Thử nghiệm LLM “Llama 3” mới nhất của Meta trong môi trường siêu tốc của Groq

Xin chào, tôi là Kakeya, đại diện của Scuti.

Công ty chúng tôi chuyên về phát triển offshore và phát triển dựa trên phòng thí nghiệm tại Việt Nam với trọng tâm là AI tạo sinh, cũng như cung cấp các dịch vụ tư vấn AI tạo sinh. Gần đây, chúng tôi may mắn nhận được nhiều yêu cầu phát triển hệ thống kết hợp với AI tạo sinh.

Meta đã gây chú ý gần đây khi công bố LLM mới nhất của mình, “Llama 3”, dưới dạng mã nguồn mở. Tôi đã thử nghiệm Llama 3 thông qua một dịch vụ gọi là Groq, cho phép vận hành siêu tốc suy luận LLM. Tôi rất vui được chia sẻ quá trình bao gồm cả các hoạt động thực tế. Llama 3 rất ấn tượng, nhưng điều làm tôi thực sự kinh ngạc là khả năng phản hồi quá nhanh của Groq…!

Mục lục

  • Giới thiệu Groq
  • Giới thiệu Llama 3
  • Thử nghiệm Llama 3 với Groq
  • Thử nghiệm phiên bản trình duyệt của Chatbot UI
  • Thử nghiệm phiên bản cục bộ của Chatbot UI

Giới thiệu Groq

Groq là gì?

Groq là công ty phát triển phần cứng tùy chỉnh và Đơn vị Xử lý Ngôn ngữ (LPU) được thiết kế đặc biệt để tăng tốc suy luận của các mô hình ngôn ngữ lớn (LLM). Công nghệ này được đặc trưng bởi tốc độ xử lý nhanh hơn đáng kể so với phần cứng thông thường.

LPU của Groq đạt hiệu suất suy luận nhanh hơn đến 18 lần so với các nhà cung cấp dịch vụ dựa trên đám mây khác. Groq hướng đến việc tối đa hóa hiệu suất của các ứng dụng AI thời gian thực bằng cách sử dụng công nghệ này.

Ngoài ra, gần đây Groq đã thành lập một bộ phận mới có tên là “Groq Systems,” hỗ trợ triển khai chip vào các trung tâm dữ liệu và xây dựng các trung tâm dữ liệu mới. Hơn nữa, Groq đã mua lại Definitive Intelligence, một nhà cung cấp giải pháp AI cho doanh nghiệp, để tăng cường thêm sức mạnh công nghệ và ảnh hưởng thị trường của mình.

LPU của Groq là gì?

LPU (Đơn vị Xử lý Ngôn ngữ) là một bộ xử lý chuyên dụng được thiết kế để cải thiện hiệu suất và độ chính xác suy luận của các mô hình ngôn ngữ AI. Khác với GPU truyền thống, LPU được chuyên môn hóa cho các nhiệm vụ tính toán tuần tự, loại bỏ các điểm nghẽn về tính toán và băng thông bộ nhớ cần thiết cho các mô hình ngôn ngữ lớn. Ví dụ, LPU giảm thời gian tính toán cho mỗi từ so với GPU và không gặp phải điểm nghẽn băng thông bộ nhớ bên ngoài, do đó cung cấp hiệu suất đáng kể tốt hơn so với bộ xử lý đồ họa.

Về mặt kỹ thuật, LPU chứa hàng ngàn yếu tố xử lý đơn giản (PE), được sắp xếp theo mảng Single Instruction, Multiple Data (SIMD). Điều này cho phép cùng một lệnh được thực hiện đồng thời cho mỗi điểm dữ liệu. Ngoài ra, một đơn vị điều khiển trung tâm (CU) phát lệnh và quản lý dòng chảy dữ liệu giữa phân cấp bộ nhớ và PE, duy trì giao tiếp đồng bộ nhất quán.

Hiệu quả chính của việc sử dụng LPU là tăng tốc các nhiệm vụ AI và học máy. Ví dụ, Groq đã chạy mô hình Llama-2 70B với tốc độ 300 token mỗi giây cho mỗi người dùng trên hệ thống LPU, đây là một cải tiến đáng kể so với 100 token và 240 token trước đây. Như vậy, LPU có thể xử lý các nhiệm vụ suy luận AI theo thời gian thực với độ trễ thấp và cung cấp chúng trong gói hiệu quả năng lượng. Điều này cho phép các thay đổi đột phá trong các lĩnh vực như tính toán hiệu năng cao (HPC) và tính toán cạnh.

API Giá rẻ bất ngờ của Groq

Tài nguyên phần cứng của Groq có thể được truy cập thông qua API do Groq cung cấp. Một trong những đặc điểm nổi bật của dịch vụ Groq là chi phí sử dụng API này cực kỳ thấp

Nói cách khác, Groq vừa nhanh vừa rẻ

Biểu đồ dưới đây, do ArtificialAnalysis.ai cung cấp, so sánh các nhà cung cấp API, với giá cả trên trục ngang và thông lượng (token mỗi giây) trên trục dọc. Từ biểu đồ này, có thể thấy rõ ràng rằng Groq nhanh và rẻ đáng kể, đặc biệt nổi bật về thông lượng so với các nhà cung cấp API khác.

Giới thiệu Llama 3

Tổng quan và các đặc điểm của Llama 3

Llama 3 là mô hình ngôn ngữ lớn (LLM) mới nhất do Meta phát triển. AI này được đào tạo trên một tập dữ liệu văn bản khổng lồ, cho phép nó hiểu và phản hồi ngôn ngữ một cách toàn diện. Llama 3 phù hợp với nhiều nhiệm vụ khác nhau bao gồm tạo nội dung sáng tạo, dịch thuật ngôn ngữ và cung cấp thông tin cho các truy vấn. Mô hình này có sẵn trên các nền tảng như AWS, Databricks và Google Cloud, phục vụ làm cơ sở để các nhà phát triển và nhà nghiên cứu tiếp tục thúc đẩy AI. Bằng cách cung cấp Llama 3 dưới dạng mã nguồn mở, Meta nâng cao tính minh bạch của công nghệ và thúc đẩy sự hợp tác với nhiều nhà phát triển.

