I can't tell you how much I appreciate your comprehensive model card notes. It is so nice to see someone that actually explains all the details of what they have created and you go way above and beyond what almost anyone else does.
I'm curious about whether the mlx conversions of these Ornith models has an impact on the self-learning capabilities of the model. It has been explained to me that there's a python script that runs with the model that is supposedly improving it as it runs so that it is learning in the context of the model's use in an agentic system.
Thanks again for creating this and for the in-depth explainers.
The self-improving description in the model refers to the process used during the model's training. DeepReinforce AI trains Ornith-1.0-35B using reinforcement learning (RL), and during this process, it jointly optimizes both the solutions the model produces and the agentic scaffolding that drives those solutions. So self-improvement here is something that happens during the training stage, not a mechanism that continues after the model is released.
I added this information to the model card to explain the self-improvement process that is also mentioned in the base model's own card. On my end, I did not use any additional self-improvement script or agentic solution, I only performed the quantization.
That said, in agentic coding tasks with LLMs, solutions like skill files or memory.md are already commonly used under the umbrella of context engineering. What you're referring to as a script that learns at runtime might be related to this, but that would not be something that updates the model's weights, rather it would be external context/memory files used by the agent.
Please let me know if there's anything I've misunderstood or missed, thank you.
