核心结论：大模型「分类」没有唯一答案，不同维度切出来的类别完全不同，但有三条主线索最实用：① 按能力定位——基础模型（Base）、指令微调模型（Instruct）、推理模型（Reasoning），对应「会续写 → 会听话 → 会思考」三个层次；② 按应用领域——通用大模型（L0）、行业大模型（L1）、垂域大模型（L2），对应「通识教育 → 行业熟手 → 场景专家」；③ 按模态与架构——纯语言 / 多模态、稠密（Dense）/ 混合专家（MoE）。理解这套分类，你才能知道该选哪个模型、为什么同一个模型家族要分 base 和 chat 两个版本、以及「推理模型」凭什么更贵。
目录
- 一句话结论
- 为什么大模型「分类」说不清：维度不同、答案就不同
- 按能力定位分（Base / Instruct / Reasoning）：会续写、会听话、会思考
- Base 基础模型：只会续写的「百科全书式新员工」
- Instruct 指令微调模型：学会「听话」的对话助手
- Reasoning 推理模型：会「停下来想」的思考者
- 按应用领域分（L0 / L1 / L2）：通用、行业、垂域
- 按架构分：稠密 Dense vs 混合专家 MoE
- 按模态分：纯语言 vs 多模态
- 一张总表：三种分类维度的对照
- 完整可运行代码：用 HuggingFace 区分三种模型
- 对照表：症状 → 原因 → 解法
- 判断收尾
1. 一句话结论
先建立一个直觉 🎭：大模型分类像「给员工分职级」——同一个人的学历（能力定位）是一回事，他在哪个部门（应用领域）是另一回事，他是什么部门类型（架构）又是第三回事。你不能用「他是研究生」来回答「他是哪个部门的」，这两件事本就独立。
所以记三个最实用的分类框架：
- 能力定位（最重要）：Base（会续写）→ Instruct（会听话）→ Reasoning（会思考）。
- 应用领域：L0 通用（通识）→ L1 行业（熟手）→ L2 垂域（专家）。
- 架构 / 模态：Dense vs MoE、纯语言 vs 多模态。
一句话：先搞懂「能力定位」这一维，你就已经解决了 80% 的选型问题——其他维度是锦上添花。
2. 为什么大模型「分类」说不清：维度不同、答案就不同
很多人问「大模型分哪几类」，得到的答案五花八门，是因为大家都从不同维度回答：
- 有人说分「通用 / 行业 / 垂域」——这是按应用领域切。
- 有人说分「基础 / 对话 / 推理」——这是按能力定位切。
- 有人说分「稠密 / 稀疏」——这是按架构切。
- 有人说分「语言 / 视觉 / 多模态」——这是按输入类型切。
这些答案都对，但它们属于不同的坐标系。一个模型可以同时是「通用 + Instruct + MoE + 多模态」，就像一个人可以同时是「研究生 + 技术部 + 程序员」——四个标签来自四个维度，彼此不冲突。
⚠️ 理解这一点是读懂本文的前提：「大模型分几类」没有一个标准答案，只有「按这个维度分几类」。 接下来分别讲三条最常用坐标轴。
3. 按能力定位分（Base / Instruct / Reasoning）：会续写、会听话、会思考
这是最实用的分类维度，因为它直接对应「你下载模型时会看到的三个版本名」——几乎所有主流模型家族（Qwen、Llama、DeepSeek 等）都会发布这三个版本：
- Base 基础模型：预训练（pretraining）后的原始 checkpoint，只会「续写下一个 token」。
- Instruct 指令微调模型：Base 再经过指令微调（instruction tuning）+ 偏好对齐，学会「照着指令回答」。
- Reasoning 推理模型：Instruct 之上再做推理专用训练（如 RL 强化学习），学会「先思考再作答」。
打个比方 🎭：Base 像「读遍公司所有书但没上过班的新员工」——知识全有但不会办事；Instruct 像「培训过如何接指令的正式员工」——会照着规矩干活；Reasoning 像「会先想清楚再动手的老手」——遇难题会停下来琢磨。
参考资料：
- Base, Instruct, and Reasoning Models Compared -- Sebastian Raschka ⭐值得阅读
- Base vs Instruct vs Chat vs Reasoning Models -- LMVersity
4. Base 基础模型：只会续写的「百科全书式新员工」
Base 模型是预训练的直接产物——在海量文本上用「预测下一个 token」练出来的原始 checkpoint。
- 会什么：补全文本、续写故事、few-shot（给几个例子它能照着来）、计算「下一句最可能是什么」。
- 不会什么：它不会「正常回答问题」。你问「15% of 60」，它可能继续编一个问题、列一堆不相干的问题，或写成一段小作文——因为对它来说，这些都只是「合理续写」。
- 为什么存在：Base 是 Instruct / Reasoning 的起点原料，也是研究者做「继续预训练」「领域预训练」的基底。
用新员工类比 🎭：Base 是那个「把公司所有书都读了一遍、但一天班没上过」的新人——你问他怎么处理报销，他给你背一段 HR 手册的原文，而不是告诉你「填这张单子」。他知识渊博，但不会办事。
💡 对普通用户：别直接用 Base 模型。它不适合对话，你用起来会觉得「这 AI 怎么答非所问」——那是你选错了版本，不是模型差。
