minnesotanlp/Finch-4B-KTO
minnesotanlp/Finch-4B-KTO is a 4.5 billion parameter language model developed by minnesotanlp, built upon Qwen3.5-4B. It is specifically fine-tuned using Evolution Fine-Tuning (EFT) to act as a mutation operator for evolutionary search, enabling it to evolve solutions across diverse optimization tasks. Further enhanced with KTO preference learning, it gains the ability to self-judge solution quality, making it highly effective for discovery and problem-solving in complex domains like competitive programming and mathematical optimization.
Loading preview...
Overview of Finch-4B-KTO
Finch-4B-KTO is a 4.5 billion parameter model from the minnesotanlp Finch family, based on Qwen3.5-4B. It is uniquely designed for evolutionary search and problem discovery. The model undergoes a two-stage training process: first, Evolution Fine-Tuning (EFT) teaches it how to evolve solutions by acting as a mutation operator. Second, KTO (Kahneman-Tversky Optimization) preference learning is applied, enabling the model to self-judge the quality of candidate solutions, distinguishing between improvements and regressions.
Key Capabilities
- Evolutionary Search Operator: Functions as a mutation operator within evolutionary scaffolds like OpenEvolve, proposing improved candidate programs.
- Cross-Task Discovery: Learns to compose strategies across 371 diverse optimization tasks, moving discovery behavior into the model itself.
- Self-Judgment of Solution Quality: Utilizes KTO to internalize a sense of solution quality, enhancing its ability to identify promising solutions.
- Problem Solving: Excels in complex domains such as competitive programming (NP-hard problems), mathematical optimization (e.g., circle packing), and other discovery-oriented tasks.
Good For
- Automated Program Improvement: When integrated with evolutionary scaffolds, it can iteratively refine and improve code or solutions.
- Scientific Discovery: Tasks requiring the generation and evaluation of novel solutions in fields like mathematics, engineering, or computational science.
- Competitive Programming: Outperforms base models on challenging competitive programming problems by composing diverse strategies.
- Optimization Problems: Solving complex optimization tasks where iterative refinement and quality assessment are crucial.