Dưới đây là một số đặc điểm của Llama 3 do Meta công bố:

  • Mô hình tham số cao cấp: Llama 3 được phát triển như một mô hình có số lượng tham số khổng lồ, có hai phiên bản: 8B (tám tỷ) và 70B (bảy mươi tỷ), cho phép nó thể hiện hiệu suất cao trong việc hiểu và tạo ngôn ngữ.
  • Đánh giá theo tiêu chuẩn ngành: Llama 3 được đánh giá bằng nhiều tiêu chuẩn ngành như ARC, DROP và MMLU, đạt kết quả hàng đầu trong các bài kiểm tra này. Cụ thể, nó cho thấy sự cải thiện đáng kể về độ chính xác nhận dạng và suy luận so với các mô hình trước.
  • Hỗ trợ đa ngôn ngữ: Llama 3 được đào tạo với tập dữ liệu hơn 15 nghìn tỷ token bao gồm dữ liệu chất lượng cao để hỗ trợ hơn 30 ngôn ngữ, không chỉ tiếng Anh mà còn nhiều ngôn ngữ khác, phục vụ người dùng toàn cầu.
  • Cải tiến Tokenizer: Tokenizer mới được phát triển có 128,000 từ vựng, cho phép mã hóa ngôn ngữ hiệu quả hơn. Điều này nâng cao tốc độ và độ chính xác của mô hình, cho phép nó xử lý các văn bản phức tạp một cách thích hợp.
  • Cải thiện hiệu quả suy luận: Việc áp dụng công nghệ Grouped Query Attention (GQA) đã nâng cao hiệu quả suy luận của mô hình, cho phép xử lý nhanh các tập dữ liệu lớn và tạo điều kiện cho phản hồi theo thời gian thực.
  • Tăng cường độ bảo mật và đáng tin cậy: Các công cụ bảo mật mới như Llama Guard 2, Code Shield và CyberSec Eval 2 đã được giới thiệu để nâng cao an toàn của nội dung do mô hình tạo ra, giảm thiểu rủi ro sản xuất nội dung không phù hợp.
  • Khả năng truy cập mã nguồn mở: Llama 3 được cung cấp hoàn toàn dưới dạng mã nguồn mở, cho phép các nhà nghiên cứu và nhà phát triển trên toàn thế giới truy cập tự do, cải thiện và sử dụng nó để phát triển ứng dụng. Mã nguồn mở giúp thúc đẩy sự đổi mới thêm.

Đánh giá hiệu suất của Llama 3 theo các tiêu chuẩn ngành

Llama 3 đã thể hiện hiệu suất nổi bật trong các tiêu chuẩn ngành, đặc biệt là xuất sắc trong các lĩnh vực hiểu ngôn ngữ, suy luận logic, và khả năng giải quyết vấn đề. Các mô hình mới được giới thiệu với 8 tỷ (8B) và 70 tỷ (70B) tham số cho thấy sự cải thiện đáng kể về độ chính xác trong hiểu và phản hồi so với mô hình trước đó, Llama 2.

Llama 3 được đánh giá cao về khả năng tạo mã và thực hiện các nhiệm vụ dựa trên chỉ thị, và những khả năng này được củng cố thêm bởi phương pháp đào tạo độc quyền của Meta. Ngoài ra, mô hình này hỗ trợ đa ngôn ngữ, cung cấp hiệu suất chất lượng cao trong hơn 30 ngôn ngữ.

Hiện tại, Meta đã công bố hai phiên bản của mô hình Llama 3: Llama 3 8B và Llama 3 70B, trong đó “8B” có 8 tỷ tham số và “70B” có 70 tỷ tham số. Hình dưới đây cho thấy kết quả của các đánh giá tiêu chuẩn ngành do Meta công bố, với mô hình 70B thể hiện hiệu suất cao hơn Gemini Pro và Claude3.

Nguồn: https://llama.meta.com/llama3/ 

So sánh giữa Llama 3 và Llama 2

Llama 3 đã có những cải tiến đáng kể so với Llama 2 ở nhiều khía cạnh. Cụ thể, bộ dữ liệu được sử dụng để đào tạo mô hình đã tăng gấp bảy lần, và dữ liệu liên quan đến mã đã tăng gấp bốn lần.

Kết quả là, mô hình đã trở nên hiệu quả hơn trong việc xử lý các nhiệm vụ ngôn ngữ phức tạp. Đặc biệt, Llama 3 vượt trội hơn mô hình trước đáng kể trong khả năng tạo mã và thực hiện các nhiệm vụ dựa trên chỉ thị.

Việc áp dụng tokenizer mới cũng đã cải thiện hiệu quả mã hóa ngôn ngữ, góp phần nâng cao hiệu suất tổng thể. Ngoài ra, một thiết kế tính đến yếu tố an toàn đã được triển khai, giới thiệu một hệ thống lọc mới để giảm thiểu rủi ro tạo ra các phản hồi không phù hợp.

Đặc điểmLlama 3Llama 2
Số lượng tham sốCác phiên bản với 8 tỷ và 70 tỷ tham sốÍt tham số hơn (số liệu chính xác không được công bố)
Hiệu suất trong BenchmarkHiệu suất cải thiện trong các benchmark ARC, DROP, MMLUCó thể có điểm thấp hơn trong cùng các benchmark
Khả năngKhả năng suy luận mạnh mẽ, tạo văn bản sáng tạoCó thể cung cấp các khả năng tương tự nhưng ở mức độ thấp hơn
Khả năng tiếp cậnCông khai cho mục đích nghiên cứuChi tiết về khả năng tiếp cận không rõ ràng, có lẽ ít được công khai hơn

Thử nghiệm Llama 3 với Groq

Hiện tại, “Meta AI” chưa được công bố tại Nhật Bản.

Người ta có thể trải nghiệm tạo hình ảnh nhanh và trò chuyện sử dụng Llama 3 thông qua trợ lý AI “Meta AI” do Meta phát hành. Tuy nhiên, tính đến ngày 1 tháng 5 năm 2024, Meta AI chỉ được công bố ở các khu vực nói tiếng Anh và chưa được phát hành tại Nhật Bản.

Thử Nghiệm Llama 3 với Groq

Tuy nhiên, như đề cập ở trên, vì Groq đã hỗ trợ Llama 3 nên bạn có thể trải nghiệm Llama 3 bằng cách đăng nhập vào Groq. Hơn nữa, sử dụng cực kỳ nhanh và miễn phí!

Cách sử dụng rất đơn giản: chỉ cần đăng nhập vào Groq, và chọn mô hình từ menu thả xuống ở góc trên bên phải. Để thử nghiệm (?), hãy chọn mô hình 70B.

Dưới đây là một hướng dẫn tôi đã thử. Ngay cả khi được chỉ định, đầu ra là bằng tiếng Anh, vì vậy từ biểu tượng bánh răng ở góc trên bên phải, bạn có thể thiết lập Hệ thống Nhắc để xuất ra tiếng Nhật trước.

