Meta · Multimodal LLMs
Writing archive黑头呆鱼进化之旅
Notes on models,
systems, and being human.
Every story from my original archive now has a complete English and Chinese edition—not just a bilingual label—and they all wear the same warm editorial outfit.
102 posts
LLM & NLP
论文深读｜SLA2: Sparse-Linear Attention with Learnable Routing and QAT
按顺序只读四块：Abstract → Method Overview 图 → Contributions → 主实验表。
LLM & NLP
Paper Deep Dive | SLA2: Sparse-Linear Attention with Learnable Routing and QAT
Today we skip paper lists and do one deep dive.
LLM & NLP
基于生成的大语言模型（LLM）评估：从生成到判断的机遇与挑战
人工智能（AI）和自然语言处理（NLP）的评估任务长期以来一直具有挑战性。传统的评估方法，例如基于匹配或嵌入的评估方法，在评估…
LLM & NLP
evaluation-of-generation-based-large-language-models-llms-opportunities-and-challenges-from-generation-to-judgment
Evaluation tasks in artificial intelligence (AI) and natural language processing (NLP) have long been challenging. Traditional evaluation methods, such as those based on matching or embeddings, are limited in assessing c…
LLM & NLP
SeCom: 重新定义对话AI的记忆管理
最近笔者一直在研究对话AI中的内存管理问题，特别是长期对话场景下的记忆构建与检索技术。发现了一篇令人眼前一亮的ICLR 2025论文——《SeCom: On Memory Construction and Retrieval for Personalized Conversational Agents》，由Microsoft和清华大学的研究团队联合发表。
LLM & NLP
SeCom: Redefining Memory Management in Conversational AI
I’ve recently been diving into memory management for dialog-based AI, especially how to construct and retrieve memories in long-term conversations. During my exploration I came across an eye-opening ICLR 2025 paper—**”Se…
LLM & NLP
Decoder-only与Encoder-only模型Padding策略的差异
在大模型 (LLM) 中，padding 是用于将不同长度的序列调整为同一长度的方法，以便于批量 (batch) 处理。
LLM & NLP
Differences in Padding Strategies Between Decoder-only and Encoder-only Models
In Large Language Models (LLMs), padding is a method used to standardize sequence lengths for batch processing.
LLM & NLP
MoE模型的的可视化指南：揭秘 MoE 在大型语言模型中的角色
原文地址：A Visual Guide to Mixture of Experts (MoE)
LLM & NLP
推理 LLM 的可视化指南：探索推理时计算技术与 DeepSeek-R1
DeepSeek-R1、OpenAI o3-mini 和 Google Gemini 2.0 Flash Thinking 是如何通过“推理”框架将 LLM（大型语言模型, Large Language Models） 扩展到新高度的典型示例。
LLM & NLP
A visual guide to MoE models: Uncovering the role of MoE in large language models
Original address: A Visual Guide to Mixture of Experts (MoE)
LLM & NLP
A visual guide to inference LLM: Exploring inference-time computing techniques with DeepSeek-R1
DeepSeek-R1, OpenAI o3-mini, and Google Gemini 2.0 Flash Thinking are prime examples of how LLMs (Large Language Models) can be scaled to new heights through “inference” frameworks.
Life
追逐与倒影
在清晨的第一缕光洒下之前，世间一切尚未显形。光与影的边界模糊，彷佛可以交叠，又彷佛注定分离。人们常说，朝阳是希望的象征，可它升起时，必将抛下一地影子。光和影之间，究竟是追逐还是相伴？这样的思考让我想起一则古老的寓言：一匹马在沙漠中追逐远方的绿洲，却不知道那不过是海市蜃楼，它每前进一步，绿洲也随之远去。
Life
Chase and Reflection
Before the first ray of light shines in the morning, nothing in the world has yet appeared. The boundaries between light and shadow are blurred, as if they can overlap, and yet they seem destined to separate. People often say that the rising sun is a symbol of hope, but when it rises, it will inevitably cast a shadow. Between light and shadow, is it a pursuit or a companionship? This kind of thinking reminds me of an old fable: a horse is chasing an oasis in the desert, but it doesn't know that it is just a mirage. Every time it takes a step forward, the oasis also disappears.
