What I just uploaded
I uploaded the Canvas Calendar Agent, version 7! There are three pieces on Hugging Face:
- the model https://huggingface.co/kleinpanic93/canvas-calendar-agent-v7-dpo
- the dataset, kleinpanic93/canvas-calendar-preferences-v7
- a demo Space https://huggingface.co/spaces/kleinpanic93/canvas-calendar-agent-demo
It belongs to the CS3704 Canvas Project at https://github.com/kleinpanic/CS3704-Canvas-Project, where it ships as release v3.0.0.
The model
It starts from google/gemma-4-E2B-it. I trained it on a DGX Spark in two steps, both full-parameter in bf16 with no LoRA and no quantization. SFT used 181 trajectory rows. DPO used 1,071 preference pairs, labeled by a Gemma-4-31B-IT-NVFP4 teacher with a 3-vote majority. The DPO run took 9:03. Per the card, it speaks the Gemma-4 tool-call format for 18 Canvas, Calendar and Study tools.
The card reports rewards/accuracies of 0.9032 and a train loss of 0.2229. Those come from the training log. They are not a test on unseen data.
The dataset
It has preference, KTO and trajectory splits. The preference split is 1,071 train and 75 test rows. KTO is 146 and 37. Trajectories are 181 and 46. Per the model card, it comes from Virginia Tech Canvas sessions with course numbers, emails, phones and names scrubbed.
The Space
The demo is Gradio. Its tool calls are simulated with mock data. It crashed on version pins for a while. The real cause was the transformers version: the gemma4 architecture needs a newer release than the one the card lists.
What is not done
The dataset has no card yet. The demo tools are mocks. I have not published a test on unseen data.
P.S. Thanks for reading. Consider subbing to the RSS feed.
Klein Panic