Băng thông khoảng 300 token mỗi giây. Bạn sẽ thấy rằng tốc độ phản hồi cực nhanh. Tuy nhiên, đây là lệnh để tạo ra truyện ngắn, và tôi cảm thấy độ chính xác không tốt vì các câu văn giống nhau được lặp lại.

Ngoài ra, giao diện Groq không thực sự thực tiễn để sử dụng bởi vì việc nhấn phím Enter khi xác nhận chuyển đổi văn bản tiếng Nhật hoàn thành nhập liệu trò chuyện.

Thử phiên bản trình duyệt của Chatbot UI

Trong khi tìm kiếm một phương pháp tốt, tôi đã tìm thấy Chatbot UI, tương thích với API của Groq, cho phép sử dụng Llama 3 qua Groq.

Chatbot UI có giao diện giống như ChatGPT cho phép chuyển đổi giữa các LLM khác nhau.

Chatbot UI là mã nguồn mở và có thể chạy trên PC cá nhân, nhưng cũng có phiên bản trình duyệt. Vì thiết lập môi trường địa phương khá phức tạp, tôi quyết định sử dụng phiên bản trình duyệt trong thời gian này.

Để sử dụng nhiều LLM trong Chatbot UI, bạn cần kết nối với API của từng LLM, điều này đòi hỏi phải lấy các API key trước. Lần này, tôi muốn so sánh ChatGPT, Claude và Groq (Llama 3), vì vậy tôi sẽ lấy API key từ OpenAI, Anthropic và Groq. Ngoài API Key, bạn cũng cần ID Mô hình và URL Cơ sở (tùy chọn) cho kết nối API.

Lấy API key của OpenAI

Bạn có thể phát hành từ menu “API keys” trong màn hình quản lý của OpenAI.

Có một nút gọi là “Create new secret key”; khi nhấp vào đó sẽ mở một popup. Sau khi nhập thông tin cần thiết, API Key sẽ được phát hành. Hãy chắc chắn lưu key này một cách an toàn, vì về sau bạn không thể kiểm tra nó nữa.

Lấy API key của Anthropic

Bạn có thể lấy nó theo cách gần giống như OpenAI. Nó có thể được phát hành từ menu “API Keys” trong màn hình quản lý của Anthropic. Nhấp vào nút “Create Key”, một popup sẽ mở ra, và nhập tên cho key sẽ phát hành nó.

Lấy API key của Groq

Phương pháp lấy API Key của Groq tương tự như trên.

Có một liên kết gọi là “GroqCloud” ở phía dưới màn hình Groq mà bạn đã đăng nhập trước đó. Nhấp vào liên kết này mở màn hình quản lý, từ đó bạn có thể phát hành key từ “Create API Key” trong menu “API Keys”.

Đăng Nhập vào Phiên Bản Trình Duyệt của Chatbot UI

Bây giờ mọi thứ đã sẵn sàng, hãy đăng nhập vào phiên bản trình duyệt của Chatbot UI.

Khi bạn truy cập phiên bản trình duyệt của Chatbot UI, màn hình ban đầu hiển thị sẽ có nút “Start Chatting” (Bắt đầu trò chuyện). Nhấp vào nút này để tiếp tục.

Tôi muốn sử dụng chức năng tìm kiếm tệp để so sánh hiệu suất, vì vậy tôi đã chọn kế hoạch trả phí. Giá là $8 mỗi tháng với hợp đồng hàng năm, hoặc $10 mỗi tháng với việc gia hạn hàng tháng.

Khi cài đặt thanh toán hoàn tất, bạn sẽ được chuyển đến màn hình để đăng ký API Key như đã hiển thị. 

Nhập các API Key bạn đã nhận được trước đó từ OpenAI, Anthropic và Groq. Việc đăng ký ban đầu này hoàn tất và màn hình trò chuyện sẽ mở ra.

Tiếp theo, bạn cần đăng ký mô hình bạn muốn sử dụng (việc đăng ký API Key trước đó chưa đủ để sử dụng chúng). Dấu hiệu lấp lánh ở bên trái (?) là menu mô hình; nhấp vào nút “New Model” (Mô hình mới) để đăng ký mô hình bạn muốn sử dụng. 

Lần này, tôi đã nhập như sau:

  • ChatGPT 4-turbo
    • Name: ChatGPT 4-turbo (bất cứ cái gì bạn có thể nhận biết)
    • Model ID: gpt-4-turbo
    • Base URL: không có
    • API Key: Key bạn đã nhận được trước đó
  • Claude 3 Opus
    • Name: Claude 3 Opus (bất cứ cái gì bạn có thể nhận biết)
    • Model ID: claude-3-opus-20240229
    • Base URL: https://api.anthropic.com/v1
    • API Key: Key bạn đã nhận được trước đó
  • Groq (Llama-3-70B)
    • Name: Groq (Llama-3-70B) (bất cứ cái gì bạn có thể nhận biết)
    • Model ID: llama3-70b-8192
    • Base URL: https://api.groq.com/openai/v1
    • API Key: Key bạn đã nhận được trước đó

Bạn có thể tham khảo Model ID từ các nguồn sau:

Sau khi thiết lập, GPT4 và Groq đã hoạt động, nhưng đã xảy ra lỗi với Claude như đã mô tả ở trên. Mặc dù đã thử thay đổi Base URL và Model ID nhưng không giải quyết được, vì vậy tôi đã từ bỏ và quyết định xây dựng phiên bản cục bộ… Nếu bạn đã quản lý để làm cho Claude 3 Opus hoạt động trên trình duyệt, tôi sẽ rất vui nếu được nghe thêm từ bạn.

Thử nghiệm phiên bản cục bộ của Chatbot UI

Thiết lập môi trường cục bộ cho Chatbot UI

Phiên bản cục bộ có thể được thiết lập dễ dàng bằng các bước sau- bằng cách làm theo hướng dẫn trong kho lưu trữ GitHub của nhà phát triển Mckay Wrigley. Thực hiện các lệnh sau trong terminal của Mac để thiết lập.

1. Clone kho lưu trữ

Trước tiên, clone kho lưu trữ Chatbot UI từ GitHub.

$ git clone https://github.com/mckaywrigley/chatbot-ui.git

2. Cài đặt các biến phụ thuộc

Di chuyển đến thư mục đã clone và cài đặt các biến phụ thuộc cần thiết.

$ cd chatbot-ui

$ npm install

3. Cài đặt Docker

Docker là cần thiết để chạy Supabase tại cục bộ. Cài đặt Docker từ trang web chính thức.