LLM & NLP
基于生成的大语言模型（LLM）评估：从生成到判断的机遇与挑战
人工智能（AI）与自然语言处理（NLP）领域中的评估任务长期面临挑战。传统的评估方法（如基于匹配或嵌入的技术）在判断复杂属性时效果有限。近期大语言模型（LLM）的发展催生了“LLM-as-a-Judge”范式，利用LLM对任务进行评分、排序或选择。本论文对LLM评估方法进行了全面综述，包括其定义、分类框架、评估基准，以及未来的研究方向。
LLM & NLP
Generation-based large language model (LLM) evaluation: opportunities and challenges from generation to judgment
Evaluation tasks in the fields of artificial intelligence (AI) and natural language processing (NLP) have long faced challenges. Traditional assessment methods, such as matching or embedding-based techniques, have limited effectiveness in judging complex attributes. The recent development of large language models (LLM) has given rise to the "LLM-as-a-Judge" paradigm, which uses LLM to score, sort, or select tasks. This paper provides a comprehensive review of LLM assessment methods, including its definition, classification framework, assessment benchmarks, and future research directions.
Life
身份与主体性的反思
当我接到永居申请被拒的消息时，短暂的愣神之后，内心涌动的却并非单纯的挫败，而是一种奇异的“生存论困境”感。表面上，这似乎只是一次行政结果的体现，但其背后却深刻折射了当代全球流动性结构与主体性建构之间的多重张力。
Life
Reflections on Identity and Subjectivity
When I received the news of my PR application being rejected, after a brief moment of shock, what arose within me was not merely frustration but a peculiar sense of “existential dilemma.” On the surface, it seemed like j…
Code
【Leetcode Python题解】「1346. Check If N and Its Double Exist」
使用哈希表可以显著优化时间复杂度。我们只需要一次遍历数组，同时用哈希表记录已经遇到的数字。
Code
[Leetcode Python problem solution] "1346. Check If N and Its Double Exist"
Using a hash table can significantly optimize the time complexity. We only need to traverse the array once and use a hash table to record the numbers we have encountered.
Code
【Leetcode Python题解】「2097. Valid Arrangement of Pairs」
给定一个二维数组 pairs，其中 pairs[i] = [start, end]，我们需要重新排列这些数字对，使得相邻的两个数字对 [start1, end1] 和 [start2, end2] 满足以下条件：
Code
Leetcode Python Solution - 2097. Valid Arrangement of Pairs
In this technical blog, we’ll dive deep into Leetcode Problem 2097 — Valid Arrangement of Pairs. We will break down the solution step by step, from understanding the problem, modeling it as a graph theory problem, to imp…
LLM & NLP
LoRA, DPO, KTO 与 SFT 技术详解
本篇文档将详细介绍几种在大型语言模型（如 LLAMA3）微调和优化中的重要技术，包括 SFT（Supervised Fine-Tuning）、LoRA（Low-Rank Adaptation）、Alignment 技术、KTO（Kahneman-Tversky Optimization） 和 DPO（Direct Preference Optimization）。文中还将详细阐述每种技术的原理、具体实现方法以及相应的损失函数与优化器选择…
LLM & NLP
Detailed Explanation of LoRA, DPO, KTO, and SFT Technologies
This document provides a detailed introduction to several important techniques used in fine-tuning and optimizing large language models (such as LLAMA3), including SFT (Supervised Fine-Tuning), LoRA (Low-Rank Adaptation)…
LLM & NLP
使用压缩有限状态机进行本地 LLM 的快速 JSON 解码
作者: Liangsheng Yin, Ying Sheng, Lianmin Zheng日期: 2024 年 2 月 5 日
LLM & NLP
Fast JSON decoding of native LLM using compressed finite state machines
Author: Liangsheng Yin, Ying Sheng, Lianmin Zheng Date: February 5, 2024
LLM & NLP
使用vLLM运行微调后的Gemma-2
在这里分享一下我运行微调后的Gemma-2-2b-it模型并使用vLLM的步骤，希望对其他人有所帮助。本文将详细介绍安装过程、环境配置以及常见问题的解决方法。
LLM & NLP
Detailed Steps for Running Fine-tuned Gemma-2-2b-it with vLLM
In this post, I will share the steps to run the fine-tuned Gemma-2-2b-it model using vLLM. This guide will cover the installation process, environment configuration, and common troubleshooting tips.
