Introduction

In DSH workflows, model tuning and experiment validation often involve many repetitive operations. Developers need a mechanism that can automatically modify code, run training, and make decisions based on results, thereby quickly identifying effective experimental approaches.

Plugin Overview

This is a DSH skill maintained by satan9394, essentially an autonomous research agent. Inspired by karpathy/autoresearch, it aims to automate the experimental process: by modifying code, fixing training to 5 minutes, using val_bpb to determine quality, and operating in a retain-or-discard loop.

Core Capabilities

  • Automated loop: Implements a closed-loop process of “modify code → train → evaluate → decide”.
  • Lightweight design: Uses an ultra-lightweight program.md skill and supports single-file modifications.
  • High-efficiency experiments: Combined with fixed-duration training, it can perform around 100 experiments per night.

Installation and Reference

Because the official documentation does not provide explicit installation commands, it is recommended to visit the community directory page to view or install it directly:
Community Directory Page

Use Cases and Cautions

  • Intended users: Developers or researchers who need extensive experimental iteration.
  • Permission note: The plugin runs with the permissions of the current DSH process.
  • Security check: Before installing, it is recommended to review the source code and license.

Summary

This plugin brings autonomous research capabilities to DSH and reduces manual intervention through automated loops, making it suitable for scenarios that require high-efficiency experimental iteration.

References