点击任意处关闭 · Esc

Skip to content

AI 原生开放教材

Main Navigation

1. AI 基础与模型认知

2. 办公与知识工作

3. AI 编程与 Vibe Coding

4. 智能体工程

5. Harness 与编码 Agent

6. 上下文、记忆与技能

7. 个人智能体

8. 本地与端侧 AI

课程库来源总表关于

外观

Sidebar Navigation

The Context Course

projects

inference

.agents

skills

huggingface-local-models

.claude

agents

benchmarker

inference lab

optimizer

.opencode

agent

benchmarker

inference lab

optimizer

.pi

agents

benchmarker

optimizer

llama.cpp Inference Pi Coordinator

prompts

inference lab

docs

Claude Code Subagents Guide

Codex Subagents Guide

OpenCode Workflow

Pi Subagents Guide

Program

llama.cpp Inference Autoresearch

research

Do Not Repeat

experiments

Experiments

Inference Notes

templates

Experiment:

post-training

.pi

agents

experiment worker

memory keeper

planner

reporter

researcher

reviewer

Post-Training Pi Coordinator

prompts

posttrain

docs

Pi Subagents Guide

Program

Post-Training Autoresearch

research

Post-Training Notes

src

eval

tasks

pre-training

.agents

skills

autolab-hermes-delegation

autolab-managed-experiment

autolab-reporter

hf-cli

.claude

agents

autolab

experiment worker

memory keeper

planner

reporter

researcher

reviewer

.opencode

agent

autolab

experiment worker

memory keeper

planner

reporter

researcher

reviewer

.pi

agents

experiment worker

memory keeper

planner

reporter

researcher

reviewer

Autolab Pi Coordinator

prompts

autolab

We made an open source AI lab of agents to train models

docs

archive

Historical Hosted Backend

Legacy Control Planes

Claude Code Subagents Guide

Codex Subagents Guide

Getting Started

Winning Autolab With Hermes Delegation

OpenCode Workflow

Pi Subagents Guide

Script Reference

Troubleshooting

Program

Multi-Agent Autoresearch

research

campaigns

Optimizer Tuning Campaign

Campaign Notes

Regularization Tuning Campaign

Autolab Do-Not-Repeat Ledger

experiments

Experiment: Increase embedding weight decay from 0.0005 to 0.001 to test whether stronger regularization on token embeddings improves validation bpb, following the success of the lm_head weight decay increase.

Experiment: Lower FINAL_LR_FRAC from 0.025 to 0.018 to continue the winning direction of lower final LR floor.

Experiment: Increase lm_head weight_decay from 0.003 to 0.004 to continue the winning direction of stronger output layer regularization.

Experiment: Lower MATRIX_LR from 0.042 to 0.038 to test whether slower learning of Muon matrix parameters improves validation bpb.

Experiment Notes

Experiment: Lower SCALAR_LR from 0.7 to 0.6 to test whether slower learning of per-layer residual and initial scaling parameters improves validation bpb.

Experiment: Increase SCALAR_LR from 0.7 to 0.8 to test whether faster learning of per-layer residual and initial scaling parameters improves validation bpb.

Experiment: Increase UNEMBEDDING_LR from 0.01 to 0.015 to test whether a higher learning rate for the output layer improves validation bpb.

Experiment: Increase value_embeds weight_decay from 0.004 to 0.005 to test whether stronger regularization on value embeddings improves validation bpb.

Experiment: Decrease WARMDOWN_RATIO from 0.825 to 0.75 to allow more time at higher learning rates before the final LR floor.

Autolab Notes

Hugging Face Paper Ideas

templates

Campaign:

Do-Not-Repeat Template

Experiment:

Multi-Autoresearch

The Context Course

units

en

unit0

Welcome to The Context Course

unit1

Building Your First Skill

Unit 1: Agent Skills

Quiz 1: Understanding Skills and the Specification

Quiz 2: Building and Using Skills

The SKILL.md Format

Using Skills with Code Agents

What Are Agent Skills?

unit2

Building MCP Servers with Python

Gradio MCP Integration: Web UIs + MCP Servers

Hands-On: Build and Deploy an MCP Server

Introduction to Model Context Protocol

MCP Key Concepts and Architecture

Configuring Agents as MCP Clients

Quiz 1: MCP Fundamentals

Quiz 2: MCP in Practice

unit3

Plugin Anatomy

Building Your Own Plugin

Unit 3: Plugins

Quiz 1: Plugin Fundamentals

Quiz 2: Building and Distributing Plugins

Using Plugins

unit4

Hands-On: Multi-Agent Workflow

Unit 4: Subagents

Subagent Patterns

Quiz 1: Subagent Concepts

Quiz 2: Multi-Agent Workflows

Using Subagents

unit5

Hands-On: Build an Agent Activity Dashboard with Gradio

Hook Events and the Agent Lifecycle

Unit 5: Hooks

Quiz 1: Hook Fundamentals

Quiz 2: Hooks in Practice

unit6

The Agentic Loop Deep Dive

Hands-On: Extending Nano Harness

Introduction to Nano Harness

Quiz: Nano Harness and Agent Internals

Tools and Sandboxing in Detail

课程首页

出处与许可

来源信息

本页目录

课程库›上下文、记忆与技能›The Context Course

The Context Course系统课程·英文原文 ↗原件 ↓仓库 ↗

Do-Not-Repeat Template

Use this template when turning a failed or stale experiment into durable guidance.

Entry

Parent master hash: <hash>

Pager

上一篇Campaign:

下一篇Experiment:

AI 原生开放教材·GitHub 仓库·@faruheaisha·署名与许可编排与站点代码 MIT · 课程内容著作权归各上游作者