LLM & NLP
如何准确计算固定长度模型的困惑度（PPL）
困惑度（PPL）是评估语言模型最常用的指标之一。在深入探讨之前，我们应该注意这个指标特别适用于传统语言模型（有时被称为自回归或因果语言模型），而对于像 BERT 这样的 masked language models 则没有明确定义（见模型总结）。
LLM & NLP
How to accurately calculate the perplexity (PPL) of a fixed-length model
Perplexity Level (PPL) is one of the most commonly used metrics for evaluating language models. Before we delve deeper, we should note that this metric applies specifically to traditional language models (sometimes called autoregressive or causal language models) and is not clearly defined for masked language models like BERT (see model summary).
Code
【Python题解】2834. 找出美丽数组的最小和
给定两个正整数 n 和 target，目标是找到一个长度为 n 的数组，满足以下条件：
Code
[Python Problem Solution] 2834. Find the minimum sum of a beautiful array
Given two positive integers n and target, the goal is to find an array of length n that satisfies the following conditions:
Life
跟着GPT老师学小聊：如何做一个好的捧哏
英语小聊 (small talk) 是日常生活中不可或缺的交流形式，它不仅有助于打破沉默，也能在轻松的氛围中促进理解和友谊。在这篇文章中，我们将结合 10 个主题，提供相关的词汇、短语，并分享生活故事的开场白。同时，学习如何成为一名优秀的捧哏，通过提问和接话，让对话更加流畅和有趣。
Life
Follow the GPT teacher to learn how to chat: how to be a good commentator
English small talk (small talk) is an indispensable form of communication in daily life. It not only helps to break silence, but also promotes understanding and friendship in a relaxed atmosphere. In this article, we'll combine 10 themes, provide relevant vocabulary, phrases, and share life story starters. At the same time, learn how to be a good moderator and make the conversation smoother and more interesting by asking questions and answering questions.
Code
【Python题解】100226. 在带权树网络中统计可连接服务器对数目
你被给定一个未定根的加权树，它有 n 个顶点，代表从 0 到 n - 1 编号的服务器，一个数组 edges，其中 edges[i] = [ai, bi, weighti] 代表顶点 ai 和 bi 之间的双向边，边的权重为 weighti。你还被给定一个整数 signalSpeed。
Code
[Python problem solution] 100226. Count the number of connectable server pairs in a weighted tree network
You are given an unrooted weighted tree with n vertices representing servers numbered from 0 to n - 1, and an array edges where edges[i] = [ai, bi, weighti] represents a bidirectional edge between vertices ai and bi with edge weight weighti. You are also given an integer signalSpeed.
Code
【Python题解】100232. 超过阈值的最少操作数 II
给定一个从 0 开始的整数数组 nums 和一个整数 k。你可以进行如下操作：
Code
[Python problem solution] 100232. Minimum number of operations exceeding the threshold II
Given an array of integers nums starting from 0 and an integer k. You can do the following:
Code
【Python题解】100231. 超过阈值的最少操作数 I
给定一个下标从 0 开始的整数数组 nums 和一个整数 k。在每次操作中，你可以删除 nums 中的最小元素。目标是通过最少的操作次数使数组中的所有元素都大于或等于 k。需要返回实现此目标所需的最少操作次数。
Code
[Python problem solution] 100231. Minimum number of operations exceeding the threshold I
Given an array of integers nums indexed starting from 0 and an integer k. In each operation, you remove the smallest element in nums. The goal is to make all elements in the array greater than or equal to k with the minimum number of operations. Needs to return the minimum number of operations required to achieve this goal.
