experiment-analysis

Diagnose GRPO training runs using WandB metrics and Axiom logs.

1|Updated Dec 2, 2025
One-click install
npx skills add https://github.com/bglick13/diplomacy-v2 --skill experiment-analysis
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: experiment-analysis
Source: https://github.com/bglick13/diplomacy-v2/tree/main/.claude/skills/experiment-analysis
Command: npx skills add https://github.com/bglick13/diplomacy-v2 --skill experiment-analysis

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires wandb, and includes scripts (resource) components.

What problem does it solve?

This Skill helps diagnose training runs by extracting Elo metrics, tracking learning dynamics, and surface issues in the GRPO pipeline.

Core Features & Use Cases

  • Elo trajectory analysis: Retrieve and interpret Elo progression across checkpoints.
  • Metrics extraction: Pull training metrics from WandB and Axiom logs for diagnostics.
  • Reports & dashboards: Generate summaries and highlight actionable insights for experiment trackers.

Quick Start

Use the provided tool to fetch Elo metrics for a run, for example: uv run python .claude/skills/experiment-analysis/analyze_elo.py <run-name>

Frequently Asked Questions about experiment-analysis

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I analyze Elo progression across GRPO training checkpoints?

Elo trajectory analysis retrieves and interprets Elo progression metrics from WandB across training checkpoints. Run the provided script with your run name to fetch Elo data, then extract initial, peak, and final values to diagnose learning dynamics and identify performance plateaus or regressions in your experiment.

What metrics can I extract from WandB and Axiom logs for experiment diagnostics?

Extract training metrics, Elo scores, throughput data, and error logs from WandB and Axiom to diagnose GRPO pipeline issues. The Skill pulls these metrics to surface learning dynamics, adapter performance, baseline comparisons, and actionable insights for experiment tracking across multiple runs.

Can I use this Skill to compare Elo performance across different adapters and baselines?

Yes, the Skill identifies checkpoints and baseline runs, then generates formatted comparison tables showing Elo metrics across adapters. This lets you diagnose relative performance, training stability, and experiment outcomes side-by-side to isolate which configurations drive learning gains.

How do I diagnose issues in my GRPO training pipeline using experiment logs?

Diagnose GRPO training runs by fetching WandB metrics and Axiom logs to extract agent trajectories, compute Elo progressions, and identify errors or throughput bottlenecks. The Skill generates summaries and highlighted reports that surface pipeline issues and anomalies affecting experiment performance.

What do I need to set up before analyzing training runs with this Skill?

You need WandB integration configured for your GRPO runs and access to Axiom logs. The Skill requires run metadata, Elo metric availability, and properly tagged experiments to fetch and compute diagnostics; ensure your training pipeline logs to both platforms.

Does this Skill work with other tools in my metrics and monitoring stack?

The Skill integrates with WandB for metrics extraction and Axiom for log analysis. It outputs formatted tables and summaries for dashboards and experiment trackers, making results compatible with downstream reporting and visualization tools in your analytics workflow.