4. Cài đặt Supabase CLI

Để cài đặt Supabase CLI, thực hiện lệnh sau.

$ brew install supabase/tap/supabase

5. Khởi động Supabase

Khởi động Supabase bằng cách thực hiện lệnh sau.

$ supabase start

6. Thiết lập các biến môi trường

Sao chép tệp .env.local.example thành .env.local và thiết lập các giá trị cần thiết.

$ cp .env.local.example .env.local

7. Kiểm tra thông tin liên quan đến Supabase

Nhập các giá trị cần thiết như API URL vào tệp .env.local.

$ supabase status

[Results of supabase status]

$ vi ./.env.local

  • NEXT_PUBLIC_SUPABASE_URL: Chỉ định “API URL” từ status
  • NEXT_PUBLIC_SUPABASE_ANON_KEY: Chỉ định “anon key” từ status
  • SUPABASE_SERVICE_ROLE_KEY: Chỉ định “service_role key” từ status

8. Thiết lập cơ sở dữ liệu

Chỉnh sửa tệp SQL để thiết lập ban đầu trong Supabase. Tham khảo kết quả supabase status cho các giá trị cụ thể. Trong trường hợp của tôi, nó hoạt động mà không cần thay đổi gì. Hãy chắc chắn rằng “service_role_key” ở dòng 54 khớp với service_role_key từ kết quả thực hiện supabase status.

9. Chạy ứng dụng cục bộ

Thực hiện lệnh sau để khởi động ứng dụng cục bộ

$ npm run chat

Bây giờ, bạn có thể truy cập ứng dụng Chatbot UI trong trình duyệt tại http://localhost:3000.

Thử nghiệm Llama 3 Qua Groq trên Chatbot UI

Khi bạn truy cập http://localhost:3000, màn hình thiết lập ban đầu giống như phiên bản trình duyệt sẽ được hiển thị. Đăng ký theo cách và nội dung giống như trước.

Tuy nhiên, Claude 3 Opus hiển thị lỗi “404 Not Found” và không thể sử dụng. Claude 3 Sonnet và Claude 3 Haiku, có thể chỉ định mặc định, hoạt động bình thường, vì vậy có vẻ như Chatbot UI có thể chưa hỗ trợ Opus.

Hãy chỉ định “Groq (Llama 3)” làm mô hình, và nhập một số lời nhắc để xem đầu ra của Llama 3.

Trước tiên, tôi đã hỏi một câu hỏi liên quan đến AI hiệu suất cao “GPT2” mới xuất hiện gần đây, câu hỏi không có câu trả lời dứt khoát.

Phản hồi rất mơ hồ. 

Tiếp theo, hãy kiểm tra tốc độ của đầu ra.

Tốc độ thực sự rất nhanh! Ngay cả bao gồm thời gian truyền thông API, không có bất kỳ sự chờ đợi căng thẳng nào!

Llama 3 vs GPT-4 Turbo vs Claude 3 Sonnet – Trận Đấu 3 Vòng!

Hãy nhập cùng ba lệnh vào Llama 3, GPT-4 Turbo và Claude 3 Sonnet và so sánh kết quả đầu ra.

Tạo Văn Bản Để Giải Thích Một Từ Khóa

Lệnh là:

“Retrieval-Augmented Generation (RAG) là gì? Hãy tóm tắt nó trong khoảng 500 ký tự để một sinh viên năm tư ngành công nghệ thông tin có thể hiểu.”

So sánh kết quả phản hồi đầu tra (từ trái: Llama 3, GPT-4 Turbo, Claude 3 Sonnet. Xác nhận lệnh từ phải trong chế độ nhanh gấp đôi)

So sánh kết quả phản hồi đầu ra (từ trái: Llama 3, GPT-4 Turbo, Claude 3 Sonnet)

Không có sự khác biệt đáng kể về độ chính xác của văn bản được tạo bởi bất kỳ mô hình nào trong số chúng. Tuy nhiên, tốc độ phản hồi đầu ra thì Groq vẫn nhanh hơn một cách áp đảo.

Giải Một Câu Đố Logic

Đây là một câu đố logic điển hình. Lệnh là:

“Ba người đàn ông ở lại khách sạn. Giá phòng là $10 mỗi người. Các khách hàng đã trả tổng cộng $30 cho lễ tân. Sau đó, lễ tân nhận ra rằng có một chiến dịch và giá phòng là $25 cho ba người, vì vậy họ quyết định trả lại $5. Tuy nhiên, lễ tân nghĩ rằng $5 không thể chia đều cho ba người, nên họ bỏ $2 vào túi và trả lại $3 cho khách hàng. Các khách hàng đã trả $9 mỗi người, tổng cộng $27. Thêm $2 mà lễ tân đã bỏ túi là $29. Vậy $1 còn lại đã đi đâu? Hãy cung cấp câu trả lời và giải thích.”

So sánh kết quả phản hồi đầu ra (từ trái: Llama 3, GPT-4 Turbo, Claude 3 Sonnet. Xác nhận lệnh từ phải trong chế độ nhanh gấp đôi.)

So Sánh Kết Quả Đầu Ra (từ trái: Llama 3, GPT-4 Turbo, Claude 3 Sonnet)

Kết quả là chiến thắng hoàn toàn dành cho GPT-4 Turbo. 

GPT-4 Turbo đã cung cấp một lời giải thích rõ ràng và logic về câu trả lời. 

Claude 3 Sonnet cũng đã cung cấp một lời giải thích, nhưng không rõ ràng về lý do.

Llama 3 không giải thích được lý do quan trọng, vì vậy được đánh giá là không thành công.

Tóm tắt một bài báo đã được đăng tải

Tôi đã tải lên một bài báo về một mô hình gọi là RAPTOR cải thiện độ chính xác tìm kiếm tài liệu trong RAG và yêu cầu tóm tắt nội dung. Lệnh là:

“Hãy tóm tắt nội dung của tệp PDF đã tải lên theo cách mà một sinh viên năm tư ngành công nghệ thông tin có thể hiểu, sử dụng các biểu hiện cụ thể và định lượng nhất có thể, trong khoảng 1000 ký tự tiếng Nhật. Vui lòng tóm tắt ngắn gọn nội dung kỹ thuật, tập trung vào các điểm quan trọng và nêu rõ cái gì là đổi mới và cái gì là vượt trội so với công nghệ truyền thống.”

Do thời gian tải tệp khá lâu và kết quả đầu ra dài, tôi sẽ dán trực tiếp văn bản được tạo ra dưới đây thay vì ảnh chụp màn hình.

