learns to Evolve: It internalizes the "discovery know-how" of evolutionary search, moving this behavior into the model itself, which is a significant departure from traditional LLM applications in search.
- Strong Performance: Outperforms its base model by up to +10.24% across 22 held-out tasks, with per-task gains reaching +290%. On NP-hard competitive programming, Finch-9B (a related model) averages 46.01 compared to Qwen3.5-9B's 32.46.
- Synergy with Test-Time RL: Demonstrates strong synergy with test-time reinforcement learning, particularly in mathematical tasks, and can surpass human scores on certain competitive programming problems when combined with KTO preference learning.
When to Use Finch-8B
- Optimization and Discovery Tasks: Ideal for scenarios requiring iterative improvement and discovery of solutions, especially when integrated into an evolutionary search framework.
- Code Generation for Complex Problems: Particularly effective for competitive programming and other NP-hard problems where composing strategies across diverse domains is crucial.
- Research in AI Evolution: Useful for researchers exploring how LLMs can learn and contribute to evolutionary algorithms and automated discovery.