Large Language Models

•

Introduction to LLMs

History of LLMs

A Survey of Large Language Models

大型語言模型(>10B)的時間軸

計算記憶體的成長與Transformer大小的關係

AI and Memory Wall

Scaling Law

我們可以用模型大小、Dataset大小、總計算量，來預測模型最終能力。（通常以相對簡單的函數型態, ex: Linear relationship）
GPT-4 Technical Report. OpenAI. 2023

Blog: 【LLM 10大觀念-1】Scaling Law

Papers:

  • Hestness et al. 於2017發現在Machine Translation, Language Modeling, Speech Recognition和Image Classification都有出現Scaling law.
  • OpenAI Kaplan et al.2020 於2020年從計算量、Dataset大小、跟參數量分別討論了Scaling Law。
  • Rosenfeld et al. 於2021年發表了關於Scaling Law的survey paper。在各種architecture更進一步驗證Scaling Law的普適性。

Chinchilla Scaling Law

Paper: Training Compute-Optimal Large Language Models

如果我們接受原本Scaling Law的定義（模型性能可藉由參數量、Dataset大小、計算量預測），馬上就會衍伸出兩個很重要的問題:

Return（收益）： 在固定的訓練計算量之下，我們所能得到的最好性能是多好？
Allocation（分配）：我們要怎麼分配我們的模型參數量跟Dataset大小。
（假設計算量 = 參數量 * Dataset size，我們要大模型 * 少量data、中模型 * 中量data、還是小模型 * 大量data）

2022年DeepMind提出Chinchilla Scaling Law，同時解決了這兩個問題，並且依此改善了當時其他大模型的訓練方式。 他們基於三種方式來找到訓練LLM的Scaling Law：

  1. 固定模型大小，變化訓練Data數量。
  2. 固定計算量（浮點運算），變化模型大小。
  3. 對所有實驗結果，直接擬合參數化損失函數。

Method 3 result from Chinchilla Scaling Law，N是模型參數量、D是數據量、其他都是係數

LLM最終的Loss（Perplexity），會隨著模型放大、數據量變多而下降，並且是跟他們呈現指數映射後線性關係。

Chinchilla最大的貢獻更是在解決Allocation的問題，他們發現

  • 數據量（Tokens數）應該要約等於模型參數量的20倍
  • 並且數據量跟模型參數量要同比放大（Ex: 模型放大一倍，數據也要跟著增加一倍）

Large Language Models

Open LLM Leaderboard

Transformer

Paper: Attention Is All You Need

ChatGPT

ChatGPT: Optimizing Language Models for Dialogue
ChatGPT is fine-tuned from a model in the GPT-3.5 series, which finished training in early 2022.

GPT4

Paper: GPT-4 Technical Report
Paper: From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting
Blog: GPT-4 Code Interpreter: The Next Big Thing in AI

LLaMA

Paper: LLaMA: Open and Efficient Foundation Language Models
Blog: Building a Million-Parameter LLM from Scratch Using Python
Kaggle: LLM LLaMA from scratch

BloombergGPT

Paper: BloombergGPT: A Large Language Model for Finance
Blog: Introducing BloombergGPT, Bloomberg’s 50-billion parameter large language model, purpose-built from scratch for finance

Pythia

Paper: Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling
Dataset:
The Pile: An 800GB Dataset of Diverse Text for Language Modeling
Datasheet for the Pile
Code: Pythia: Interpreting Transformers Across Time and Scale

MPT-7B

model: mosaicml/mpt-7b-chat
Code: https://github.com/mosaicml/llm-foundry
Blog: Announcing MPT-7B-8K: 8K Context Length for Document Understanding
Blog: Introducing MPT-7B: A New Standard for Open-Source, Commercially Usable LLMs

Falcon-40B

model: tiiuae/falcon-40b
Paper: The RefinedWeb Dataset for Falcon LLM: Outperforming Curated Corpora with Web Data, and Web Data Only

Orca

Paper: Orca: Progressive Learning from Complex Explanation Traces of GPT-4

OpenLLaMA

model: openlm-research/open_llama_3b_v2
model: openlm-research/open_llama_7b_v2
Code: https://github.com/openlm-research/open_llama
Kaggle: https://www.kaggle.com/code/rkuo2000/llm-openllama

Vicuna

model: lmsys/vicuna-7b-v1.5
Paper: Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
Code: https://github.com/lm-sys/FastChat

