training

Aggregate BetaOne training experiment progress, win/loss records, and eval-signal metrics.

1|2|Updated Apr 3, 2026

Search Tags:#metrics #history #training #benchmarks #experiments #eval #betaone

Ready to run

npx skills add https://github.com/dbyrne/ClawTheSpire --skill training-dbyrne

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: training
Source: https://github.com/dbyrne/ClawTheSpire/tree/main/.agents/skills/training
Command: npx skills add https://github.com/dbyrne/ClawTheSpire --skill training-dbyrne

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

BetaOne training monitoring aggregates and interprets experimental data to reveal training progress, performance trends, and potential issues fast.

Core Features & Use Cases

Eval-centric progress tracking: compares P-Eval and V-Eval across experiments to distinguish true skill gains from incidental improvements.
Historical context and risk signals: surfaces history windows, peak metrics, momentum, and buffer status to guide decision making.
Decision-support reporting: provides structured reports with status, architecture, encounter sets, and recommended next steps for experiments.

Quick Start

Invoke the BetaOne training monitor for a given experiment to view its current progress and eval trajectory.

Frequently Asked Questions about training

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

FAQPage Schema

How do I monitor BetaOne training experiments and track eval metrics?▼

To monitor BetaOne training experiments, you aggregate progress, win/loss records, and eval-signal metrics. This skill analyzes live and finalized experiments by comparing P-Eval and V-Eval trajectories to distinguish true skill gains from incidental improvements.

What data files are required to evaluate BetaOne training history and benchmarks?▼

Evaluating BetaOne training history requires access to four specific files: betaone_history.jsonl, benchmarks/eval.jsonl, benchmarks/value_eval.jsonl, and progress.json. These files provide the data needed to compute current state, trends, and actionable insights across gen cycles.

Can I analyze configuration impacts across multiple BetaOne training cycles?▼

Yes, you can analyze configuration impacts across multiple gen cycles. The skill evaluates history trends and configuration changes by aggregating experimental data to reveal performance trends and potential issues for rapid assessment and debugging.

What is the best way to generate decision-support reports for BetaOne experiments?▼

The best way to generate decision-support reports is by aggregating eval trajectories and historical context. This provides structured reports detailing status, architecture, encounter sets, and recommended next steps to guide training decision making.

How do I identify risk signals and momentum issues in BetaOne training?▼

To identify risk signals and momentum issues in BetaOne training, the skill surfaces history windows, peak metrics, and buffer status. This historical context helps reveal training progress and potential issues fast for rapid debugging.