개발자를 위한_Large Language Model(LLM) _ 한성민 [Hello World]

Transcript

1. Large Language Model(LLM) SungMin Han GDE for ML | GDG Golang Korea
2. Speaker (SungMin Han) - MLOps Lead at Riiid - GDE for ML - GDG Golang Korea Organizer - Python Mentor at F-Lab - Former) Research Engineer at Naver Clova - Former) Software Engineer at 심심이
3. Agenda • Generative AI • Large Language Model(LLM) • ChatGPT (and GPT 4) • Instruct GPT • GPT in Go • Appendix
4. Generative AI
5. Keywords GPT 4 GPT 3.5 ChatGPT LLaMA Stanford Alpaca T5 Google Bard Megatron-TLG Stable Diffusion MidJourney BingChat DALL·E 2 Microsoft 365 Copilot LaMDA BLOOM
6. Keywords in 2023 GPT 4 GPT 3.5 ChatGPT LLaMA Stanford Alpaca T5 Google Bard Megatron-TLG Stable Diffusion MidJourney BingChat DALL·E 2 Microsoft 365 Copilot LaMDA BLOOM
7. Image Midjourney, NovelAI, Stable Diffusion…
8. Audio Riffusion MurfAI
9. Text Google Bard OpenAI ChatGPT
10. Landscapes https://www.antler.co/blog/generative-ai
11. Large Language Model (LLM)
12. (2020) https://www.microsoft.com/en-us/research/blog/turing-nlg-a-17-billion-parameter-language-model-by-microsoft/
13. 2022 https://huggingface.co/blog/large-language-models
14. ChatGPT
15. Reinforcement Learning with Human Feedback (RLHF) https://github.com/Hannibal046/Awesome-LLM/blob/main/resources/creepy_llm.jpeg
16. Intruct GPT (GPT 3.5) ChatGPT (GPT 3.5 turbo) https://matt-rickard.com/gpt-lineage https://lifearchitect.ai/
17. Instruct GPT: 3 Step Training Method https://openai.com/blog/chatgpt *PPO => Proximal Policy Optimization Algorithm in the field of Reinforcement Learning
20. CTO
22. Manim
23. Mermaid
24. Do Anything Now (DAN): ChatGPT
25. Do Anything Now (DAN): Jailbreak
26. Prompt Base Awesome ChatGPT Prompt
27. GPT4: Multi-modal (Image)
28. LLM 복잡한 문제
29. 한계: 할루시네이션1
31. : regurgitation engine
32. Instruct GPT
34. LeetCode
37. GPT in Go
39. package main import ( "fmt" gpt35 "github.com/AlmazDelDiablo/gpt3-5-turbo-go" ) func main() { api_key := "<REDACTED>" instruct := ` 내가 앞으로 음식 재료들을 입력하면, 너가 입력된 재료들을 조합한 음식 키워드를 하나 뽑아줘. 줄 넘김은 하면 안돼 ! 예시 1: INPUT: 시금치, 고추장, 계란, 밥, 콩나물, 고사리 OUTPUT: 비빔밥 예시 2: INPUT: 면, 라면스프, 물 OUTPUT: 라면 시작해볼까?` c := gpt35.NewClient(api_key) fmt.Println("음식 재료를 입력하세요 (쉼표로 구분):") var ingredients string fmt.Scanln(&ingredients)
40. req := &gpt35.Request{ Model: gpt35.ModelGpt35Turbo, Messages: []*gpt35.Message{ { Role: gpt35.RoleUser, Content: instruct, }, { Role: gpt35.RoleSystem, Content: "네, 시작해봅시다! 어떤 재료를 사용하실 건가요?", }, { Role: gpt35.RoleUser, Content: ingredients, }, }, } resp, err := c.GetChat(req) if err != nil { panic(err) } fmt.Println("추천 음식:", resp.Choices[0].Message.Content) }
41. $ go run main.go 음식 재료를 입력하세요 (쉼표로 구분): 시금치,고추장, 나물 추천 음식: 추천하는 음식 키워드는 된장찌개입니다! $ go run main.go 음식 재료를 입력하세요 (쉼표로 구분): 옥수수, 고춧가루, 소금 추천 음식: 옥수수, 고춧가루, 소금으로 만드는 음식은 '옥수수전'입니다. 즐거운 조리되 세요!
42. Appendix
43. Appendix: Stanford Alpaca LLaMA GPT 3.5(text-davinci-003) 7B LLM.