一键在 Manus 中运行任何 Skill

autoresearch

星标69

分支6

更新时间2026年3月17日 01:32

Autonomous ML research loop inspired by Karpathy's autoresearch. Agent modifies code, trains on single GPU, evaluates metric, keeps or discards, repeats indefinitely. Use when asked to "run autoresearch", "autonomous training loop", or "overnight experiments".

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

在 Manus 中运行

来源

yxbian23

yxbian23/research-mate

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

Autoresearch — Autonomous ML Experiment Loop

Adapted from karpathy/autoresearch. Give an AI agent a GPU, let it experiment autonomously overnight.

Core Concept

The agent modifies training code → trains for a fixed time budget → checks if metric improved → keeps or discards → repeats. The human sleeps, the agent researches.

Quick Start

To set up autoresearch in the current project:

Copy the reference implementation (optional, for new projects):

cp third-party/autoresearch/prepare.py ./
cp third-party/autoresearch/train.py ./
cp third-party/autoresearch/pyproject.toml ./
uv run prepare.py  # one-time data setup

Or adapt to your existing project — you need:
- A training script that trains for a fixed time and outputs a metric
- A clear metric to optimize (e.g., val_loss, val_bpb, accuracy)

Setup

Agree on a run tag with the user (e.g., mar17).
Create branch: git checkout -b autoresearch/<tag>
Read source files to fully understand the codebase.
Initialize results.tsv with header: commit\tval_bpb\tmemory_gb\tstatus\tdescription
Run baseline: execute the training script as-is, record results.

The Experiment Loop

LOOP FOREVER:

Look at git state and past results in results.tsv.
Modify the training script with an experimental idea.
git commit the change.
Run: uv run train.py > run.log 2>&1 (redirect all output).
Read results: grep "^val_bpb:\|^peak_vram_mb:" run.log
If empty → crash. Run tail -n 50 run.log for traceback.
Log results to results.tsv (do NOT commit this file).
If metric improved → keep the commit, advance the branch.
If metric equal or worse → git reset --hard HEAD~1 to revert.

Rules

NEVER STOP. The user may be asleep. Keep going until manually interrupted.
Never ask "should I continue?" — you are autonomous.
Simplicity criterion: simpler is better. Removing code for equal perf = keep.
Crashes: fix trivial bugs (typos, imports), skip fundamentally broken ideas.
Timeout: if a run exceeds 2x the time budget, kill it and treat as failure.
One file: prefer modifying only the training script to keep diffs reviewable.

What You CAN Modify

Model architecture, optimizer, hyperparameters, training loop, batch size, model size.

What You CANNOT Modify

Data preparation, evaluation harness, tokenizer, fixed constants.
Do not install new packages.

Logging Format (results.tsv)

Tab-separated, 5 columns:

commit	val_bpb	memory_gb	status	description
a1b2c3d	0.997900	44.0	keep	baseline
b2c3d4e	0.993200	44.2	keep	increase LR to 0.04
c3d4e5f	1.005000	44.0	discard	switch to GeLU activation
d4e5f6g	0.000000	0.0	crash	double model width (OOM)

Reference

For the complete methodology, see: third-party/autoresearch/program.md For the reference training script: third-party/autoresearch/train.py

同仓库更多 Skills

同仓库

timesfm-forecasting

yxbian23/research-mate

Zero-shot time series forecasting with Google's TimesFM foundation model. Use for any univariate time series (sales, sensors, energy, vitals, weather) without training a custom model. Supports CSV/DataFrame/array inputs with point forecasts and prediction intervals. Includes a preflight system checker script to verify RAM/GPU before first use.

2026-03-1769

ai-research-slides

yxbian23/research-mate

为 AI 研究论文生成专业的分析演示文稿。按照标准化的学术分析框架，从 TLDR 到 Limitation 全面解读论文。适用于论文阅读汇报、组会分享、研究讨论等场景。

2026-03-1769

autoresearch-loop

yxbian23/research-mate

Domain-agnostic autonomous optimization loop. Try ideas, measure results, keep what works, discard what doesn't, repeat forever. Works for any quantifiable metric — ML training, build speed, bundle size, test performance, etc. Use when asked to "optimize X in a loop", "autoresearch for X", or "autonomous optimization".