LLaMA-2

model: meta-llama/Llama-2-7b-chat-hf
Paper: Llama 2: Open Foundation and Fine-Tuned Chat Models
Code: https://github.com/facebookresearch/llama

Sheared LLaMA

model_name = “princeton-nlp/Sheared-LLaMA-1.3B”, princeton-nlp/Sheared-LLaMA-2.7B | princeton-nlp/Sheared-Pythia-160m
Paper: Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning
Code: https://github.com/princeton-nlp/LLM-Shearing

Neural-Chat-7B (Intel)

model_name = “Intel/neural-chat-7b-v3-1”
Blog: Intel neural-chat-7b Model Achieves Top Ranking on LLM Leaderboard!

Mistral

model_name = “mistralai/Mistral-7B-Instruct-v0.2”
Paper: Mistral 7B
Code: https://github.com/mistralai/mistral-src
Kaggle: https://www.kaggle.com/code/rkuo2000/llm-mistral-7b-instruct

Mistral 8X7B

model: mistralai/Mixtral-8x7B-v0.1
Paper: Mixtral of Experts

Starling-LM

model: Nexusflow/Starling-LM-7B-beta
Paper: RLAIF: Scaling Reinforcement Learning from Human Feedback with AI Feedback
Blog: Starling-7B: Increasing LLM Helpfulness & Harmlessness with RLAIF

Zephyr

model: HuggingFaceH4/zephyr-7b-beta
Paper: Zephyr: Direct Distillation of LM Alignment
Kaggle: https://www.kaggle.com/code/rkuo2000/llm-zephyr-7b
Blog: Zephyr-7B : HuggingFace’s Hyper-Optimized LLM Built on Top of Mistral 7B

Orca 2

model: microsoft/Orca-2-7b
Paper: https://arxiv.org/abs/2311.11045
Blog: Microsoft’s Orca 2 LLM Outperforms Models That Are 10x Larger

BlueLM (VIVO)

model: vivo-ai/BlueLM-7B-Chat-4bits
Code: https://github.com/vivo-ai-lab/BlueLM/

Taiwan-LLM (優必達+台大)

model: yentinglin/Taiwan-LLM-7B-v2.1-chat
Paper: TAIWAN-LLM: Bridging the Linguistic Divide with a Culturally Aligned Language Model
Blog: 專屬台灣！優必達攜手台大打造「Taiwan LLM」，為何我們需要本土化的AI？
Code: https://github.com/MiuLab/Taiwan-LLM

Phi-2 (Transformer with 2.7B parameters)

model: microsoft/phi-2
Blog: Phi-2: The surprising power of small language models
Kaggle: https://www.kaggle.com/code/rkuo2000/llm-phi-2

Mamba

model: Q-bert/Mamba-130M
Paper: Mamba: Linear-Time Sequence Modeling with Selective State Spaces
Kaggle: https://www.kaggle.com/code/rkuo2000/llm-mamba-130m
Kaggle: https://www.kaggle.com/code/rkuo2000/llm-mamba-3b

SOLAR-10.7B ~ Depth Upscaling

Paper: SOLAR 10.7B: Scaling Large Language Models with Simple yet Effective Depth Up-Scaling
Code: https://huggingface.co/upstage/SOLAR-10.7B-v1.0
Depth-Upscaled SOLAR-10.7B has remarkable performance. It outperforms models with up to 30B parameters, even surpassing the recent Mixtral 8X7B model.
Leveraging state-of-the-art instruction fine-tuning methods, including supervised fine-tuning (SFT) and direct preference optimization (DPO), researchers utilized a diverse set of datasets for training. This fine-tuned model, SOLAR-10.7B-Instruct-v1.0, achieves a remarkable Model H6 score of 74.20, boasting its effectiveness in single-turn dialogue scenarios.

Qwen (通义千问)

model: Qwen/Qwen1.5-7B-Chat
Blog: Introducing Qwen1.5
Code: https://github.com/QwenLM/Qwen1.5
Kaggle: https://www.kaggle.com/code/rkuo2000/llm-qwen1-5

Yi (零一万物)

model: 01-ai/Yi-6B-Chat
Paper: CMMMU: A Chinese Massive Multi-discipline Multimodal Understanding Benchmark
Paper: Yi: Open Foundation Models by 01.AI

Orca-Math

Paper: Orca-Math: Unlocking the potential of SLMs in Grade School Math
Dataset: https://huggingface.co/datasets/microsoft/orca-math-word-problems-200k

BitNet

Paper: BitNet: Scaling 1-bit Transformers for Large Language Models
Paper: The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits
bitnet.cpp is the official inference framework for 1-bit LLMs (e.g., BitNet b1.58).