ーーー

<Llama 3>

The provided text appears to be a collection of research papers and articles related to natural language processing, specifically focusing on language models, question answering, and text summarization. Here is a summary of the main points:

Language Models: The text discusses the development of language models, such as GPT-4, that can generate coherent and context-specific text. These models are trained on large datasets and can be fine-tuned for specific tasks.

Question Answering: The text explores various question answering models, including UnifiedQA, that can retrieve relevant information from a large corpus of text. These models are evaluated on datasets such as QuALITY, QASPER, and NarrativeQA.

Text Summarization: The text touches on the topic of text summarization, highlighting the importance of summarizing long documents and the challenges associated with it. Models such as Longformer and Hybrid Hierarchical Retrieval are mentioned as approaches to tackle this task.

Retrieval-Augmented Generation: The text discusses the concept of retrieval-augmented generation, which involves using a retrieval mechanism to augment language models with relevant information from a knowledge base.

Evaluation Metrics: The text mentions various evaluation metrics used to assess the performance of language models and question answering systems, including BLEU score, ROUGE score, and METEOR score.

Applications: The text hints at the potential applications of these technologies in areas such as information retrieval, text classification, and chatbots.

In summary, the text provides an overview of the current state of natural language processing research, highlighting advancements in language models, question answering, and text summarization.

ーーー

<GPT-4 Turbo>

Tóm tắt nội dung của tệp PDF, nó mô tả nghiên cứu về một hệ thống tìm kiếm thông tin mới gọi là “RAPTOR.” Hệ thống này có thể cung cấp thông tin liên quan và toàn diện hơn so với các hệ thống tìm kiếm truyền thống. Đặc biệt, RAPTOR có thể thu được các giải thích chi tiết về không chỉ các mô tả ngắn mà còn các phần chính của câu chuyện, cải thiện hiệu suất trong các nhiệm vụ hạ lưu.

Cụ thể, so sánh RAPTOR với DPR (Dense Passage Retrieval) truyền thống, RAPTOR được cho là có thể tạo ra các câu trả lời xem xét ngữ cảnh rộng hơn. Ví dụ, trong câu chuyện về Lọ Lem, DPR tạo ra các câu trả lời chỉ xem xét phần đầu của câu chuyện, trong khi RAPTOR nắm bắt chủ đề của câu chuyện từ góc nhìn toàn diện hơn, cung cấp một câu trả lời đầy đủ hơn.

Sự đổi mới của công nghệ này nằm ở khả năng trích xuất hiệu quả các phần quan trọng của thông tin, đặc biệt là trong các tài liệu dài hoặc văn bản chứa thông tin phức tạp, cho phép cung cấp thông tin chính xác hơn. Điều này cho phép người dùng truy cập thông tin mà họ tìm kiếm nhanh chóng và chính xác.

Ngoài ra, việc giới thiệu RAPTOR dự kiến sẽ cải thiện đáng kể độ chính xác và tốc độ của việc tìm kiếm thông tin so với các công nghệ thông thường. Đây sẽ là một lợi thế lớn, đặc biệt đối với các công ty và tổ chức nghiên cứu xử lý lượng dữ liệu lớn.

Nghiên cứu này đại diện cho một bước tiến quan trọng trong lĩnh vực tìm kiếm thông tin và dự kiến sẽ có nhiều ứng dụng hơn trong tương lai. Đối với một sinh viên chuyên ngành kỹ thuật thông tin, hiểu được công nghệ tiên tiến như vậy và khám phá tiềm năng của nó sẽ là một trải nghiệm rất có giá trị.

ーーー

<Claude 3 Sonnet>

Để tóm tắt nội dung của tệp PDF này, nó trình bày các điểm sau:

Công nghệ tìm kiếm và tóm tắt văn bản dài tiên tiến, RAPTOR, được giới thiệu. So với các phương pháp tìm kiếm truyền thống, nó có thể truy xuất chính xác thông tin liên quan hơn từ toàn bộ tài liệu, cải thiện đáng kể độ chính xác của các nhiệm vụ hỏi đáp.

Cụ thể, ba điểm sau đây là đổi mới:

Áp dụng phương pháp tìm kiếm hiệu quả xem xét cấu trúc phân cấp của tài liệu, cho phép trích xuất chính xác thông tin liên quan ngay cả từ các tài liệu dài và lớn.

Tóm tắt tự nhiên thông tin đã truy xuất, tạo ra các nút cấp cao có thể trả lời trực tiếp các câu hỏi.

Kết hợp với mô hình ngôn ngữ lớn GPT-4, đạt được độ chính xác cao trong việc hỏi đáp.

Các thí nghiệm đánh giá cho thấy độ chính xác của việc hỏi đáp trên tập dữ liệu NarrativeQA đã được cải thiện đáng kể so với các phương pháp truyền thống. Trong việc trích xuất thông tin và trả lời câu hỏi từ các tài liệu dài, RAPTOR đã đạt được những kết quả đột phá.

ーーー

Mặc dù tôi đã chỉ định “bằng tiếng Nhật”, nhưng Llama 3 đã phản hồi bằng tiếng Anh… Tuy nhiên, Llama 3 dường như có độ chi tiết kỹ thuật cao nhất. Có vẻ như Llama 3 và GPT-4 Turbo hòa, còn Claude 3 Sonnet đứng ở vị trí tiếp theo.

Tóm Tắt

Bài viết đã trở nên khá dài khi tôi thử nhiều điều khác nhau, nhưng để tóm tắt, đây là những ấn tượng của tôi:

  • Groq nhanh…! Và rẻ…!
  • Chatbot UI tiện lợi…!
  • Mặc dù Llama 3 được đánh giá cao hơn Claude 3 Sonnet trong các bảng xếp hạng, tôi đánh giá là gần như tương đương. Tuy nhiên, điểm cộng tuyệt vời là nó được công bố dưới dạng mã nguồn mở…!

Trying Out Meta’s Latest LLM “Llama 3” in Groq’s Ultra-Fast Environment

Hello, I am Kakeya, the representative of Scuti.

Our company specializes in Vietnamese offshore development and lab-based development with a focus on generative AI, as well as providing generative AI consulting services. Recently, we have been fortunate to receive numerous requests for system development in collaboration with generative AI.

Meta’s release of the latest LLM “Llama 3” as open source has become a hot topic. I immediately tried Llama 3 through a service called Groq, which allows for ultra-fast operation of LLM inference. I will share the process up to confirming its actual operation. Llama 3 was impressive, but I was astonished by Groq’s overwhelmingly fast response…!