Code
【Python题解】2369. 检查数组是否存在有效划分
Leetcode 的题目 2369 要求我们检查一个整数数组nums是否可以划分为一个或多个满足特定条件的连续子数组。有效的划分条件包括：
Code
[Python problem solution] 2369. Check whether the array has valid divisions
Leetcode question 2369 asks us to check whether an integer array nums can be divided into one or more consecutive subarrays that meet certain conditions. Valid partition conditions include:
LLM & NLP
Gorilla LLM 大语言模型简介
🦍 Gorilla: Large Language Model Connected with Massive APIsLink: https://gorilla.cs.berkeley.edu/blogs/7_open_functions_v2.html
LLM & NLP
Introduction to Gorilla LLM Large Language Model
🦍 Gorilla: Large Language Model Connected with Massive APIsLink: https://gorilla.cs.berkeley.edu/blogs/7_open_functions_v2.html
Life
英语学习日记：De Facto
含义：’De facto’ 这个短语用来描述一些实际上存在的事物，即使它们没有被官方认可或法律确立。就像是在说“实际上”或“实践中”，而不是“理论上”或“官方上”。
LLM & NLP
FastChat 训练脚本代码逐行解析-Train.py 【FastChat 系列第 1 篇】
在本文中，我们将深入探讨 FastChat 的 train.py 脚本，这是一个用于训练和优化大型语言模型的关键组件。FastChat 是一个先进的开源平台，专注于开发、部署和评估基于大型语言模型（LLM）的聊天机器人。该平台不仅提供对顶尖模型如 Vicuna 和 MT-Bench 的支持，还包括一个分布式的多模型服务系统，配备了 Web UI 和与 OpenAI 兼容的 RESTful API，使用户能够高效地训练和评估他们的模型。
Life
De Facto: Unveiling the Power of an Intriguing Phrase
Hello, language enthusiasts! 🌐 Today, we’re diving into a fascinating phrase that often pops up in English conversations and writings: de facto. Let’s explore its meaning, origins, and how to use it effectively in senten…
LLM & NLP
FastChat Training Script Code Analysis - Train.py 【FastChat Series Part 1】
In this article, we delve into the train.py script of FastChat (https://github.com/lm-sys/FastChat) (https://github.com/lm-sys/FastChat/blob/main/fastchat/train/train.py), a key component for training and optimizing larg…
Life
英语学习日记：探索金融术语 'Giro Date'
今天的英语学习之旅中，我遇到了一个有趣的金融术语：“Giro date”。这个探索过程不仅丰富了我的词汇，还加深了我对英语中金融概念的理解。让我和你分享一下我是如何分析这个术语及其在金融世界中的重要性。
Life
English Learning Journey: Unraveling the Term 'Giro Date'
Today in my English learning journey, I encountered an intriguing financial term: “Giro date”. This exploration not only expanded my vocabulary but also deepened my understanding of financial concepts in English. Let me…
Debugging
如何解决多个 Git 服务的 SSH 密钥问题
在使用 Git 和不同的 Git 服务（如 GitHub 和 GitLab）时，可能会遇到 SSH 密钥的问题。本文将指导你如何设置和配置 SSH 密钥，以便可以同时与多个服务顺利工作。
LLM & NLP
理解大型语言模型中Fine-tuning和Further Pretraining的区别
在自然语言处理（NLP）领域，大型语言模型，如 GPT 和 BERT 的出现，彻底改变了我们处理文本分类、情感分析和问答等任务的方式。在这些模型的应用中，Fine-tuning（微调）和 Further Pretraining（进一步预训练）是两种关键技术。虽然它们看起来相似，但实际上服务于 NLP 流程中的不同需求和场景。
Debugging
How to Resolve SSH Key Issues with Multiple Git Services
When using Git with different Git services such as GitHub and GitLab, you may encounter SSH key issues. This article will guide you on how to set up and configure SSH keys so that you can work smoothly with multiple serv…
LLM & NLP
Understanding the Differences Between Fine-tuning and Further Pretraining in Large Language Models