5. Instruct 指令微调模型：学会「听话」的对话助手
Instruct 模型在 Base 之上，用「指令-回答」的配对数据做了指令微调（instruction tuning），再用偏好数据（如 RLHF，人类反馈强化学习）对齐，让模型「学会听话」。
- 会什么：直接回答问题、遵循你指定的格式（JSON、列表、摘要）、按对话模板（chat template）跟你对话。
- 关键点：它要求特定的对话格式——特殊的 token、角色标记、turn 结束标记。格式给错了，质量会断崖式下降，看起来就像「这个模型很烂」。
- 为什么重要：Instruct 是绝大多数人日常用的模型——你打开 ChatGPT、Kimi、文心一言，背后基本都是 Instruct / Chat 类模型。
⚠️ 常见误区：很多人下载了 Base 模型当对话用，效果差就骂模型不行——其实是该用 Instruct 版本。这也是我在同样的模型为什么本地和云API差别那么大：量化精度、聊天模板与采样参数的三大元凶里强调的「聊天模板」问题的根源。
参考资料：
6. Reasoning 推理模型：会「停下来想」的思考者
推理模型（Reasoning Model）是近两年最热的一类，代表如 OpenAI 的 o 系列、DeepSeek-R1、Qwen 的 Thinking 版。它在 Instruct 基础上，用强化学习专门训练「先思考再回答」。
- 核心差异：普通模型「看到问题立刻答」，推理模型会先生成一段长思维链（chain-of-thought），把中间推理过程写出来，再给最终答案。
- 擅长什么：数学题、逻辑题、编程、复杂推理——需要「多步推导」的任务。
- 代价是什么：慢、贵、token 消耗大。它每个问题都「想半天」，推理成本比 Instruct 高好几倍。
用类比 🎭：Instruct 像「反应快的实习生」，问啥秒答，但复杂问题容易答错；Reasoning 像「会打草稿的老手」，先在草稿纸上算一遍再报答案，慢但准。
💡 选型判断：日常聊天、摘要、普通问答 → Instruct 够用且快；数学、推理、疑难编程 → 上 Reasoning。别为所有问题都用推理模型，那是拿大炮打蚊子。
参考资料：
7. 按应用领域分（L0 / L1 / L2）：通用、行业、垂域
第二条主坐标轴，按「服务于谁」切分，国内常用 L0 / L1 / L2 三层的说法（由行业专家提出，如中科院王金桥）：
- L0 通用大模型（基础通用）：泛化能力强、能「举一反三」，相当于完成「通识教育」。如 GPT、DeepSeek、Qwen 的通用旗舰。
- L1 行业大模型（行业）：在 L0 基础上，用特定行业数据（医疗、金融、法律语料）微调，掌握行业术语与知识体系，相当于「进入行业的熟手」。如金融领域的专用对话模型。
- L2 垂域大模型（垂直场景）：面向更具体的业务场景深度优化，相当于「解决特定问题的专家」。如针对某种单据识别的 OCR 大模型。
用类比 🎭：L0 是「上过通识课的大学毕业生」，L1 是「入职某行业干了三年的熟手」，L2 是「专门修某一类设备的老师傅」——越往下越专、越窄，但解决具体问题的效率越高。
参考资料：
8. 按架构分：稠密 Dense vs 混合专家 MoE
第三条坐标轴是「模型内部怎么组织」，主要分两类：
- 稠密模型（Dense）：所有参数在每个输入上都被激活。结构简单、行为可预期，但参数多、推理成本高。
- 混合专家模型（MoE，Mixture of Experts）：由多个「专家子网络」组成，每个输入只激活其中一部分专家。同样的总参数量下，单次推理只走一小部分专家，又快又省，还能把参数堆到万亿级。
用类比 🎭：Dense 像「一支全队都要上场的球队」，每个球员都参与；MoE 像「一个分工明确的大团队」，接到任务只派最对口的几个小组上场，其他人待命。
💡 在本专栏的《MoE 混合专家模型深度解析》中我们详细讲过 MoE 的机制，这里只记住结论：MoE 用「总参数大、单次计算小」换来了「大模型 + 低推理成本」，是当前万亿级模型的默认选择。
9. 按模态分：纯语言 vs 多模态
第四条坐标轴看「模型能处理什么输入」：
- 纯语言大模型（LLM，Large Language Model）：只处理文本，输入是 token，输出也是 token。
- 多模态大模型（LMM，Large Multimodal Model）：能处理文本、图像、音频、视频，比如「看图说话」「语音理解」。常见形式是视觉语言模型（VLM），如 GPT-4V、Qwen-VL。
用类比 🎭：纯语言模型像「只认字的编辑」，多模态模型像「能看图说话的记者」——都能产出文字，但后者多了「看图」这个输入通道。
💡 这是最独立的维度：多模态和「通用 / 专用」不冲突，也和「Base / Instruct」不冲突——一个模型可以同时是「多模态 + Instruct + 通用」。
10. 一张总表：三种分类维度的对照
把四条坐标轴放在一起看，才不会乱：
关键认知：这四个维度是正交的（互相独立），一个模型在这四个维度上各有一个标签。理解它们，你就看懂了模型命名里的后缀含义——-Instruct 是能力定位，-VL 是多模态，-MoE 是架构，-L2（如果标注）是应用层级。
11. 完整可运行代码（Runnable Code）：用 HuggingFace 区分三种模型
