Skip to content

LLM Atlas

Main Navigation At a Glance Base Models

Overview

Black-Box Distillation

OPD (On-Policy Distillation)

English

English

Appearance

Sidebar Navigation

Guide

At a Glance

How to Read This Atlas

Notation

Information Theory Basics

GPU Basics

A/B Testing & Statistics

Base Models

Overview

Qwen (Alibaba)

DeepSeek

GLM (Zhipu)

Llama (Meta)

Kimi (Moonshot)

MiniMax

Step (StepFun)

Gemini (Google)

Claude (Anthropic)

GPT (OpenAI)

Post-training

Overview: Two Dimensions

Supervised Fine-Tuning

Overview

Full Fine-Tuning

Data Construction

Chat Template

Sequence Packing

Loss Masking

LoRA Family

Overview & Comparison

LoRA

QLoRA

DoRA

AdaLoRA

rsLoRA

LoRA+

PiSSA

DPO Family

Overview

DPO

IPO

KTO

ORPO

SimPO

CPO

PPO / GRPO Family

Overview

Reward Model

PPO

GRPO

DAPO

GSPO

CISPO

RLOO

REINFORCE++

Black-Box Distillation

Overview

Black-Box (Data/CoT)

OPD (On-Policy Distillation)

Overview

White-Box (Logits KL)

Inference & Decoding

Overview

KV Cache & PagedAttention

Quantization (GPTQ/AWQ/FP8)

Speculative Decoding

Harness

Overview

Agent Loop & Context

Sandbox & Tool Execution

Systems Compared

Agent

Overview

Tool Use Training

Agentic RL

Deep Research

Overview

Mind DeepResearch (Li Auto)

AgentDisCo

DR-Rubric

Multi-Agent

Skills

Agent Skills

Skill Design & Evaluation

Skills vs RAG vs Fine-Tuning

On this page

gspo

Translation pending

Full translation pending; see the Chinese version.

Edit this page on GitHub

Last updated:

Pager

Previous pageDAPO

Next pageCISPO

知乎小红书 X GitHub

gspo ​

gspo