`

AI domain related

 
阅读更多
Instruct learning/prompt learning
Reinforcement Learning from Human Feedback

Prompt Engineering from Microsoft
https://learn.microsoft.com/en-us/semantic-kernel/prompt-engineering/

https://zhuanlan.zhihu.com/p/605451419
实现对话模型的第一步就需要依赖SFT(Supervised FineTune),技术上是对GPT-3进行有监督的微调,本质目的是「对数据集进行优化」,用人类喜欢的答案作为训练数据,机器才能生成人类更喜欢的内容。

Supervised FineTune: 靠人工标注
RM:靠人工打分
Proximal Policy Optimization: SFT +RM

How to Train Really Large Models on Many GPUs
https://lilianweng.github.io/posts/2021-09-25-train-large/#:~:text=How%20to%20Train%20Really%20Large%20Models%20on%20Many,to%20push%20the%20limit%20of%20model%20size.%20

Training Parallelism
Data Parallelism
Model Parallelism
Pipeline Parallelism
Tensor Parallelism:partition the computation for one tensor operation across multiple devices

State of GPT:大神Andrej揭秘OpenAI大模型原理和训练过程
https://mp.weixin.qq.com/s/zmEGzm1cdXupNoqZ65h7yg

超大语言模型LLM发展历程
https://zhuanlan.zhihu.com/p/612181615

GPT-2: 由transformer解码器块构建,并在非常大的文本语料库上进行训练,以预测段落中的下一个单词
T5: 用于回答问题、总结、翻译和分类

语料: 顾名思义就是我们平时所说的文本,带有文字描述性的文本都可以归类于语料
语料预处理方法: 主要包括数据清洗、分词、词性标注(每个词标注正确的词性,即确定每个词是名词、动词、形容词或其他词性的过程)、去停用词等。

数据标注
https://www.v7labs.com/blog/data-labeling-guide#what-is-data-labeling

https://docs.v7labs.com/docs

Conda Commands:
conda info -e: list all envs

conda create -n johnson --clone glm6b_ljz
conda remove -n johnson --all

conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/
conda config --set show_channel_urls yes
conda config --show channels
conda config --remove channels https://mirrors.tuna.tsinghua.edu.cn

示例
channels:
  - https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/
  - https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/
  - defaults
show_channel_urls: true
proxy_servers:
  https: http://proxy.huawei.com:8080

virtualenv:
1. pip install virtualenv(if not installed)
2. mkdir -p {virtualPath}
3. cd {virtualPath}
4. virtualenv {virtualName}
5. {virtualPath}/{virtualName}/Scripts/activate.bat
   or {virtualPath}/{virtualName}/Scripts/deactivate.bat


Types of evaluations:
1. Metrics
A metric measures the performance of a model on a given dataset.
Examples of metrics include: Accuracy, Exact Match, Mean Intersection over union (IoUO)
2. Comparisons
3. Measurements

LangChain:
https://mp.weixin.qq.com/s/KrWM3cMywMvYUiawRZ94Gg

Build the prompt with RAG
https://learn.deeplearning.ai/building-applications-vector-databases/lesson/3/retrieval-augmented-generation-(rag)

Building Agentic RAG with LlamaIndex
https://www.deeplearning.ai/short-courses/building-agentic-rag-with-llamaindex/

building your own database agent
https://www.deeplearning.ai/short-courses/building-your-own-database-agent/


补充参考:
Pip package installing ways:
https://packaging.python.org/en/latest/tutorials/installing-packages/#installing-from-local-archives

CUDA与Pytorch适配版本列表:
https://blog.csdn.net/weixin_44842318/article/details/127492491
快速了解 OpenAI 的 fine-tune 和 Embedding 能力
https://zhuanlan.zhihu.com/p/609359047

python:
https://www.runoob.com/python3/python3-tutorial.html


分享到:
评论

相关推荐

Global site tag (gtag.js) - Google Analytics