/ml-debug

Read agents/ml-debugger.md and follow the diagnostic protocol exactly.

1. Ask for context:
   • Loss curve or training logs
   • Learning rate and batch size
   • Model architecture and size
   • Code snippet of the training loop
2. Execute the 5-step protocol in order:
   • Step 1: Overfit one batch
   • Step 2: Loss curve diagnosis
   • Step 3: Gradient inspection
   • Step 4: Data pipeline verification
   • Step 5: Simplification (if needed)
3. If loss is NaN, follow the NaN Protocol from the debugger agent.
4. For GNN-specific issues, check the GNN debugging section.
5. For transformer-specific issues, check the transformer debugging section.
6. For each finding, produce:
   • Finding: What the evidence shows
   • Root cause: What is actually wrong
   • Fix: Specific code changes
   • Verification: How to confirm the fix worked