2026-03-1769

course-learning

yxbian23/research-mate

CUHK PhD course learning assistant for lecture summarization, homework assistance, and exam review. Triggers on "课程", "作业", "复习", "homework", "review", "lecture", "考试", "/course". Supports bilingual output (Chinese main + English terms) for summaries, LaTeX homework with human-like writing, and comprehensive review documents.

2026-03-1769

dataset-processing

yxbian23/research-mate

Use this skill when processing large-scale ML datasets. Covers data loading, preprocessing, augmentation, multimodal data handling, and streaming/sharding techniques.

2026-03-1769

diffusers-workflow

yxbian23/research-mate

Use this skill when working with diffusion models for image/video generation. Covers Diffusers library pipelines, custom samplers, ControlNet, model training, and optimization techniques.

2026-03-1769

name	autoresearch
description	Autonomous ML research loop inspired by Karpathy's autoresearch. Agent modifies code, trains on single GPU, evaluates metric, keeps or discards, repeats indefinitely. Use when asked to "run autoresearch", "autonomous training loop", or "overnight experiments".

Autoresearch — Autonomous ML Experiment Loop

Adapted from karpathy/autoresearch. Give an AI agent a GPU, let it experiment autonomously overnight.

Core Concept

The agent modifies training code → trains for a fixed time budget → checks if metric improved → keeps or discards → repeats. The human sleeps, the agent researches.

Quick Start

To set up autoresearch in the current project:

Copy the reference implementation (optional, for new projects):

cp third-party/autoresearch/prepare.py ./
cp third-party/autoresearch/train.py ./
cp third-party/autoresearch/pyproject.toml ./
uv run prepare.py  # one-time data setup

Or adapt to your existing project — you need:
- A training script that trains for a fixed time and outputs a metric
- A clear metric to optimize (e.g., val_loss, val_bpb, accuracy)

Setup

Agree on a run tag with the user (e.g., mar17).
Create branch: git checkout -b autoresearch/<tag>
Read source files to fully understand the codebase.
Initialize results.tsv with header: commit\tval_bpb\tmemory_gb\tstatus\tdescription
Run baseline: execute the training script as-is, record results.

The Experiment Loop

LOOP FOREVER:

Look at git state and past results in results.tsv.
Modify the training script with an experimental idea.
git commit the change.
Run: uv run train.py > run.log 2>&1 (redirect all output).
Read results: grep "^val_bpb:\|^peak_vram_mb:" run.log
If empty → crash. Run tail -n 50 run.log for traceback.
Log results to results.tsv (do NOT commit this file).
If metric improved → keep the commit, advance the branch.
If metric equal or worse → git reset --hard HEAD~1 to revert.

Rules

NEVER STOP. The user may be asleep. Keep going until manually interrupted.
Never ask "should I continue?" — you are autonomous.
Simplicity criterion: simpler is better. Removing code for equal perf = keep.
Crashes: fix trivial bugs (typos, imports), skip fundamentally broken ideas.
Timeout: if a run exceeds 2x the time budget, kill it and treat as failure.
One file: prefer modifying only the training script to keep diffs reviewable.

What You CAN Modify

Model architecture, optimizer, hyperparameters, training loop, batch size, model size.

What You CANNOT Modify

Data preparation, evaluation harness, tokenizer, fixed constants.
Do not install new packages.

Logging Format (results.tsv)

Tab-separated, 5 columns:

commit	val_bpb	memory_gb	status	description
a1b2c3d	0.997900	44.0	keep	baseline
b2c3d4e	0.993200	44.2	keep	increase LR to 0.04
c3d4e5f	1.005000	44.0	discard	switch to GeLU activation
d4e5f6g	0.000000	0.0	crash	double model width (OOM)

Reference

For the complete methodology, see: third-party/autoresearch/program.md For the reference training script: third-party/autoresearch/train.py