In the world of Natural Language Processing (NLP), the advent of large language models like GPT and BERT has revolutionized how we approach tasks such as text classification, sentiment analysis, and question-answering. T…
LLM & NLP
🤖 十分钟用 Python 和 OpenAI 创建 Telegram 机器人！ 🚀
在这个有趣的教程中，我们将向您展示如何创建一个具有的 Telegram 机器人，该机器人能够与用户聊天并生成幽默回复。我们将详细解释每一步，让您轻松入门！
LLM & NLP
🤖 Create a Telegram Bot with Python and OpenAI in 10 Minutes! 🚀
In this fun tutorial, we’ll show you how to create a Telegram bot that can chat with users and generate witty responses. We’ll explain each step in detail, making it easy for you to get started!
Debugging
解决端口冲突：查找并终止进程
有时候，当你尝试启动一个应用程序或服务时，可能会遇到”Address already in use”（地址已经在使用）的错误，这意味着指定的端口已经被另一个进程占用。为了解决这个问题，你需要确定哪个进程正在使用该端口，并可以选择终止该进程或更改应用程序的端口配置。
Debugging
Resolving Port Conflicts: Identifying and Terminating Processes
Sometimes, when you try to start an application or service, you may encounter an “Address already in use” error. This means that the specified port is already in use by another process. To resolve this issue, you need to…
Code
无断开烦恼！远程服务器后台运行程序的3种方法：`nohup`、`tmux`和`screen`
在数据分析或机器学习项目中，经常需要在远程服务器上运行耗时长、计算密集型的任务。通过SSH连接到远程服务器是常见的操作方式。但是，如何确保在断开SSH连接之后，远程服务器上的程序能够继续运行呢？本文详细介绍了三种方法：nohup、tmux和screen。
Code
No disconnection worries! 3 ways to run programs in the background on remote servers: `nohup`, `tmux` and `screen`
In data analysis or machine learning projects, it is often necessary to run long, computationally intensive tasks on remote servers. Connecting to a remote server via SSH is a common way of doing this. But how to ensure that the program on the remote server can continue to run after disconnecting the SSH connection? This article details three methods: nohup, tmux and screen.
Code
超越Python的边界：`subprocess` 助你一键执行外部命令
在日常开发中，有时候我们希望能够从 Python 脚本中执行系统命令或者其他程序。Python 提供了 subprocess 模块，使得这一操作变得既简单又安全。
Code
Beyond the boundaries of Python: `subprocess` helps you execute external commands with one click
In daily development, sometimes we want to be able to execute system commands or other programs from Python scripts. Python provides the subprocess module to make this operation easy and safe.
LLM & NLP
conver pytorch model to onnx format
在机器学习和深度学习的开发过程中，模型的互操作性变得越来越重要。ONNX (Open Neural Network Exchange) 是一种开放格式，用于表示机器学习和深度学习模型。它允许开发者在各种深度学习框架之间轻松地共享模型，从而提高了模型的可移植性和互操作性。
LLM & NLP
conver pytorch model to onnx format
Model interoperability is becoming increasingly important during the development of machine learning and deep learning. ONNX (Open Neural Network Exchange) is an open format for representing machine learning and deep learning models. It allows developers to easily share models among various deep learning frameworks, thereby improving model portability and interoperability.
