๋ณธ๋ฌธ์œผ๋กœ ๊ฑด๋„ˆ๋›ฐ๊ธฐ
buildradar
๋กœ๊ทธ์ธ

datawhalechina/diy-llm

@datawhalechina

๐ŸŽ“ ์ฒด๊ณ„์ ์ธ LLM ๊ตฌ์ถ• ์ฝ”์Šค๏ฝœ๐Ÿ› ๏ธ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์—”์ง€๋‹ˆ์–ด๋ง, Tokenizer, Transformer, MoE, GPU ํ”„๋กœ๊ทธ๋ž˜๋ฐ (CUDA/Triton), ๋ถ„์‚ฐ ํ•™์Šต, Scaling Laws, ์ถ”๋ก  ์ตœ์ ํ™” ๋ฐ ์ •๋ ฌ (SFT/RLHF/GRPO) ํฌํ•จ๏ฝœ๐Ÿš€ 6๊ฐœ์˜ ์ ์ง„์  ๊ณผ์ œ + ์ฝ”๋“œ ๊ธฐ๋ฐ˜, LLM ํ’€์Šคํƒ ์ธ์ง€ ์ฒด๊ณ„ ๊ตฌ์ถ•

์Šคํƒ€
1,338
ํฌํฌ
135
์–ธ์–ด
Jupyter Notebook
๋ผ์ด์„ ์Šค
โ€”
๋งˆ์ง€๋ง‰ ํ‘ธ์‹œ
1์ผ ์ „
Jupyter Notebookllmrltransformernlpgpu-programminggrposftdatawhale

์•„์ง ๊ด€๋ จ ์ธํ…”์ด ์—†์Šต๋‹ˆ๋‹ค

์ด ์ €์žฅ์†Œ๋Š” radar๊ฐ€ ์ถ”์ ํ•˜๋Š” ์–ด๋–ค ์†Œ์Šค์—๋„ ์•„์ง ๋“ฑ์žฅํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ˆ˜์ง‘๊ธฐ๋Š” ์˜ˆ์•ฝ์— ๋”ฐ๋ผ ์‹คํ–‰๋ฉ๋‹ˆ๋‹คโ€”โ€”์ด ์ €์žฅ์†Œ๋ฅผ ๋‹ค๋ฃจ๊ฒŒ ๋˜๋ฉด ๋‹ค์‹œ ํ™•์ธํ•˜์„ธ์š”.