MCP Agent Fine tuning Workshop
summary
Agent runs on MCP servers using a Playwright browser, generates and saves MCP browser tool traces (tools list + multi‑turn messages), pushes an unrolled dataset to Hugging Face (nine rows from four sessions), and fine‑tunes a Qwen3 model via Qwen3 fine‑tune with unsloth using LoRA adapters (4B base, train adapters only, rank 32, batch size 1, one epoch). Setup uses an OpenAI‑style endpoint (port 8000) and RunPod/H100 templates; success is whether the fine‑tuned model reliably emits JSON tool calls.
AI-written summary · about
chapters (18)
- 0:00MCP agent fine-tuning intro
- 1:02Model Context Protocol
- 5:14Workshop environment setup
- 6:38Run MCP agent with Quen model
- 9:25Demo: Trellis.com navigation
- 12:16Agent logs for fine-tuning
- 13:20High-quality traces generation
- 16:10Improving traces with prompts
- 17:58Push traces to Hugging Face Hub
- 21:07MCP agent code deep dive
- 22:54Preparing for fine-tuning
- 24:16Inspect fine-tuning data
- 26:02Benchmark before fine-tuning
- 27:48Apply LoRA adapters
- 28:53Training the fine-tuned model
- 31:20Post-training performance eval
- 32:23Supervised vs reinforcement learning
- 34:30Workshop recap
related talks
50mBeyond local tools: Dive into the Model Context Protocol (MCP)
01h 01mAI Agents In-Depth – Function Calling, MCP and Tool Use Under the Hood
01h 03mAI Coding Agents and how to code them
36mPutting AI Agents to Work: A New Era of Open Source Connectivity with MCP
46mMCP for LLM Agents, Infrastructure APIs & Data Services
46mMCP and A2A Revolutionizing Interconnection and Collaboration of AI Agents
30mComposable AI: Building Next-Gen AI Agents with MCP
40mBuilding Agents with Agent Bricks and MCP