wandb-report

W&B Report Generator

Create Weights & Biases reports for weekly mentor reviews and experiment updates.

Workflow

Follow these 3 phases in order. Never skip the outline phase.

Phase 1: Gather Context

Before drafting anything, gather this information:

1. Project/Entity: Ask or detect from context/environment

   • Check for WANDB_API_KEY in environment or .env
   • Look for wandb config in the project

2. Runs to Include: Ask what experiments/runs to include

   • Examples: "last 4 runs", specific run names, runs matching a pattern
   • Query wandb API to get available run names and metrics

3. Story to Tell: Ask what the report should communicate

   • What's the main finding or comparison?
   • Example: "FooIntervention works better than BarIntervention"

Phase 2: Draft Outline (CRITICAL)

ALWAYS draft a markdown outline BEFORE writing any code.

Present the outline for user critique. Include [bracketed descriptions] of where plots will go.

Outline Template

# Report: [Title]

## Key Finding
- Main takeaway: [one sentence summary]
- [CalloutBlock: "quote the key result here"]

## Setup
- Brief bullet points describing what was compared
- [No plot - just text]

## Training Curves
- Show how loss evolved over training
- [Full-width LinePlot: train/loss for all runs, x=Step, y=Loss]

## Results Comparison
- Compare final metrics across runs
- [Half-width LinePlot: eval/metric_A] [Half-width LinePlot: eval/metric_B]

## Conclusion
- Brief interpretation
- Next steps (if any)

Wait for user approval. Iterate on the outline until the user is satisfied with:

• The story structure
• What plots are included
• Chart layouts (full-width vs side-by-side)

Phase 3: Create Report

Only after outline approval:

1. Generate Python code using wandb_workspaces API
2. Execute the code to create the actual report
3. Return the report URL to the user

Layout Rules

Layout	Width	Use Case
Full-width	w=24	Default for all charts
Half-width	w=12	Side-by-side comparisons only
Never use	w<12	Too small to read