Archive 8

[λ²ˆμ—­] The History of Open-Source LLMs: Part β…‘. Better Base Models

The History of Open-Source LLMs: Part β… . Early days The History of Open-Source LLMs: Part β…‘. Better Base Models The History of Open-Source LLMs: Part β…’. Imitations and alignment ⭐ 글에 μ‚½μž…λœ λͺ¨λ“  μ΄λ―Έμ§€μ˜ μΆœμ²˜λŠ” μ›λ¬Έμž…λ‹ˆλ‹€. 초기 μ˜€ν”ˆμ†ŒμŠ€ LLM - κ³΅κ°œλ˜μ§€ μ•Šμ€ 사전 ν•™μŠ΅ λͺ¨λΈλ“€κ³Ό λΉ„κ΅ν•˜λ©΄ μ„±λŠ₯이 많이 λ–¨μ–΄μ§„λ‹€λŠ” 단점이 μ‘΄μž¬ν•¨ 1) LLM ν•™μŠ΅ νŒŒμ΄ν”„λΌμΈ β‘  λŒ€λŸ‰μ˜ μ›μ‹œ 데이터λ₯Ό μ΄μš©ν•΄ λͺ¨λΈ 사전 ν•™μŠ΅ β‘‘ SFT와 RLHF 같은 κΈ°μˆ μ„ μ΄μš©ν•΄ alignment μˆ˜ν–‰ β‘’ LLM을 νŠΉμ • νƒœμŠ€ν¬μ— μ μš©ν•˜κΈ° μœ„ν•΄ 파인 νŠœλ‹ λ˜λŠ” in-context learning μˆ˜ν–‰..

Archive 2023.10.25

[λ²ˆμ—­] The History of Open-Source LLMs: Part β… . Early days

연ꡬ싀 μ˜€λΉ κ°€ κ³΅μœ ν•΄μ€€ μ˜€ν”ˆμ†ŒμŠ€ LLM λ°œμ „ 과정에 λŒ€ν•œ 글을 읽고 λ²ˆμ—­ 및 μΆ”κ°€ κ³΅λΆ€ν•œ λ‚΄μš©μ„ 정리해본닀. The History of Open-Source LLMs: Part β… . Early days The History of Open-Source LLMs: Part β…‘. Better Base Models The History of Open-Source LLMs: Part β…’. Imitations and alignment ⭐ 글에 μ‚½μž…λœ λͺ¨λ“  μ΄λ―Έμ§€μ˜ μΆœμ²˜λŠ” μ›λ¬Έμž…λ‹ˆλ‹€. LLM λ“±μž₯ λ°°κ²½ - μ–Έμ–΄ λͺ¨λΈ μžμ²΄λŠ” 역사가 였래 λμ§€λ§Œ self-supervised pre-trainingκ³Ό in-context learning을 μ‘°ν•©ν•˜μ—¬ μ—¬λŸ¬ νƒœμŠ€ν¬μ—μ„œ 인상 κΉŠμ€ few-shot learning μ„±λŠ₯을 보인 GP..

Archive 2023.10.23

[λ²ˆμ—­] Micro, Macro & Weighted Averages of F1 Score, Clearly Explained

πŸ’¬ μ΅œλŒ€ν•œ λ§€λ„λŸ½κ²Œ ν•΄μ„ν•˜κ³ μž λ…Έλ ₯ν–ˆμ§€λ§Œ μ–΄μƒ‰ν•œ λ¬Έμž₯이 μžˆμ„ 수 μžˆμŠ΅λ‹ˆλ‹€. ν”Όλ“œλ°±μ€ μ–Έμ œλ‚˜ ν™˜μ˜μž…λ‹ˆλ‹€ πŸ™‚ 원본 κΈ€ μ£Όμ†Œ: https://towardsdatascience.com/micro-macro-weighted-averages-of-f1-score-clearly-explained-b603420b292f Micro, Macro & Weighted Averages of F1 Score, Clearly Explained Understanding the concepts behind the micro average, macro average and weighted average of F1 score in multi-class classification towardsdatascience.com F1 Score(..