下面是一段复制粘贴就能跑的代码，用 HuggingFace 加载同一个模型的三个版本，直观对比 Base / Instruct / Reasoning 对同一句 prompt 的响应差异。运行前 pip install transformers。
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
# 用 Qwen 家族演示：同一个模型系列的三个版本（能力定位三档）
MODELS = {
"Base": "Qwen/Qwen2.5-0.5B", # 基础模型：只会续写
"Instruct": "Qwen/Qwen2.5-0.5B-Instruct", # 指令微调：会听话
# Reasoning 版用 DeepSeek-R1-Distill 演示（小参数推理模型）
"Reasoning": "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B",
}
prompt = "15% of 60 is" # 一个需要"先想再答"的数学问题
for kind, name in MODELS.items():
print(f"\n===== {kind} 模型：{name} =====")
tokenizer = AutoTokenizer.from_pretrained(name) # 加载对应分词器
model = AutoModelForCausalLM.from_pretrained(name) # 加载对应模型
inputs = tokenizer(prompt, return_tensors="pt") # 编码 prompt
out = model.generate(**inputs, max_new_tokens=40, # 生成 40 个 token 的续写
do_sample=False) # 贪心解码，结果确定
print("输出：", tokenizer.decode(out[0], skip_special_tokens=True))
怎么读这段代码：同一个 prompt 15% of 60 is，Base 版大概率会继续「编一句合理的续写」（答非所问）；Instruct 版会规规矩矩给出「15% of 60 is 9」；Reasoning 版会先输出一段「思考」再给答案。三行代码就能看出「能力定位」三档模型的本质差异——这也是第 3～6 节讲的概念的可运行落地。
12. 对照表：症状 → 原因 → 解法
13. 判断收尾
什么时候该用这套分类：你需要在「选模型 / 理解模型版本后缀 / 判断为什么某个模型不听话」时——先看能力定位（Base/Instruct/Reasoning）定版本，再看应用领域（L0/L1/L2）定专业度，最后用架构与模态补充约束。什么时候不必纠结：你只是用现成的 API 聊天、不自己部署模型时，这些分类对你只是「知道即可」的背景知识，不用逐个深究。判断标准只有一条——你是在「选模型」还是在「用模型」，前者需要这套分类，后者直接挑口碑好的通用 Instruct 版即可。
本专栏《LLM 大模型技术专栏》从 Transformer 注意力机制一路讲到模型选型、微调、推理优化与 Agent 落地，每篇都配一段可直接运行的代码，每周更新一篇。想按顺序读完，可在专栏页订阅。
参考资料（综合）：
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Core findings: LLM "Classes" do not have a single answer, and the categories cut across different dimensions are completely different, but there are three main lines of best practice: (1) Positioning by Competency — Base Model (Base), Command fine-tuning Model (Instract), Delineation Model (Reasoning), corresponding to the three levels of "continuing, listening, thinking"; (2) By application area — General LLM (L0), Industry LLM (L1), Facing LLM (L2), corresponding to "General Awareness Education Quiz Zun industry Quest Expert"; and (3) Models and architecture — Pure language / Multimodulated, dense (Dense)/ Mixed Expert (MoE). In order to understand this classification, you can see which model to choose, why the same model family needs to divide the base and chat versions, and why "the reasoning model" is more expensive.
