Update on building LLMs tools for desci onboarding
#da0-desci
Open-Assistant
https://github.com/LAION-AI/Open-Assistant
Stability-AI/ StableLM
https://github.com/Stability-AI/StableLM
https://huggingface.co/blog/rlhf
Open Source Tools for RLHF
• Transformers Reinforcement Learning
https://github.com/lvwerra/trl
• Reinforcement Learning for Language models (RL4LMs)
https://github.com/CarperAI/trlx
• CarperAI trlx
https://github.com/CarperAI/trlx
◦ Transformer Reinforcement Learning X
https://github.com/CarperAI/trlx
◦
模型開發的工具
• 基礎模型
◦ OpenAI, Google...
• 垂直領域模型
◦ Einstein, Baize, FireFly, PERT, Lamini, huggingface
• 多模態/擴散模型
◦ Stability, LoRa, ControlNet
• 輕模型
◦ 知識蒸餾
◦ 量化優化:FP32 -> Int4
◦ 結構優化(避免Padding)
◦ 內存優化
• 開源模型/數據集
◦ 編碼器
▪︎ BERT, ALBERT, ROBERTA, DEBERTA
◦ 解碼器
▪︎ LLaMa, GPT早期, BLLOM, FLAN
◦ 編碼器-解碼器
▪︎ T5, T0, BART, FLAN-T5
◦ 語言模型
▪︎ LLaMa, Alpaca, Vicuna, Koala, LLaMa Adaptor, CerebaseGTP, MPT-7B(MosiacAI), Dolly(Databrick)
◦ 重要數據集/系統
▪︎ Wikipedia, Common Crawl, MSCOCO, VQA
應用開發
• 工具箱&工具鍊
◦ 基礎開發能力
▪︎ 代碼框架:PyTorch, TensorFlow, MXNet
▪︎ 基礎服務:AWS, Azure, Google Colab
▪︎ 工作流:LangChain, Cohere, Helicone, Stack AI
▪︎ 代理:Auto-GPT
▪︎ 編制:BabyAGI
▪︎ 整合:Jarvis
▪︎ 提示:Github Copilot, Tabnine
▪︎ 調適:Syn Code
▪︎ 記憶:Pinecore, Zilliz