Build a tool-using agent from scratch around a small open LLM (Qwen2.5-1.5B-Instruct). Write three tools (an AST-based safe calculator, a knowledge-base lookup and a unit converter), a ReAct prompt, a robust parser that turns malformed model output into error observations, and the Thought → Action → Observation loop with a max-steps guard. Then score the agent on ten tasks, compare it with answering directly, and chart every trace. The first run downloads the model (about 3 GB).
Tools & Model
Tasks & Parsing
The ReAct Loop
Evaluation
Premium Project
This guided, GPU-powered project is available to Premium members. Upgrade to train models end-to-end in an interactive notebook.