LLM & NLP
Training Llama 2 Model on Single GPU with int8 Quantization and LoRA
Llama 2 是一个包含预训练和微调的生成式文本模型的集合，其规模从 70 亿到 700 亿个参数不等。Llama2模型是由Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert等人在Llama 2: Open Foundation and Fine-Tuned Chat Models中提出的。
LLM & NLP
Training Llama 2 Model on Single GPU with int8 Quantization and LoRA
Llama 2 is a collection of pretrained and fine-tuned generative text models ranging in size from 7 billion to 70 billion parameters. The Llama2 model was proposed by Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert and others in Llama 2: Open Foundation and Fine-Tuned Chat Models.
LLM & NLP
LONGNET - Scaling Transformers to 1,000,000,000 Tokens
在本篇文章中，我们将详细讨论一个近期发布的先进模型——“LongNet”。该模型由微软亚洲研究院研发，于大约两周前正式公布。LongNet基于Transformer模型构建，其核心理念在于拓展Transformer的应用规模。值得一提的是，研究团队成功地将其扩展至处理10亿个令牌的规模。对于熟悉语言模型的人来说，会明白序列长度对模型性能的影响，因为序列长度决定了在执行注意力机制时，能够关联的令牌数量，从而影响模型可以获取的上下文信息长度…
LLM & NLP
LONGNET - Scaling Transformers to 1,000,000,000 Tokens
In this article, we will discuss in detail a recently released advanced model - "LongNet". The model was developed by Microsoft Research Asia and was officially announced about two weeks ago. LongNet is built based on the Transformer model, and its core concept is to expand the application scale of Transformer. It is worth mentioning that the research team successfully scaled it to handle 1 billion tokens. For those who are familiar with language models, they will understand the impact of sequence length on model performance, because the sequence length determines the number of tokens that can be associated when executing the attention mechanism, thus affecting the length of contextual information that the model can obtain...
LLM & NLP
Prompt Engineering
Prompt Engineering, 也被称为上下文提示，是指在不更新模型权重的情况下，与LLM（语言模型）进行交互以引导其产生期望输出的方法。它是一门实证科学，提示工程方法的效果在不同模型之间可能会有很大的差异，因此需要进行大量的实验和试探。
LLM & NLP
Prompt Engineering
Prompt Engineering, also known as contextual prompting, is a method of interacting with an LLM (language model) to guide it to produce a desired output without updating the model weights. It is an empirical science, suggesting that the effects of engineering methods can vary widely between models, requiring extensive experimentation and heuristics.
Code
ColoredLogger-彩色打印日志到控制台并记录到文件
本文档介绍了一个名为 ColoredLogger 的日志记录器类，它可以根据不同的消息类型以不同的颜色打印日志，并将日志记录到文件中。该类使用了 colorama 库来实现在控制台中显示带颜色的文本。为了使控制台输出的日志更加易于阅读和理解，我们通常会使用彩色的输出。同时，将日志记录到文件中可以方便我们后续的调试和分析。在Python中，我们可以使用logging和colorama库来实现这样的功能。
Code
ColoredLogger - prints logs in color to the console and records to files
This document introduces a logger class called ColoredLogger, which can print logs in different colors according to different message types and log to a file. This class uses the colorama library to display colored text in the console. In order to make the console output logs easier to read and understand, we usually use colored output. At the same time, recording logs to files can facilitate our subsequent debugging and analysis. In Python, we can use the logging and colorama libraries to achieve such functionality.
LLM & NLP
详解梯度下降算法
梯度是一个在微积分中使用的重要概念，它用于衡量函数在给定点上的方向导数沿各个方向最大时的最大值。对于一个标量函数，梯度的方向是函数增长最快的方向，而梯度的反方向则是函数减小最快的方向。
LLM & NLP
Detailed explanation of gradient descent algorithm
Gradient is an important concept used in calculus that measures the maximum value of a function at a given point when its directional derivative is maximum in all directions. For a scalar function, the direction of the gradient is the direction in which the function grows fastest, and the opposite direction of the gradient is the direction in which the function decreases fastest.
