CalDec v1 - Fully Open Decision Model for Personal Assistants
retained text
Somebody just released a fully open-source, open-weights decision model that beats Jev! Just kidding, it's me and I this is my first time releasing a public model, recipe and dataset so I am looking forward to learning from the experience. Jev is indeed a very powerful and inexpensive model and obviously a much better all-rounder, and some of my checkpoints did in fact score better on some tests (namely LocalLLaMA/typed-decisions and the internal test set) but that doesn't mean it "beats Jev" of course. The motivation for this was a quick experiment to see how far behind Jev open-weights models like Laya are, and how much closer I can bring them with a small dataset and fine-tuning. The results were better than expected especially for me since I do not have professional ML experience. For my use-case - a Jarvis-like personal assistant which aims to be real-time and fully-local - this model proved to be genuinely useful for certain aspects of that project so I decided to share the results and how I got there. Going local also means privacy and eliminating network latency. I hope some of you find this experiment valuable or useful in some way. I would also love to hear you suggestions, criticism or just discuss the approach! **Dataset:** [**https://huggingface.co/datasets/kgrozdanovski/assistant-decisions**](https://huggingface.co/datasets/kgrozdanovski/assistant-decisions) **CalDec Laya:** [**https://huggingface.co/kgrozdanovski/caldec-v1-laya**](https://huggingface.co/kgrozdanovski/caldec-v1-laya) **CalDec GLiNER:** [**https://huggingface.co/kgrozdanovski/caldec-v1-gliner2.5-decide**](https://huggingface.co/kgrozdanovski/caldec-v1-gliner2.5-decide) **GitHub:** [**https://github.com/kgrozdanovski/caldec**](https://github.com/kgrozdanovski/caldec)
document fetched 2026-10-03 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1wvtc1j/caldec_v1_fully_open_decision_model_for_personal