Archive 2022.12.20

[λ²ˆμ—­] Foundations of NLP Explained Visually: Beam Search, How It Works

πŸ’¬ μ΅œλŒ€ν•œ λ§€λ„λŸ½κ²Œ ν•΄μ„ν•˜κ³ μž λ…Έλ ₯ν–ˆμ§€λ§Œ μ–΄μƒ‰ν•œ λ¬Έμž₯이 μžˆμ„ 수 μžˆμŠ΅λ‹ˆλ‹€. ν”Όλ“œλ°±μ€ μ–Έμ œλ‚˜ ν™˜μ˜μž…λ‹ˆλ‹€ πŸ™‚ 원본 κΈ€ μ£Όμ†Œ : https://towardsdatascience.com/foundations-of-nlp-explained-visually-beam-search-how-it-works-1586b9849a24 Foundations of NLP Explained Visually: Beam Search, How it Works A Gentle Guide to how Beam Search enhances predictions, in Plain English towardsdatascience.com NLP λͺ¨λΈμ˜ Output 생성 방법 기계 λ²ˆμ—­ μž‘μ—…μ— 자주 μ‚¬μš©λ˜λŠ” sequence-to-sequence λͺ¨..

Archive 2022.08.01

[λ²ˆμ—­] Word2Vec Research Paper Explained

πŸ’¬ μ΅œλŒ€ν•œ λ§€λ„λŸ½κ²Œ ν•΄μ„ν•˜κ³ μž λ…Έλ ₯ν–ˆμ§€λ§Œ μ–΄μƒ‰ν•œ λ¬Έμž₯이 μžˆμ„ 수 μžˆμŠ΅λ‹ˆλ‹€. ν”Όλ“œλ°±μ€ μ–Έμ œλ‚˜ ν™˜μ˜μž…λ‹ˆλ‹€ πŸ™‚ 원본 κΈ€ μ£Όμ†Œ : https://towardsdatascience.com/word2vec-research-paper-explained-205cb7eecc30 Word2Vec Research Paper Explained An Intuitive understanding and explanation of the word2vec model. towardsdatascience.com μ„œλ‘  λ§Žμ€ NLP ν™œμš© μ‚¬λ‘€μ—μ„œ, λ‹¨μ–΄λŠ” μ„œλ‘œ κ°„μ˜ 관계λ₯Ό ν¬μ°©ν•˜μ§€ μ•ŠλŠ” 원-ν•« μΈμ½”λ”©μœΌλ‘œ ν‘œν˜„λœλ‹€. 원-ν•« 인코딩을 μ‚¬μš©ν•˜λŠ” μ΄μœ λŠ” "간단함, 견고함과 λŒ€λŸ‰μ˜ λ°μ΄ν„°λ‘œ ν›ˆλ ¨λœ κ°„λ‹¨ν•œ λͺ¨λΈμ΄ μ†ŒλŸ‰μ˜ λ°μ΄ν„°λ‘œ ν›ˆλ ¨λœ λ³΅μž‘ν•œ λͺ¨λΈλ³΄λ‹€..