Engineering
Hexo博客添加可交互式足迹地图
这篇文章将向你展示如何在基于Hexo和Next的GitHub Pages博客中创建一个交互式的世界地图页面，这个地图将展示你曾经访问过的城市，你可以根据你对每个城市访问的频率在地图上显示不同颜色的标记，你还可以点击这些标记来显示更多关于这个城市的信息。
Engineering
Hexo blog adds interactive trail map
This article will show you how to create an interactive world map page in your GitHub Pages blog based on Hexo and Next. This map will display the cities you have visited. You can display different colored markers on the map based on the frequency of your visits to each city. You can also click on these markers to display more information about the city.
Debugging
VS Code Remote SSH连接失败问题
本文档针对VS Code中的Remote SSH插件在尝试连接远程服务器时出现”Failed to parse remote port from server output”错误的情况提供解决方案。作者在经过一系列的排查和尝试后，最终找到了解决的方法。
Life
SUP体验
立式单桨冲浪 (Stand-Up Paddleboarding, 简称SUP) 是一项起源于夏威夷现代冲浪运动的水上活动。冲浪者站在漂浮在水面上的木板上，通过使用单桨来推动自己在水中前进。这项运动近年来在世界各地迅速流行起来，成为许多人喜爱的水上活动之一。
Life
SUP体验
立式单桨冲浪 (Stand-Up Paddleboarding, 简称SUP) 是一项起源于夏威夷现代冲浪运动的水上活动。冲浪者站在漂浮在水面上的木板上，通过使用单桨来推动自己在水中前进。这项运动近年来在世界各地迅速流行起来，成为许多人喜爱的水上活动之一。
LLM & NLP
DSSM 模型详解
DSSM（Deep Structured Semantic Models）是微软提出的一个深度学习模型，用于学习文本的语义表达。DSSM首次在信息检索领域中被提出，用于处理查询-文档匹配任务，但很快被应用到了各种其他场景，如广告点击率预测，推荐系统等。
LLM & NLP
推荐系统离线评估指标详解
本文将详细介绍推荐系统中常用的离线评估指标，包括精确率（Precision）、召回率（Recall）、准确率（Accuracy）、F1-Score、NDCG、命中率（Hit Rate）、AUC、GAUC和对数损失（Log Loss）。这些指标对于评估推荐系统的性能和效果至关重要。这些指标对于评估推荐系统的性能和效果至关重要。
Debugging
VS Code Remote SSH connection failure problem
This document provides solutions to the problem of "Failed to parse remote port from server output" error when the Remote SSH plug-in in VS Code attempts to connect to a remote server. After a series of investigations and attempts, the author finally found a solution.
Life
SUP experience
Stand-Up Paddleboarding (SUP) is a water sport that originated from modern surfing in Hawaii. Surfers stand on a board that floats on the water and use a paddle to propel themselves through the water. This sport has rapidly grown in popularity around the world in recent years and has become one of many people's favorite water activities.
Life
SUP Experience
Stand-Up Paddleboarding (SUP) is a water activity that originated from modern surfing in Hawaii. Surfers stand on a floating board and use a single paddle to propel themselves forward in the water. This sport has rapidly…
LLM & NLP
Detailed explanation of DSSM model
DSSM (Deep Structured Semantic Models) is a deep learning model proposed by Microsoft for learning the semantic expression of text. DSSM was first proposed in the field of information retrieval to handle query-document matching tasks, but was quickly applied to various other scenarios, such as advertising click-through rate prediction, recommendation systems, etc.
LLM & NLP
Detailed explanation of offline evaluation indicators for recommendation systems
This article will introduce in detail the offline evaluation indicators commonly used in recommendation systems, including precision, recall, accuracy, F1-Score, NDCG, hit rate, AUC, GAUC and log loss. These metrics are crucial for evaluating the performance and effectiveness of recommendation systems. These metrics are crucial for evaluating the performance and effectiveness of recommendation systems.
