ornith-ai Releases Ornith-1.5-122B Qwen Finetune with Self-Improvement Loop
Ornith-ai has introduced Ornith-1.5-122B, a large Qwen finetune that expands the end-to-end self-improvement loop by jointly optimizing task generation, scaffold construction, and solution rollouts. The model builds upon previous iterations by continuously generating its own training tasks and improving its policy through reinforcement learning without relying solely on human-curated data. This release represents a significant step toward reducing reliance on extensive human-annotated datasets by enabling autonomous skill development and recursive self-improvement in large language models. Such capabilities allow open-source models to evolve more dynamically in specialized domains. Ornith-1.5-122B is available in GGUF format on Hugging Face, enabling local execution and deployment by community developers. It extends prior work by dynamically discovering effective strategies for solving self-generated tasks rather than relying on fixed harnesses.
## BACKGROUND
A self-improvement loop in large language models allows a system to generate its own practice problems and evaluate its solutions, establishing a closed-loop training process. Reinforcement learning solution rollouts involve generating multiple trial paths or outputs and using reward signals to guide the model toward optimal strategies.