kleinpanic93/canvas-calendar-agent-v7-dpo

VISIONConcurrency Cost:1Model Size:5.1BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:May 5, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

The kleinpanic93/canvas-calendar-agent-v7-dpo is a 5.1 billion parameter model, fine-tuned from google/gemma-4-E2B-it (2.7B parameters) using Direct Preference Optimization (DPO). It is designed to act as an intelligent agent for Canvas LMS users, generating actionable scheduling plans by interpreting Canvas state and calendar data. This model excels at producing tool calls in the native Gemma-4 format for 18 Canvas/Calendar/Study tools, making it highly specialized for academic planning and task management.

Loading preview...

Overview

kleinpanic93/canvas-calendar-agent-v7-dpo is a specialized language model fine-tuned from google/gemma-4-E2B-it (2.7B parameters) using Direct Preference Optimization (DPO). Its primary function is to assist students by analyzing their Canvas LMS data (assignments, courses, deadlines) and personal calendar to generate concrete, actionable scheduling plans, including study blocks and exam preparation.

Key Capabilities

  • Intelligent Scheduling Agent: Reads Canvas LMS state and calendar information to produce personalized study and scheduling plans.
  • Native Tool Calling: Generates tool calls using the native Gemma-4 format (<|tool_call>call:tool.name{args}<tool_call|>) for 18 distinct Canvas, Calendar, and Study-related tools.
  • DPO Fine-tuning: Trained with Direct Preference Optimization on 1,071 preference pairs, achieving a 90.3% accuracy in ranking held-out preference pairs, indicating strong alignment with desired outputs.
  • Robust Data Handling: Training data underwent a two-pass anonymization process to scrub CRNs and PII, ensuring privacy.
  • Gemma-4 Base: Built on a google/gemma-4-E2B-it decoder-only Transformer architecture, supporting an inference context of 8,192 tokens.

Limitations

  • Limited Training Data: Trained on a relatively small dataset (1,071 preference pairs, 181 trajectory rows), which may limit coverage of diverse Canvas course structures.
  • Virginia Tech Specific: Training data exclusively from Virginia Tech Canvas instances; generalization to other institutions is untested.
  • Gemma-4 Specific Tool Format: The tool-call format is specific to Gemma-4 and requires compatible chat templates.
  • Qualitative Reasoning: As a 2.7B parameter model, its qualitative reasoning ability is below that of larger models, though it benefits from DPO training guided by a 31B-IT teacher model.

Good for

  • Developing automated academic planning tools for Canvas LMS users.
  • Integrating intelligent scheduling and task management features into educational applications.
  • Researching DPO applications for specialized agentic tasks and tool-use generation.

Finetunes
2 models