Engineering
Kubernetes基础
前言：本文将介绍Kubernetes基础中的重要概念——Pod，以及它在Kubernetes中的应用和功能。首先，我们将详细解释Pod的基本概念和构成，包括它作为最小可部署单元的特点和包含的资源。然后，我们将探讨Pod在Kubernetes中的作用与功能，包括调度单位、网络单元、存储单元、生命周期管理和水平扩展。通过深入了解Pod，您将对Kubernetes中的核心概念有更全面的理解。
Engineering
Kubernetes basics
Preface: This article will introduce the important concept in the foundation of Kubernetes - Pod, as well as its application and functions in Kubernetes. First, we will explain in detail the basic concept and composition of Pod, including its characteristics as the smallest deployable unit and the resources it contains. Then, we will explore the role and functions of Pod in Kubernetes, including scheduling unit, network unit, storage unit, life cycle management and horizontal expansion. By gaining a deeper understanding of Pods, you will gain a more comprehensive understanding of the core concepts in Kubernetes.
Engineering
GitLog
当使用Git进行版本控制时，**git log**命令是一个有用的工具，它可以显示提交历史记录和分支之间的关系。
Engineering
GitLog
When using Git for version control, the git log command is a useful tool that can show commit history and the relationship between branches.
Code
FAISS向量查询简介
IndexFlatL2测量查询向量与加载到索引中的向量之间所有给定点之间的 L2（或欧几里得）距离。它很简单，非常准确，但也不会太快。
Code
Introduction to FAISS vector queries
IndexFlatL2 measures the L2 (or Euclidean) distance between all given points between the query vector and the vector loaded into the index. It's simple, very accurate, but not too fast.
Code
Intersection & Union
Python中的intersection()方法是用于获取两个集合的交集。这个方法是set对象的方法，因此它只能用于set对象中。
Code
Intersection & Union
The intersection() method in Python is used to obtain the intersection of two sets. This method is a method of the set object, so it can only be used in the set object.
Code
Python yield
在 Python 中，**yield** 是一个关键字，它通常用于生成器函数中，用于生成序列化的值而不需要将整个序列保存在内存中。
Code
Python -m 参数解释
当在命令行中输入 “python -m “ 时，这表示使用Python的内置模块运行程序，其中是要运行的模块的名称。这类似于在Python脚本中使用 “import “ 并运行 “module-name.run()”。
Code
Python yield
In Python, **yield** is a keyword that is commonly used in generator functions to generate serialized values without holding the entire sequence in memory.
Code
Python -m parameter explanation
When "python -m " is entered on the command line, this means to run the program using Python's built-in modules, where is the name of the module to be run. This is similar to using "import " and running "module-name.run()" in a Python script.
Engineering
Hexo博客的创建与部署
（生活反思）（代码编年史）（流浪癖笔记）NLP Insights（自然语言处理洞察）Tech Toolbox（技术工具箱）Travel Tales（旅行故事）Debugging Diaries
Engineering
Creation and deployment of Hexo blog
(Life Reflections) (Code Chronicles) (Wanderlust Notes) NLP Insights (Natural Language Processing Insights) Tech Toolbox (Technical Toolbox) Travel Tales (Travel Stories) Debugging Diaries
Travel
南洋的椰风海韵
轻柔的椰风，吹拂着南洋的海滨，翠绿的棕榈，舞动在碧蓝的天际。大海的波涛，如歌如泣，耳畔回荡。潮起潮落间，悠远涛声如诗韵鸣响，
Travel
The coconut breeze and sea charm of Nanyang
The gentle coconut breeze blows on the Nanyang seaside, and the emerald green palms dance in the blue sky. The waves of the sea, like songs and tears, echo in the ears. Between the ebb and flow of the tide, the distant sound of the waves resounds like the rhyme of poetry,