Table of Contents

  • About Groq
  • About Llama 3
  • Trying out Llama 3 with Groq
  • Testing the browser version of Chatbot UI
  • Testing the local version of Chatbot UI

About Groq

What is Groq?

Groq is a company that develops custom hardware and Language Processing Units (LPUs) specifically designed to accelerate the inference of large language models (LLMs). This technology is characterized by significantly enhanced processing speeds compared to conventional hardware.

Groq’s LPUs achieve inference performance up to 18 times faster than other cloud-based service providers. Groq aims to maximize the performance of real-time AI applications using this technology.

Additionally, Groq has recently established a new division called “Groq Systems,” which assists in the deployment of chips to data centers and the construction of new data centers. Furthermore, Groq has acquired Definitive Intelligence, a provider of AI solutions for businesses, to further enhance its technological prowess and market impact.

What is Groq’s LPU?

Groq’s Language Processing Unit (LPU) is a dedicated processor designed to improve the inference performance and accuracy of AI language models. Unlike traditional GPUs, the LPU is specialized for sequential computation tasks, eliminating bottlenecks in computation and memory bandwidth required by large language models. For instance, the LPU reduces per-word computation time compared to GPUs and does not suffer from external memory bandwidth bottlenecks, thus offering significantly better performance than graphics processors.

Technically, the LPU contains thousands of simple processing elements (PEs), arranged in a Single Instruction, Multiple Data (SIMD) array. This allows the same instructions to be executed simultaneously for each data point. Additionally, a central control unit (CU) issues instructions and manages the flow of data between the memory hierarchy and PEs, maintaining consistent synchronous communication.

The primary effect of using an LPU is the acceleration of AI and machine learning tasks. For example, Groq has run the Llama-2 70B model at a rate of 300 tokens per second per user on the LPU system, which is a notable improvement over the previous 100 tokens and 240 tokens. Thus, the LPU can process AI inference tasks in real-time with low latency and provide them in an energy-efficient package. This enables groundbreaking changes in areas such as high-performance computing (HPC) and edge computing.

Groq’s Overwhelmingly Affordable API

Groq’s hardware resources can be accessed through an API provided by Groq. One of the notable features of Groq’s service is the incredibly low cost of using this API.

In other words, Groq is fast and affordable

The figure below, provided by ArtificialAnalysis.ai, compares API providers, mapping price on the horizontal axis and throughput (tokens per second) on the vertical axis. From this chart, it is visually evident that Groq is overwhelmingly fast and affordable, especially standing out in terms of throughput compared to other API providers.

About Llama 3

Overview and Features of Llama 3

Llama 3 is the latest large language model (LLM) developed by Meta. This AI is trained on an extensive dataset of text data, enabling it to comprehensively understand and respond to language. Llama 3 is suited for a wide array of tasks such as creating creative content, translating languages, and providing information for queries. The model is available on platforms like AWS, Databricks, and Google Cloud, serving as a foundation for developers and researchers to further advance AI. By offering Llama 3 as open source, Meta aims to enhance the transparency of the technology and foster collaboration with a broad range of developers.

Here are some features of Llama 3 as disclosed by Meta:

  • Advanced Parameter Model: Llama 3 is developed as a model with a vast number of parameters, available in two versions: 8B (eight billion) and 70B (seventy billion), which allow it to perform highly in language understanding and generation.
  • Industry-Standard Benchmark Evaluations: Llama 3 is evaluated using several industry-standard benchmarks such as ARC, DROP, and MMLU, achieving state-of-the-art results in these tests. Specifically, it shows significant improvements in recognition and reasoning accuracy compared to previous models.
  • Multilingual Support: Llama 3 is trained with a dataset of over 15 trillion tokens that includes high-quality data for supporting over 30 languages, not only in English but also various other languages, thus catering to users worldwide.
  • Tokenizer Innovation: The newly developed tokenizer has 128,000 vocabularies, enabling more efficient encoding of languages. This enhances the processing speed and accuracy of the model, allowing it to handle more complex texts appropriately.
  • Improved Inference Efficiency: The adoption of Grouped Query Attention (GQA) technology has enhanced the model’s inference efficiency, enabling high-speed processing of large datasets and facilitating real-time responses.
  • Enhanced Security and Reliability: New security tools like Llama Guard 2, Code Shield, and CyberSec Eval 2 have been introduced to enhance the safety of content generated by the model, minimizing the risk of producing inappropriate content.
  • Open Source Availability: Llama 3 is offered entirely as open source, allowing researchers and developers worldwide to freely access, improve, and use it for the development of applications. This openness promotes further innovation.

Performance Evaluation of Llama 3 According to Industry Benchmarks

Llama 3 has demonstrated remarkable performance in industry-standard benchmarks, particularly excelling in areas of language understanding, logical reasoning, and problem-solving abilities. The newly introduced models with 8 billion (8B) and 70 billion (70B) parameters show a significant improvement in understanding and response accuracy compared to the previous model, Llama 2.

Llama 3 is highly rated for its ability to generate code and perform tasks based on instructions, and these capabilities are further enhanced by Meta’s proprietary training methods. Additionally, this model supports multiple languages, delivering high-quality performance in over 30 languages.

Currently, Meta has released two versions of the Llama 3 model: Llama 3 8B and Llama 3 70B, where “8B” has 8 billion parameters and “70B” has 70 billion parameters. The figure below shows the results of the benchmark evaluations published by Meta, with the 70B model exhibiting higher performance than Gemini Pro and Claude3.

Source: https://llama.meta.com/llama3/ 

Comparison between Llama 3 and Llama 2

Llama 3 shows significant improvements over Llama 2 in many aspects. Specifically, the dataset used for training the model has increased sevenfold, and the data related to code has quadrupled.

As a result, the model has become more efficient at handling complex language tasks. Particularly, Llama 3 significantly surpasses the previous model in its ability to generate code and perform tasks based on instructions.

The adoption of a new tokenizer has also improved the efficiency of language encoding, contributing to an overall performance enhancement. Additionally, a design considering safety has been implemented, introducing a new filtering system to reduce the risk of generating inappropriate responses.

FeaturesLlama 3Llama 2
Parameter CountVersions with 8 billion & 70 billion parametersFewer parameters (exact numbers not disclosed)
Benchmark PerformanceImproved performance in ARC, DROP, MMLU benchmarksLikely lower scores in the same benchmarks
CapabilitiesStrong reasoning abilities, creative text generationMay offer similar abilities but to a lesser extent
AvailabilityPublicly available for researchDetails on availability unclear, probably less publicly accessible

Trying Out Llama 3 with Groq

As of now, “Meta AI” is not yet available in Japan

Ideally, one would be able to experience fast image generation and chatting using Llama 3 through the AI assistant “Meta AI” released by Meta. However, as of May 1, 2024, when this article was written, Meta AI is only available in English-speaking regions and has not been released in Japan.