Contents
- One conclusion.
- Why can't LLM say "classification": different dimensions, different answers.
-Base/Instract/Resoning: Relay, listen, think.
- Base Base Model: Only continues with Encyclopedia New Worker.
- Instract Command fine-tuning model: A dialogue assistant who learns to listen.
Ressoning Logic Model: Thinkers who "stop and think"
- By field of application (L0/L1/L2); common, industrial, perigee
- By structure: dense Dense vs Mixed Expert MoE
- By quartile: pure language vs multimodular
- A summary table: a comparison of three classification dimensions
- Full runable code: Distinguishing three models with HuggingFace
- Checklist: Symptoms, causes, solutions
- Finish your judgment.
1. Conclusion Comment
🎭: It's one thing to create a hunch: LLM is sorted like "segregation of employees" — the same individual's education (placement of competence) is one thing, the other thing about which department he is in (application area), the third thing is what type of sector he is (structure). You can't use "he's a graduate student" to say "who's his department," which is independent.
So remember the three most useful classification frameworks:
- Capability positioning (most important): Base (resumed) Instract (behaved) Reasing (thinking).
- Application area: L0 General (literacy) L1 Industry (advanced) L2 Peri field (expert).
- Structure / Moe: Dense vs MoE, pure language vs multimodular.
One thing: first you understand the dimensions of "capacity positioning," you've solved 80% of the selection problem -- the other dimensions are added.
2. Why is LLM "classification" unclear: different dimensions, different answers
A lot of people ask, "What kind of LLM is" and the answers come from a variety of different dimensions:
- Some say it's "general/industry/ field" -- it's cut by field of application.
- Some say "basics/dialogues/ reasoning" — it's position by ability.
- Some say it's dense/ thin -- it's structured.
- Some say "Language / Visual / Multi-modular" -- this is by type of input.
These answers are correct, but they belong to different coordinates. A model can be "Universal + Instract + MoE + Multimodular" as if a person could be "Postgraduate + Technical + Programmer" at the same time -- four labels from four dimensions, without conflict.
⚠️ Understanding this is the premise of reading this paper: "LLM is classified in several categories" does not have a standard answer, but only "by this dimension". Next is the three most commonly used axis.
3. Positioning by capacity (Base / Information / Ressoning): Relaying, listening, thinking
This is the most practical dimension of the classification, because it corresponds directly to "three versions you will see when you download the model" -- I'm sorry. Almost all mainstream model families (Qwen, Llama, DeepSeek, etc.) have three versions of publish:
- Base Basic Model: Original checkpoint after pre-training, only token.
- Instruct command fine-tuning model: Base is then fine-tuned by command + preferences and learns to "follow the instructions."
Reasoning Logic Model: On the Instract, then specialized training in reasoning (e.g. RL intensive learning), learning to "think first and answer."
For example: Base is like "new employee who read all the books of the company but never worked" -- knowledge is all but impossible; Instract is like "training official staff to take orders" -- Reasoning, like a "old man who'll think first and then do it" -- in a difficult situation, stop thinking.
References:
- Base, Instract, and Ressoning Models Company- Sebastian Raschka is worth reading.
