跳过主要内容
N

tunetune

nanshine/tunetune

TuneTune 在 DeepSeek Harness 对话旁提供工作台,可与 Agent 一起训练、测试和改进小模型。

安装

dsh plugin --profile web add github:nanshine/tunetune

README

English · 简体中文

TuneTune

Work with your Agent to train, test, and improve small models for your own tasks.

CI npm License: Apache-2.0

https://github.com/user-attachments/assets/878c9c3f-b6c5-428e-8026-65cab7ba558d

Quick start

Requirements: macOS / Linux, Python 3.10+; Windows via WSL2

Send this to your Agent:

Install TuneTune following https://github.com/nanshine/tunetune/blob/main/INSTALL.md

Then, in the folder where you want to start a study, try a request like:

Use TuneTune to train a model that names Pokémon from silhouettes in the anime’s “Who’s That Pokémon?” segment.

Already installed? Start with the request. Prefer to install it yourself? See Installation.

How it works

Tell your Agent what you want the model to do. It looks into the task, compares approaches, and proposes a plan with estimated time, downloads, and costs for you to review.

It then prepares test examples, measures a simple starting model, and improves it through rounds of experiments. Follow the plans, experiments, and results in the workbench, and try your own inputs to see how well the model works for you. Your Agent asks for your judgment when reference answers need review or the budget or direction needs to change.

Research workflow: you describe the need, approve the plan and budget, review reference answers, and finally try the model; the ghost investigates and writes the proposal, prepares samples and measures a baseline, then plans, runs and analyzes experiment rounds until it verifies and delivers a model; at any time you can change the goal, question scores or suggest a method.

SkillRole
tunetune-initSet up the study, understand the task, research approaches, and agree on a plan and budget.
build-evaluationPrepare test examples and scoring, check that the tests are useful, and measure a starting model.
run-experimentsTry improvements in rounds, compare results, and verify that the final model runs locally.
research-workspaceKeep and display study records so your Agent can continue from earlier work.

When the goal is met, the budget runs out, or no worthwhile next experiment remains, your Agent explains the results, limitations, and recommended next step.

Installation

Claude Code

/plugin marketplace add nanshine/tunetune
/plugin install tunetune@tunetune

Codex

codex plugin marketplace add nanshine/tunetune
codex plugin add tunetune@tunetune

DeepSeek Harness

Install dsh-tunetune from Plugins to get the skills and a workbench beside your conversation.

For a CLI-managed profile:

dsh plugin --profile <profile> add dsh-tunetune

Other Agents with skills support

Install the skills for your Agent, replacing <application> with its identifier, such as cursor or opencode:

npx skills add nanshine/tunetune --global --agent <application> -y

Contributing

Find the source code on GitHub. Report an issue, share a real study, or submit a pull request. For code changes, see the repository's contribution guidelines.

License

Apache-2.0.

相关插件