Archive 2022.07.04

[λ²ˆμ—­] Introduction to Stemming and Lemmatization

πŸ’¬ μ΅œλŒ€ν•œ λ§€λ„λŸ½κ²Œ ν•΄μ„ν•˜κ³ μž λ…Έλ ₯ν–ˆμ§€λ§Œ μ–΄μƒ‰ν•œ λ¬Έμž₯이 μžˆμ„ 수 μžˆμŠ΅λ‹ˆλ‹€. ν”Όλ“œλ°±μ€ μ–Έμ œλ‚˜ ν™˜μ˜μž…λ‹ˆλ‹€ πŸ™‚ 원본 κΈ€ μ£Όμ†Œ : https://medium.com/geekculture/introduction-to-stemming-and-lemmatization-nlp-3b7617d84e65 μžμ—°μ–΄μ²˜λ¦¬(Natural Language Processing, NLP) ν…μŠ€νŠΈ λ°μ΄ν„°λŠ” λ‹€μ–‘ν•œ μ†ŒμŠ€λ‘œλΆ€ν„° 얻을 수 μžˆλ‹€. 우리의 λͺ©ν‘œλŠ” νƒœμŠ€ν¬μ™€ κ΄€λ ¨μ—†λŠ” μ†ŒμŠ€ 고유의 λ§ˆν¬μ—…μ΄λ‚˜ ꡬ문이 μ—†λŠ” 평문을 μΆ”μΆœν•˜λŠ” 것이닀. ꡬ두점, λŒ€λ¬Έμžμ™€ 같은 μ–Έμ–΄μ˜ λͺ‡λͺ‡ νŠΉμ„±κ³Ό "a"/"of"/"the"와 같은 일반적인 단어듀이 λ¬Έμ„œμ˜ ꡬ쑰 μ œκ³΅μ— 도움을 μ£ΌκΈ΄ ν•˜μ§€λ§Œ λ§Žμ€ 의미λ₯Ό μ£Όμ§€λŠ” μ•ŠλŠ”λ‹€. λ”°λΌμ„œ ν…μŠ€νŠΈ 데이터λ₯Ό λΆ„μ„ν•˜μ—¬ μžμ—°μ–΄ 처리 파이..

Archive 2022.04.01

[λ²ˆμ—­] Entropy, Cross-Entropy, KL-Divergence

μ—”νŠΈλ‘œν”Ό κ°œλ…μ€ λ„ˆλ¬΄λ‚˜λ„ λ³΅μž‘λ‚œν•΄ν•΄ ~ πŸ’¬ μ΅œλŒ€ν•œ λ§€λ„λŸ½κ²Œ ν•΄μ„ν•˜κ³ μž λ…Έλ ₯ν–ˆμ§€λ§Œ μ–΄μƒ‰ν•œ λ¬Έμž₯이 μžˆμ„ 수 μžˆμŠ΅λ‹ˆλ‹€. ν”Όλ“œλ°±μ€ μ–Έμ œλ‚˜ ν™˜μ˜μž…λ‹ˆλ‹€ πŸ™‚ 원본 κΈ€ μ£Όμ†Œ : https://towardsdatascience.com/entropy-cross-entropy-and-kl-divergence-explained-b09cdae917a Entropy, Cross-Entropy, and KL-Divergence Explained! Let us try to understand the most widely used loss function — Cross-Entropy. towardsdatascience.com Cross-Entropy(log-loss라고 ν•˜κΈ°λ„ 함)λŠ” λΆ„λ₯˜ λ¬Έμ œμ—μ„œ κ°€μž₯ 많이 μ“°μ΄λŠ” loss funct..

Archive 2022.03.31

[λ²ˆμ—­] Attention: Sequence 2 Sequence model with Attention Mechanism

μ—¬λŸ¬ μ‚¬μ΄νŠΈμ— 흩어져 μžˆλŠ” 글을 λͺ¨μœΌλ©΄ μ–΄λ €μš΄ κ°œλ…λ“€μ„ μ™„λ²½νžˆ 이해할 수 μžˆλ‹€κ³  λ―ΏμœΌλ©΄μ„œ μ‹œμž‘ν•œ 자체 μ½˜ν…μΈ  'Medium λ²ˆμ—­' πŸ’¬ μ΅œλŒ€ν•œ λ§€λ„λŸ½κ²Œ ν•΄μ„ν•˜κ³ μž λ…Έλ ₯ν–ˆμ§€λ§Œ μ–΄μƒ‰ν•œ λ¬Έμž₯이 μžˆμ„ 수 μžˆμŠ΅λ‹ˆλ‹€. ν”Όλ“œλ°±μ€ μ–Έμ œλ‚˜ ν™˜μ˜μž…λ‹ˆλ‹€ πŸ™‚ 원본 κΈ€ μ£Όμ†Œ : https://towardsdatascience.com/sequence-2-sequence-model-with-attention-mechanism-9e9ca2a613a Sequence 2 Sequence model with Attention Mechanism Detailed explanation about Attention mechanism in a sequence 2 sequence model suggested by Bahdanau and Luong t..

Archive 2022.03.16