- Base vs Instruct vs Chat vs Reasoning Models -- LMVersity
4. Base model: "Encyclopedia New Worker" only continues
The Base model is a direct product of pre-training -- the original checkpoint that was developed on big text with "predict the next token".
- What's it gonna be: a copy of the text, a follow-up story, a few-shot (for a few examples, it can be used) and a calculation of what's most likely to be the next sentence?
- It's nothing: it's not "normal." You ask, "15% of 60," it may continue to make up a question, list a bunch of irrelevant questions, or write a little piece of it. Because for it, these are just "reasonable continuations."
- Why exists: Base is the starting material of the Institute / Reasoning and the basis for a researcher's "continuing pre-training" "field pre-training".
🎭: Base is the new guy who "reads all the books of the company but doesn't go to class one day" -- you ask him how he handles the claims, and he gives you a copy of the HR manual, instead of telling you to fill this out. He knows everything, but he can't do anything.
💡 For ordinary users: Do not use the Base model directly. It's not for conversation, and you think, "How does this AI answer the question?" -- it's the wrong version, not the bad model.
5. Instract Command fine-tuning model: a dialogue assistant who learns to listen
The Instract model is based on Base, fine-tuned with command-responding data, then aligned with preferred data (e.g. RLHF, human feedback enhanced learning), so that the model " learns to listen."
- What is it? Answer the question directly, follow the format you have specified (JSON, List, summary), and talk to you on the dialogue template (chat text).
- Key points: It requires specific dialogue formats - special token, role tags, turn end tags. The format is wrong, the quality is falling off the cliff, and it looks like "this model sucks."
- Why is it important: Instract is a model used daily by most people -- you turn on ChatGPT, Kimi, Ernie, which is basically an Instract/ Chat model.
⚠️ Common error zone: A lot of people download the Base model for conversation, and it doesn't work well -- it's actually the Instract version. This is why the same model is so different from the cloud API: quantization precision, chat templates, and the problem of chat templates that I have highlighted among the three ingenious elements of sampling parameters.
References:
6. Ressoning Logic Model: Thinkers who "stop and think"
The Reasoning Model (Reasoning Model) is the hottest category in the last two years and represents Thinking editions of the Open AI series, DeepSeek-R1, Qwen. On the basis of Instract, it uses intensive learning training to "think and answer."
- Core differences: the normal model "Assist the question immediately" and the reasoning model will form a long line of thinking (chain-of-thought), writing out the middle reasoning process and giving the final answer.
- What are they good at: mathematical questions, logical questions, programming, complex reasoning — tasks that need to be “multi-step”.
- What's the price: slow, expensive, token consumes. It's "wanted for half a day" every question, and the cost of reasoning is several times higher than that of Instract.
Use analogy: Instract is like a quick-reactioned intern, but complex questions are easy to answer wrongly; Ressoning is like a "older who can play draft" and gives answers on the draft paper, slow but correct.
• Selective judgement: daily chat, summary, regular question and answer →Instract is sufficient and fast; mathematics, reasoning, difficult programming. Don't use reasoning models for all questions. That's using cannons to fight mosquitoes.
References:
7. By field of application (L0/L1/L2); common, industrial, peripheral
The second main axis, divided by "to whom", is commonly used at the national level on the three levels of L0/L1 / L2 (by industry experts, such as King Kim Bridge of the Middle School):
- L0 Universal LLM (basic Qualitative): a broad-based, "one-on-one" approach, which is equivalent to a "literacy education". For example, the GPT, DeepSeek, and Qwen flagships.
- L1 Industry LLM (industry): On the basis of L0, fine-tuning of industry-specific data (medical, financial, legal), mastering of trade terms and knowledge systems is equivalent to “advanced entry”. For example, a model of dedicated dialogue in the area of finance.
- L2 Peripheral LLM (vertical scenario): Optimizing the depth of the more specific business scenario, which is equivalent to "experts on specific issues". For example, OCR LLM for identification of a document.
By analogy: L0 is a "college graduate of general education", L1 is an "advanced worker in a trade for three years", L2 is a "teacher specializing in certain types of equipment" – more specialized and narrow, but more efficient in solving specific problems.
References:
8. By structure: dense Dense vs Mixed Expert MoE
The third axis, "How to organize inside the model", is divided into two main categories:
- A dense model (Dense): All parameters are activated on each input. The structure is simple, behaviour can be expected, but the parameters are numerous and the reasoning costs are high.
