julien31/Soar-qwen-7b
Soar-qwen-7b is a 7.6 billion parameter Qwen-based language model developed by Julien Pourcel, Cédric Colas, and Pierre-Yves Oudeyer. It is fine-tuned using the SOAR (Self-improving Operators for Automated program Refinements) framework, specializing in program synthesis for the Abstraction and Reasoning Corpus (ARC) tasks. This model excels at generating Python programs to solve complex reasoning problems, leveraging a self-improving evolutionary search and learning from hindsight mechanism. It supports a context length of 131072 tokens.
Loading preview...
Model Overview
julien31/Soar-qwen-7b is a 7.6 billion parameter language model based on the Qwen architecture, developed by Julien Pourcel, Cédric Colas, and Pierre-Yves Oudeyer. This model is specifically fine-tuned using the SOAR (Self-improving Operators for Automated program Refinements) framework. SOAR is designed to enable language models to learn from their own experience through a "virtuous cycle" of evolutionary search and learning, addressing complex reasoning tasks that require discovering solutions from scratch.
Key Capabilities
- Program Synthesis for ARC-AGI: The model is specialized in solving tasks from the challenging Abstraction and Reasoning Corpus (ARC) by synthesizing Python programs.
- Self-Improvement: Utilizes an evolutionary search process where the LLM generates and refines candidate programs, learning from both successful and failed attempts through hindsight relabeling.
- Outperforms Larger Models: The SOAR framework, when used with open-weight LLMs like this one, has demonstrated significant outperformance compared to much larger closed-source LLMs on ARC tasks.
- Python Code Generation: Primarily generates Python functions to solve ARC tasks, given formatted prompts containing training and test examples.
Good for
- ARC-AGI Task Solving: Ideal for researchers and developers working on the Abstraction and Reasoning Corpus (ARC) and similar program synthesis challenges.
- Exploring Self-Improving AI: Provides a practical implementation of the SOAR framework for studying evolutionary program synthesis and learning from feedback.
- Code Generation for Pattern Recognition: Useful for scenarios requiring the generation of code that identifies and transforms patterns based on examples.