Automate GRPO training ablation sweeps with YAML configuration and Modal orchestration.
npx skills add https://github.com/bglick13/diplomacy-v2 --skill perform-sweep
Please help me install this Agent Skill. Skill: perform-sweep Source: https://github.com/bglick13/diplomacy-v2/tree/main/.claude/skills/perform-sweep Command: npx skills add https://github.com/bglick13/diplomacy-v2 --skill perform-sweep
End-to-end workflow for running ablation experiments on the Diplomacy GRPO training pipeline.
Create your sweep configuration under experiments/sweeps/<name>/sweep.yaml, then launch with: python scripts/launch_sweep.py experiments/sweeps/<name>/
High-intent search queries and answers about installing and using this skill.
Ablation sweeps systematically test different hyperparameter combinations to identify optimal settings. Create a sweep configuration in experiments/sweeps/<name>/sweep.yaml, then launch with python scripts/launch_sweep.py experiments/sweeps/<name>/. The Skill automates the full pipeline including Modal orchestration, WandB tracking, and result analysis.
Yes. This Skill automates end-to-end ablation experiments for the Diplomacy GRPO pipeline, testing variations in horizon length, scoring, and KL settings. It handles sequential execution, state persistence across runs, and integrated post-sweep analysis to ensure reproducible, trackable experiments.
The Skill implements auto-resume: if Modal times out (24-hour limit), the sweep automatically respawns from saved state. Progress is persisted after each training run, enabling easy recovery and rerun without restarting from scratch.
Yes. The Skill includes built-in dry-run validation to test your configuration without launching expensive training jobs. This catches configuration errors early and confirms your sweep setup is correct before committing resources.
WandB tagging tracks all ablation runs with consistent labels and metadata, enabling side-by-side comparison of results across sweeps. This provides visibility into experiment progress, reproducibility verification, and downstream analysis of which hyperparameter combinations perform best.
Sequential execution is enforced by design to preserve infrastructure resources. One training runs at a time; the Skill queues subsequent experiments in your sweep and executes them serially while tracking progress state for each.
VYRE-StudiosTrack ML experiments and model artifacts with automated logging and visualization.
ruiyangruiyiLog ML experiment metrics and configurations to Weights & Biases.
Clay-HHKTrack and visualize ML experiments and artifacts with wandb.
xqinagOrchestrate ML experiment batches across SSH-enabled GPUs with retries and resume.