- Mixture of Experts: consisting of multiple "expert sub-networks" each input activates only part of the experts. Under the same total of parameter count, the single reasoning takes only a small number of experts, quickly and saves, and can pile the parameters to trillions of magnitudes.
By analogy, Dense is like "a team with a team on the field" and every player is involved; MoE is like "a team with a clear division of labour" and only a few of the most targeted teams are on the field, while others are on standby.
In this column's "MoE Mixed Expert Model In-depth Analysis" we have detailed MoE's mechanisms, and here we can only remember the conclusion: MoE replaced the "LLM + Low Logic Cost" with "large, one-time calculations ", which is the default option for the current trillion-degree model.
9. By quartile: pure language vs multimodular
The fourth coordinates look at "what input the model can handle":
- Pure language LLM (LLLM, Large Language Model): Processing text only, input is token, output is also token.
- Multi-modular LLM (LMM, Large Multimodal Model): capable of processing text, images, audio, videos, such as "read and speak" "spoken understanding". Common forms are visual language models (VLM), such as GPT-4V, Qwen-VL.
Use analogies: Pure language models are like "editor with word only" and multi-modular models are like "journalist with speech" -- they can produce text, but the latter have more "view" as an input channel.
This is the most independent dimension: multimodules and "general / exclusive" do not conflict with "Base / Insurance" - a model can be "multimodular + General".
10. A summary table: comparison of the three dimensions of the classification
Put the four axes together so it doesn't get messy:
Key perception: The four dimensions are active (independent) and a model has a label for each of the four dimensions. To understand them, you can see the meaning of suffix in model naming -- -- Instract is capability positioning, --VL is multi-modular, -- MoE is architecture, --L2(if labeled) is application level.
Full runable code (Runnable Code): Distinguishing three models with HuggingFace
Here is a copy of the code that you can run, which loads three versions of the same model with HuggingFace, comparing intuitively the difference between Base/Instract/ Resson's response to the same sentence. Before running, pip install transports.
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
# A Qwen family demonstration: three versions of the same model series
MODELS = {
"Base": "Qwen/Qwen2.5-0.5B", # Base model: only continues
"Instruct": "Qwen/Qwen2.5-0.5B-Instract", # Command fine-tune: listens
# Reasoning with DeepSeek-R1-Distill demonstration (small parametric reasoning model)
"Reasoning": "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B",
}
♪ A mathematical question that needs to be answered first.
for kind, name in MODELS.items():
print(f"\n====kind} Model:{name}======”)
Tokenizer = AutoTokenizer. from pretrained(name)# Load corresponding terminator
MOdel = AutoModelForCausalLM.from pretrained(name)# load matching model
# Encoding
= model.generate (**inputs, max new tokens=40, # Generate 40 token extensions Write
# Greed decoded, results confirmed
Point ("Output:", tokenizer.decode (out [0, sskip special tokens=True))
How to read the code: The same prompt 15% of 60 is, the Base version probably will continue to "prescript a reasonable continuation" (the answer is no); the Instract version will give "15% of 60 is 9" by the rules; the Reasoning version will output a "thinking" before giving the answer. The three-line code shows the difference in the essence of the three-phase "capacity positioning" model, which is also the operationalization of the concept described in sections 3 to 6.
12. Comparative table: Symptoms — causes —
13. Closure of judgement
When should this classification be used: when you need to "select the model/ understand the suffix of the model version/ judge why a model is not doing what it says" - you need to look at a definitive version of the Capability Positioning (Base/Instract/Resoning) and then you need to look at the application area (L0/L1/L2) to determine the level of specialization, and finally use the structure and model to supplement it. When you don't have to worry about it: when you're talking to an API that's available, you don't deploy models on your own, they're just "know it" background, and you don't have to look into it. There's only one test -- are you "selecting the model" or "using the model" -- the former needs this classification, and the latter can just pick out a good generic version of Instract.
This column, LLM LLM Technical Column, addresses model selection, fine-tuning, reasoning optimization and Agent landing from Transformer ' s Focusing Mechanism, each with a direct running code, updated weekly. To read in order, you can subscribe to the column page.
References (comprehensive):
Auto-fetched and translated by AI Atlas. All rights belong to the original author.