[OWASP-LLM] Top 10 List for Large Language Models version 0.1 - (1) prompt injections
ยท
๐Ÿƒ Routine
LLM01:2023 ํ”„๋กฌํ”„ํŠธ ์ธ์ ์…˜ ์„ค๋ช…: ํ”„๋กฌํ”„ํŠธ ์ธ์ ์…˜์€ ํ•„ํ„ฐ๋ฅผ ์šฐํšŒํ•˜๊ฑฐ๋‚˜ LLM์„ ์กฐ์ž‘ํ•˜์—ฌ ๋ชจ๋ธ์ด ์ด์ „ ๋ช…๋ น์„ ๋ฌด์‹œํ•˜๊ฑฐ๋‚˜ ์˜๋„ํ•˜์ง€ ์•Š์€ ๋™์ž‘์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ํ•˜๋Š” ์„ค๊ณ„๋œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์„ ๋งํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ทจ์•ฝ์ ์€ ๋ฐ์ดํ„ฐ ์œ ์ถœ, ๋ฌด๋‹จ ์ ‘๊ทผ ๋˜๋Š” ๋‹ค๋ฅธ ๋ณด์•ˆ ์นจํ•ด์™€ ๊ฐ™์€ ์˜๋„ํ•˜์ง€ ์•Š์€ ๊ฒฐ๊ณผ๋ฅผ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ํ”„๋กฌํ”„ํŠธ ์ธ์ ์…˜ ์ทจ์•ฝ์ : LLM์„ ์กฐ์ž‘ํ•˜์—ฌ ๋ฏผ๊ฐํ•œ ์ •๋ณด๋ฅผ ๋…ธ์ถœ์‹œํ‚ค๋Š” ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ž‘์„ฑํ•˜๋Š” ๊ฒƒ. ํŠน์ • ์–ธ์–ด ํŒจํ„ด์ด๋‚˜ ํ† ํฐ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•„ํ„ฐ๋‚˜ ์ œํ•œ์„ ์šฐํšŒํ•˜๋Š” ๊ฒƒ. LLM์˜ ํ† ํฐํ™” ๋˜๋Š” ์ธ์ฝ”๋”ฉ ๋ฉ”์ปค๋‹ˆ์ฆ˜์˜ ์ทจ์•ฝ์ ์„ ์ด์šฉํ•˜๋Š” ๊ฒƒ. ์ž˜๋ชป๋œ ๋ฌธ๋งฅ์„ ์ œ๊ณตํ•˜์—ฌ LLM์ด ์˜๋„ํ•˜์ง€ ์•Š์€ ๋™์ž‘์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ํ•˜๋Š” ๊ฒƒ. ์˜ˆ๋ฐฉ ๋ฐฉ๋ฒ•: ์‚ฌ์šฉ์ž๊ฐ€ ์ œ๊ณตํ•˜๋Š” ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด ์—„๊ฒฉํ•œ ์ž…๋ ฅ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ๋ฅผ ๊ตฌํ˜„ํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ..
[2023-9] ํ˜„๋ช…ํ•œ ์ดˆ๋ณด ํˆฌ์ž์ž
ยท
๐Ÿ“š Book
[2023-8] ์—ญ์‚ฌ์† ๊ฒฝ์ œ ์ด์•ผ๊ธฐ
ยท
๐Ÿ“š Book
์ค‘๊ตญ์˜ ์—ญ์‚ฌ์™€ ํ˜„๋Œ€ ๊ฒฝ์ œ๋ฅผ ์„œ๋กœ ๋น„๊ตํ•˜๋ฉฐ ์ง€๊ธˆ ๊นŒ์ง€ ์–ด๋–ป๊ฒŒ ๊ฒฝ์ œ๊ฐ€ ๋ฐœ์ „ํ•ด์™”๋Š”์ง€๋ฅผ ์‚ฌ๋ก€๋ฅผ ํ†ตํ•ด ์ด์•ผ๊ธฐํ•ด์ค€๋‹ค. ex) ์กฐ๊ฐœ ํŒจ๋ฅผ ์‚ฌ์šฉํ•˜๋‹ค ์ง„์‹œํ™ฉ ๋•Œ ๊ตญ๊ฐ€๊ฐ„ ํ†ตํ•ฉ์œผ๋กœ ํ™”ํ ๊ฐœํ˜์ด ์ผ์–ด๋‚œ๋‹ค. ์ด๋•Œ ๋ฌด๋ ฅ์œผ๋กœ ์ธํ•œ ํ†ตํ•ฉ์œผ๋กœ ๊ตญ๊ฐ€๊ฐ„ ๋นˆ๋ถ€๊ฒฉ์ฐจ๊ฐ€ ์‹ฌํ–ˆ๋Š”๋ฐ ์ด ๊ฒƒ์„ ์œ ๋Ÿฝ ์—ฐํ•ฉ์˜ ์œ ๋กœ ํ™”ํ๋ฅผ ๋น„๊ตํ•˜๋ฉฐ ์„ค๋ช…
Text Embedding + t-SNE Visualization
ยท