Trying Llama 3 with Groq

Nevertheless, since the aforementioned Groq supports Llama 3, you can experience Llama 3 by logging into Groq. Moreover, it’s super fast and free!

The usage is simple: just log into Groq, and select the model from the dropdown menu in the upper right corner. For the sake of trying (?), let’s choose the 70B model.

Here is a casual instruction I tried. Even if specified, the output is in English, so from the gear icon in the upper right, you can set the System Prompt to output in Japanese beforehand.

The throughput is about 300 tokens per second. You should be able to tell that the output is incredibly fast. However, it is a prompt for generating short stories, and I felt the precision was not good as the same sentences were repeatedly generated.

Additionally, the Groq interface is not very practical for actual use because pressing the enter key when confirming the conversion of Japanese text finalizes the chat input.

Trying the Browser Version of Chatbot UI

While searching for a good method, I found something called Chatbot UI, which is compatible with Groq’s API, enabling the use of Llama 3 through Groq.

Chatbot UI has a ChatGPT-like interface that allows switching between various LLMs.

Chatbot UI is open-source and can be run on a local PC, but a browser version is also available. Since setting up a local environment was cumbersome, I decided to use the browser version for now.

To use multiple LLMs in Chatbot UI, you need to connect to each LLM’s API, which requires obtaining API keys first. For this instance, I want to compare ChatGPT, Claude, and Groq (Llama 3), so I will obtain API keys from OpenAI, Anthropic, and Groq. Besides the API Key, you also need the Model ID and Base URL (optional) for the API connection.

Obtaining OpenAI’s API Key

You can issue it from the “API keys” menu in the OpenAI management screen.

There’s a button called “Create new secret key”; clicking it will open a popup. Once you enter the necessary information, the API Key will be issued. Make sure to save this key securely, as you cannot check it later.

Obtaining Anthropic’s API Key

You can obtain it in almost the same way as OpenAI. It can be issued from the “API Keys” menu in the Anthropic management screen. Click the “Create Key” button, a popup will open, and entering a name for the key will issue it.

Obtaining Groq’s API Key

The method to obtain Groq’s API Key is similar to the above.

There is a link called “GroqCloud” at the bottom of the Groq screen you logged into earlier. Clicking this link opens the management screen, where you can issue a key from the “Create API Key” in the “API Keys” menu.

Logging into the Browser Version of Chatbot UI

Now that everything is set up, let’s log into the browser version of Chatbot UI.

When you access the browser version of Chatbot UI, the initial screen displayed will have a “Start Chatting” button. Click this button to proceed.

I wanted to use the file search function for performance comparison, so I opted for the paid plan. It’s $8 per month with an annual contract, or $10 per month with monthly renewals.

Once the payment setup is complete, you will be taken to a screen to register the API Key as shown above. 

Enter the keys you obtained earlier from OpenAI, Anthropic, and Groq. This completes the initial registration, and the chat screen will open.

Next, you need to register the model you want to use (the previous API Key registration was not enough to use them). The sparkly mark on the left (?) is the model menu; click the “New Model” button to register the model you want to use. This time, I entered the following:

  • ChatGPT 4-turbo
    • Name: ChatGPT 4-turbo (whatever you can identify)
    • Model ID: gpt-4-turbo
    • Base URL: none
    • API Key: the one you obtained earlier
  • Claude 3 Opus
    • Name: Claude 3 Opus (whatever you can identify)
    • Model ID: claude-3-opus-20240229
    • Base URL: https://api.anthropic.com/v1
    • API Key: the one you obtained earlier
  • Groq (Llama-3-70B)
    • Name: Groq (Llama-3-70B) (whatever you can identify)
    • Model ID: llama3-70b-8192
    • Base URL: https://api.groq.com/openai/v1
    • API Key: the one you obtained earlier

You can refer to the following for the Model ID:

Having set this up, GPT4 and Groq worked, but there was an error with Claude as described above. Despite trying to change the Base URL and Model ID, it wasn’t resolved, so I gave up and decided to build a local version instead… If anyone has managed to get Claude 3 Opus working in the browser, I would be happy to hear from you.

Trying the Local Version of Chatbot UI

Setting Up the Local Environment for Chatbot UI

The local version can be set up easily using the following steps. It worked by following the instructions in the GitHub repository by developer Mckay Wrigley. Execute the following commands in the Mac terminal to set it up.

1. Clone the Repository

First, clone the Chatbot UI repository from GitHub.

$ git clone https://github.com/mckaywrigley/chatbot-ui.git 

2. Install Dependencies

Navigate to the cloned directory and install the necessary dependencies.

$ cd chatbot-ui

$ npm install

3. Install Docker

Docker is required to run Supabase locally. Install Docker from the official site.

4. Install Supabase CLI

To install the Supabase CLI, execute the following command.

$ brew install supabase/tap/supabase

5. Start Supabase

Start Supabase by executing the following command.

$ supabase start

6. Set Environment Variables

Copy the .env.local.example file to .env.local and set the necessary values.

$ cp .env.local.example .env.local

7. Check Supabase Related Information

Enter the necessary values such as the API URL in the .env.local file.

$ supabase status

[Results of supabase status]

$ vi ./.env.local

  • NEXT_PUBLIC_SUPABASE_URL: Specify the “API URL” from the status
  • NEXT_PUBLIC_SUPABASE_ANON_KEY: Specify the “anon key” from the status
  • SUPABASE_SERVICE_ROLE_KEY: Specify the “service_role key” from the status

8. Set Up the Database

Edit the SQL file for initial setup in Supabase. Refer to the supabase status results for specific values. In my case, it worked without any changes. Just make sure that the “service_role_key” in line 54 matches the service_role_key from the supabase status results.

9. Run the Application Locally

Execute the following command to start the application locally.

$ npm run chat

Now, you can access the Chatbot UI application in the browser at http://localhost:3000.

Trying Llama 3 via Groq on Chatbot UI

When you access http://localhost:3000, an initial setup screen similar to the browser version will be displayed. Register in the same way and content as before.

However, Claude 3 Opus showed a “404 Not Found” error and could not be used. Claude 3 Sonnet and Claude 3 Haiku, which can be specified by default, worked normally, so it seems that Chatbot UI may not support Opus yet.

Let’s specify “Groq (Llama 3)” as the model, and enter some prompts to see the output of Llama 3.