Gemma

model: google/gemma-1.1-7b-it
Blog: Gemma: Introducing new state-of-the-art open models
Kaggle: https://www.kaggle.com/code/nilaychauhan/fine-tune-gemma-models-in-keras-using-lora

Gemini-1.5

Claude 3

Breeze (達哥)

model: MediaTek-Research/Breeze-7B-Instruct-v0_1
Paper: Breeze-7B Technical Report
Blog: Breeze-7B: 透過 Mistral-7B Fine-Tune 出來的繁中開源模型

Bialong (白龍)

Paper: Bailong: Bilingual Transfer Learning based on QLoRA and Zip-tie Embedding
model: INX-TEXT/Bailong-instruct-7B

TAIDE

model: taide/TAIDE-LX-7B-Chat

  • TAIDE-LX-7B: 以 LLaMA2-7b 為基礎，僅使用繁體中文資料預訓練 (continuous pretraining)的模型，適合使用者會對模型進一步微調(fine tune)的使用情境。因預訓練模型沒有經過微調和偏好對齊，可能會產生惡意或不安全的輸出，使用時請小心。
  • TAIDE-LX-7B-Chat: 以 TAIDE-LX-7B 為基礎，透過指令微調(instruction tuning)強化辦公室常用任務和多輪問答對話能力，適合聊天對話或任務協助的使用情境。TAIDE-LX-7B-Chat另外有提供4 bit 量化模型，量化模型主要是提供使用者的便利性，可能會影響效能與更多不可預期的問題，還請使用者理解與注意。

InflectionAI

Blog: Inflection AI 發表新基礎模型「Inflection-2.5 」，能力逼近 GPT-4！

Phind-70B

Blog: Introducing Phind-70B – closing the code quality gap with GPT-4 Turbo while running 4x faster
Blog: Phind - 給工程師的智慧搜尋引擎
Phind-70B is significantly faster than GPT-4 Turbo, running at 80+ tokens per second to GPT-4 Turbo’s ~20 tokens per second. We’re able to achieve this by running NVIDIA’s TensorRT-LLM library on H100 GPUs, and we’re working on optimizations to further increase Phind-70B’s inference speed.

Llama-3

model: meta-llama/Meta-Llama-3-8B-Instruct
Code: https://github.com/meta-llama/llama3/

Phi-3

model: microsoft/Phi-3-mini-4k-instruct”
Blog: Introducing Phi-3: Redefining what’s possible with SLMs

Octopus v4

model: NexaAIDev/Octopus-v4
Paper: Octopus v4: Graph of language models
Code: https://github.com/NexaAI/octopus-v4
design demo

Llama 3.1

mode: meta-llama/Meta-Llama-3.1-8B-Instruct

Grok-2

Grok-2 & Grok-2 mini, achieve performance levels competitive to other frontier models in areas such as graduate-level science knowledge (GPQA), general knowledge (MMLU, MMLU-Pro), and math competition problems (MATH). Additionally, Grok-2 excels in vision-based tasks, delivering state-of-the-art performance in visual math reasoning (MathVista) and in document-based question answering (DocVQA).

Phi-3.5

model: microsoft/Phi-3.5-mini-instruct
model: microsoft/Phi-3.5-vision-instruct
model: microsoft/Phi-3.5-MoE-instruct
News: Microsoft Unveils Phi-3.5: Powerful AI Models Punch Above Their Weight

OpenAI o1

Blog: Introducing OpenAI o1-preview

Qwen2.5

model: Qwen/Qwen2.5-7B-Instruct

  • Qwen2.5: 0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B
  • Qwen2.5-Coder: 1.5B, 7B, coming 32B
  • Qwen2.5-Math: 1.5B, 7B, 72B

Blog: 阿里雲AI算力大升級！發佈100個開源Qwen 2.5模型及視頻AI模型

NVLM 1.0

Paper: NVLM: Open Frontier-Class Multimodal LLMs

Llama 3.2

Blog: Llama 3.2: Revolutionizing edge AI and vision with open, customizable models
model: meta-llama/Llama-3.2-1B-Instruct
model: meta-llama/Llama-3.2-3B-Instruct
model: meta-llama/Llama-3.2-11B-Vision-Instruct