๐Ÿ—ฃ๏ธ Natural Language Processing
https://betterprogramming.pub/openais-embedding-model-with-vector-database-b69014f04433 OpenAI’s Embedding Model With Vector Database The updated Embedding model offers State-of-the-Art performance with 4x longer context window. Thew new model is 90% cheaper. The smaller… betterprogramming.pub Introduction OpenAI๋Š” 2022๋…„ 12์›” ์ž„๋ฒ ๋”ฉ ๋ชจ๋ธ์„ ํ…์ŠคํŠธ ์ž„๋ฒ ๋”ฉ-ada-002๋กœ ์—…๋ฐ์ดํŠธํ–ˆ์Šต๋‹ˆ๋‹ค. ์ƒˆ ๋ชจ๋ธ์€ ๋‹ค์Œ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค: 90%-99.8% ์ €๋ ดํ•œ ๋น„์šฉ 1/8..
On the amortized complexity of approximate counting
ยท
๐Ÿ Python
https://arxiv.org/abs/2211.03917 On the amortized complexity of approximate counting (์ƒ๊ฐ ๋ณต์žก๋„ ๊ทผ์‚ฌ ๊ณ„์‚ฐ) Abstract ๊ฐ’ n๊นŒ์ง€์˜ ์นด์šดํ„ฐ๋ฅผ ๋‹จ์ˆœํžˆ ์ €์žฅํ•˜๋Š” ๊ฒƒ์€ โ„ฆ(log n)๋น„ํŠธ์˜ ๋ฉ”๋ชจ๋ฆฌ๊ฐ€ ํ•„์š”. Nelson๊ณผ Yu [NY22]๋Š” Morris์˜ ์—ฐ๊ตฌ [Mor78]๋ฅผ ๋”ฐ๋ผ, ์ฟผ๋ฆฌ ์‘๋‹ต์ด (1+ฯต)-๊ทผ์‚ฌ๊ฐ’์ด๋ฉฐ ํ™•๋ฅ ์ ์œผ๋กœ ์ตœ์†Œ 1-δ๋ฅผ ๋งŒ์กฑํ•œ๋‹ค๋ฉด O(log(logn)+log(log(1/δ)+log(1/ฯต))๋น„ํŠธ๋งŒ์ด ํ•„์š”ํ•˜๋ฉฐ, ์‚ฌ์‹ค ์ด ๊ฒฝ๊ณ„๋Š” ์ตœ์ ์ด๋‹ค. Morris์˜ ์›๋ž˜ ์—ฐ๊ตฌ ๋™๊ธฐ์™€ ํ˜„๋Œ€ ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์€ ํ•˜๋‚˜์˜ ์นด์šดํ„ฐ๋ฅผ ์œ ์ง€ํ•˜๋Š” ๊ฒƒ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ k๊ฐœ์˜ ์นด์šดํ„ฐ๋ฅผ ์œ ์ง€ํ•ด์•ผํ•œ๋‹ค. ์ด๋กœ ์ธํ•ด ๋‹ค์Œ ์งˆ๋ฌธ์ด ์ƒ๊ธด๋‹ค. k๊ฐœ์˜ ์นด์šดํ„ฐ๋ฅผ ์œ ์ง€ํ•˜๋ฉด์„œ ..
[Langchain] paper-translator
ยท
๐Ÿ—ฃ๏ธ Natural Language Processing
https://github.com/seohyunjun/paper-translator GitHub - seohyunjun/paper-translator: pdf paper translator pdf paper translator. Contribute to seohyunjun/paper-translator development by creating an account on GitHub. github.com Version History v0.1.2 2023/6/15 ChatGPT API Update : gpt-3.5-turbo-16k token 4k -> 16k (about 3 pages cover per 1 request) https://openai.com/blog/function-calling-and-ot..
๋‹คํ–ˆ๋‹ค
B's