First, I asked a question related to the recently emerging mysterious high-performance AI “GPT2,” which has no definitive answer.

The response was quite ambiguous. 

Next, let’s check the speed of the output.

It’s undeniably fast! Even including the API communication time, there is no stress of waiting at all.

Llama 3 vs GPT-4 Turbo vs Claude 3 Sonnet – 3 Match Battle!

Let’s input the same three prompts to Llama 3, GPT-4 Turbo, and Claude 3 Sonnet and compare the outputs.

Generating a Text to Explain a Keyword

The prompt is:

“What is Retrieval-Augmented Generation (RAG)? Summarize it in about 500 characters in a way that a fourth-year university student majoring in computer engineering can understand.”

Comparison of Outputs (from left: Llama 3, GPT-4 Turbo, Claude 3 Sonnet. Prompts confirmed from right in double speed. Double speed.)

Comparison of Output Results (from left: Llama 3, GPT-4 Turbo, Claude 3 Sonnet)

There was no significant difference in the precision of the text output by any of them. However, the output speed was overwhelmingly faster with Groq.

Solving a Logical Quiz

Here is a typical logical quiz. The prompt is:

“Three men stay at a hotel. The room rate is $10 per person. The guests paid a total of $30 to the receptionist. Later, the receptionist realized that there was a campaign and the room rate was $25 for three people, so they decided to return $5. However, the receptionist thought that $5 couldn’t be evenly divided among three people, so they put $2 in their pocket and returned $3 to the guests. The guests paid $9 each, totaling $27. Adding the $2 the receptionist pocketed makes $29. Where did the remaining $1 go? Provide an answer and explanation.”

Comparison of Outputs (from left: Llama 3, GPT-4 Turbo, Claude 3 Sonnet. Prompts confirmed from right in double speed. Double speed.)

Comparison of Output Results (from left: Llama 3, GPT-4 Turbo, Claude 3 Sonnet)

The result was a complete victory for GPT-4 Turbo. 

GPT-4 Turbo provided a logical and clear explanation of the answer. 

Claude 3 Sonnet also provided an explanation, but it felt somewhat unclear in its reasoning. Llama 3 did not explain the essential “why,” so it was judged to be unsuccessful.

Summary of an Uploaded Article

I uploaded a paper about a model called RAPTOR that improves document retrieval accuracy in RAG and asked for a summary. The prompt is:

“Summarize the content of the uploaded PDF file in a way that a fourth-year university student majoring in computer engineering can understand, using as concrete and quantitative expressions as possible, in about 1000 characters in Japanese. Please briefly summarize the technical content focusing on important points and clearly state what is innovative and what is superior compared to conventional technology.”

Due to the considerable time it took to upload the file and the length of the output, I will paste the generated text directly below instead of screenshots.

ーーー

<Llama 3>

The provided text appears to be a collection of research papers and articles related to natural language processing, specifically focusing on language models, question answering, and text summarization. Here is a summary of the main points:

Language Models: The text discusses the development of language models, such as GPT-4, that can generate coherent and context-specific text. These models are trained on large datasets and can be fine-tuned for specific tasks.

Question Answering: The text explores various question answering models, including UnifiedQA, that can retrieve relevant information from a large corpus of text. These models are evaluated on datasets such as QuALITY, QASPER, and NarrativeQA.

Text Summarization: The text touches on the topic of text summarization, highlighting the importance of summarizing long documents and the challenges associated with it. Models such as Longformer and Hybrid Hierarchical Retrieval are mentioned as approaches to tackle this task.

Retrieval-Augmented Generation: The text discusses the concept of retrieval-augmented generation, which involves using a retrieval mechanism to augment language models with relevant information from a knowledge base.

Evaluation Metrics: The text mentions various evaluation metrics used to assess the performance of language models and question answering systems, including BLEU score, ROUGE score, and METEOR score.

Applications: The text hints at the potential applications of these technologies in areas such as information retrieval, text classification, and chatbots.

In summary, the text provides an overview of the current state of natural language processing research, highlighting advancements in language models, question answering, and text summarization.

ーーー

<GTP 4 Turbo>

Summarizing the content of the PDF file, it describes research on a new information retrieval system called “RAPTOR.” This system can provide more relevant and comprehensive information compared to traditional retrieval systems. Particularly, RAPTOR can obtain detailed explanations of not only short descriptions but also the main parts of a story, improving performance in downstream tasks.

Specifically, comparing RAPTOR with traditional DPR (Dense Passage Retrieval), RAPTOR is said to be able to generate answers considering a broader context. For example, in the story of Cinderella, DPR generates answers considering only the early part of the story, while RAPTOR captures the theme of the story from a more holistic perspective, providing a more complete answer.

The innovation of this technology lies in its ability to efficiently extract important parts of information, especially in long documents or texts containing complex information, enabling more accurate information provision. This allows users to access the information they seek quickly and accurately.

Additionally, the introduction of RAPTOR is expected to significantly improve the accuracy and speed of information retrieval compared to conventional technologies. This would be a significant advantage, especially for companies and research institutions dealing with large amounts of data.

This research represents an important advancement in the field of information retrieval and is expected to have many more applications in the future. For a student majoring in computer engineering, understanding such cutting-edge technology and exploring its potential will be a highly valuable experience.

ーーー

<Claude 3 Sonnet>

To summarize the content of this PDF, it presents the following points:

The innovative long-text retrieval and summarization technology, RAPTOR, is introduced. Compared to traditional retrieval methods, it can accurately retrieve more relevant information from entire documents, significantly improving the accuracy of question-answering tasks.

Specifically, the following three points are innovative:

It adopts an efficient retrieval method that considers the hierarchical structure of documents, enabling the precise extraction of relevant information even from large, long documents.

It naturally summarizes the retrieved information, generating top-level nodes that can directly answer questions.

Combined with the large language model GPT-4, it achieves high question-answering accuracy.

Evaluation experiments show that the question-answering accuracy on the NarrativeQA dataset has significantly improved compared to traditional methods. In extracting information and answering questions from long documents, RAPTOR has achieved groundbreaking results.

ーーー

Even though I specified “in Japanese,” Llama 3 responded in English… However, Llama 3 seems to have the highest technical specificity. It feels like a draw between Llama 3 and GPT-4 Turbo, with Claude 3 Sonnet coming in next.

Summary

The article has become quite long as I tried various things, but to summarize, here are my impressions:

  • Groq is fast…! And affordable…!
  • Chatbot UI is convenient…!
  • Although Llama 3 was rated slightly better than Claude 3 Sonnet in benchmarks, it feels almost equivalent in practice. However, it is amazing that this is available as open source…!