LFM Liquid-3B

Try Liquid

safe AI

Constitutional AI

Paper: Constitutional AI: Harmlessness from AI Feedback Two key phases:

  1. Supervised Learning Phase (SL Phase)
    • Step1: The learning starts using the samples from the initial model
    • Step2: From these samples, the model generates self-critiques and revisions
    • Step3: Fine-tine the original model with these revisions
  2. Reinforcement Learning Phase (RL Phease)
    • Step1. The model uses samples from the fine-tuned model.
    • Step2. Use a model to compare the outputs from samples from the initial model and the fine-tuned model
    • Step3. Decide which sample is better. (RLHF)
    • Step4. Train a new “preference model” from the new dataset of AI preferences. This new “prefernece model” will then be used to re-train the RL (as a reward signal). It is now the RLHAF (Reinforcement Learning from AI feedback)

Attack LLM

Blog 如何攻擊 LLM (ChatGPT) ?

  • JailBreak
  • Prompt Injection
  • Data poisoning

LLM running locally

ollama

ollama -v
ollama
ollama pull llama3.2
ollama run llama3.2

Code: Github
Examples:

LM Studio

PrivateGPT

Code: https://github.com/zylon-ai/private-gpt

GPT4All

chmod +x gpt4all-installer-linux.run
./gpt4all-installer-linux.run
cd ~/gpt4all
./bin/chat

GPT4FREE

pip install g4f

Reasoning

我在杯子裡放一個草莓, 把杯子倒過來放在桌上, 再把杯子放到微波爐裡, 請問草莓在哪？
  • Gemini
  • ChatGPT
  • Llama3.2

Chain-of-Thought Prompting

Paper: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

ReAct Prompting

Paper: ReAct: Synergizing Reasoning and Acting in Language Models
Code: https://github.com/ysymyth/ReAct

Tree-of-Thoughts

Paper: Tree of Thoughts: Deliberate Problem Solving with Large Language Models
Code: https://github.com/princeton-nlp/tree-of-thought-llm
Code: https://github.com/kyegomez/tree-of-thoughts

Tabular CoT

Paper: Tab-CoT: Zero-shot Tabular Chain of Thought
Code: https://github.com/Xalp/Tab-CoT

Survey of Chain-of-Thought

Paper: A Survey of Chain of Thought Reasoning: Advances, Frontiers and Future

Chain-of-Thought Hub

Paper: Chain-of-Thought Hub: A Continuous Effort to Measure Large Language Models’ Reasoning Performance
Code: https://github.com/FranxYao/chain-of-thought-hub

Everything-of-Thoughts

Paper: Everything of Thoughts: Defying the Law of Penrose Triangle for Thought Generation
Code: https://github.com/microsoft/Everything-of-Thoughts-XoT

R3

Paper: Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning
Code: https://github.com/WooooDyy/LLM-Reverse-Curriculum-RL

Time-series LLM

Paper: Large Language Models for Time Series: A Survey

TimeGPT-1

Paper: TimeGPT-1

Lag-LLaMA

Paper: Lag-Llama: Towards Foundation Models for Probabilistic Time Series Forecasting
Blog: From RNN/LSTM to Temporal Fusion Transformers and Lag-Llama
Code: https://github.com/time-series-foundation-models/lag-llama
Colab:

Applications

FunSearch

DeepMind發展用LLM解困難數學問題的方法

Automatic Evaluation

Paper: Can Large Language Models Be an Alternative to Human Evaluation?

Paper: A Closer Look into Automatic Evaluation Using Large Language Models
Code: https://github.com/d223302/A-Closer-Look-To-LLM-Evaluation

BrainGPT

Paper: DeWave: Discrete EEG Waves Encoding for Brain Dynamics to Text Translation
Blog: New Mind-Reading “BrainGPT” Turns Thoughts Into Text On Screen

Designing Silicon Brains using LLM

Paper: Designing Silicon Brains using LLM: Leveraging ChatGPT for Automated Description of a Spiking Neuron Array

Robotic Manipulation

Paper: Language-conditioned Learning for Robotic Manipulation: A Survey
Paper: Human Demonstrations are Generalizable Knowledge for Robots

ALTER-LLM

Paper: From Text to Motion: Grounding GPT-4 in a Humanoid Robot “Alter3”



This site was last updated November 15, 2024.