content.dataiku.com
Open in
urlscan Pro
18.235.231.235
Public Scan
Submitted URL: https://pages.dataiku.com/e3t/Ctc/GA+113/cfvmy04/VW_gYs5jMn3WVK-7_m3xTlPxW7swsJQ5fqDN0N56dXgq5nR32W50kH_H6lZ3q5W7Y6PD16DFY...
Effective URL: https://content.dataiku.com/the-llm-cheatsheet-bundle?utm_campaign=GLO+CONTENT+DB+Emails+Unengaged+2024&utm_medium=email&_hs...
Submission: On May 23 via api from IN — Scanned from DE
Effective URL: https://content.dataiku.com/the-llm-cheatsheet-bundle?utm_campaign=GLO+CONTENT+DB+Emails+Unengaged+2024&utm_medium=email&_hs...
Submission: On May 23 via api from IN — Scanned from DE
Form analysis
0 forms found in the DOMText Content
Miniaturansichten Dokumentstruktur Anhänge Ebenen Aktuelles Struktur-Element Definitions Transformers Types of LLMs Configuration settings Zurück Weiter Alle hervorheben Groß-/Kleinschreibung beachten Akzente Ganze Wörter Farbe Größe Farbe Dicke Deckkraft Präsentationsmodus Öffnen Drucken Speichern Aktuelle Ansicht Erste Seite anzeigen Letzte Seite anzeigen Im Uhrzeigersinn drehen Gegen Uhrzeigersinn drehen Textauswahl-Werkzeug Hand-Werkzeug Einzelseitenanordnung Vertikale Seitenanordnung Horizontale Seitenanordnung Kombinierte Seitenanordnung Einzelne Seiten Ungerade + gerade Seite Gerade + ungerade Seite Dokumenteigenschaften… Sidebar umschalten Suchen Zurück Vor von 1 Präsentationsmodus Öffnen Drucken Speichern Aktuelle Ansicht FreeText-Annotation Ink-Annotation Werkzeuge Verkleinern Vergrößern Automatischer Zoom Originalgröße Seitengröße Seitenbreite 50 % 75 % 100 % 125 % 150 % 200 % 300 % 400 % DEFINITIONS Generative AI AI systems that can produce realistic content (text, image, etc.) Large Language Models ( LLMs) Large neural networks trained at internet scale to estimate the probability of sequences of words Ex: GPT, FLAN-T5, LLaMA, PaLM, BLOOM (transformers with billions of parameters) Abilities (and computing resources needed) tend to rise with the number of parameters USE CASES – Standard NLP tasks (classification, summarization, etc.) – Content generation – Reasoning (Q&A, planning, coding, etc.) In-context learning Specifying the task to perform directly in the prompt Introduction to LLMs TRANSFORMERS – Can scale efficiently to use multi-core GPUs – Can process input data in parallel – Pay attention to all other words when processing a word Transformers’ strength lies in understanding the context and relevance of all words in a sentence Token Word or sub-word The basic unit processed by transformers Encoder Processes input sequence to generate a vector representation (or embedding) for each token Decoder Processes input tokens to produce new tokens Embedding layer Maps each token to a trainable vector Positional encoding vector Added to the token embedding vector to keep track of the token’s position Self-Attention Computes the importance of each word in the input sequence to all other words in the sequence TYPES OF LLMS Encoder only = Autoencoding model Ex: BERT, RoBERTa These are not generative models. PRE-TRAINING OBJECTIVE To predict tokens masked in a sentence (= Masked Language Modeling) OUTPUT Encoded representation of the text USE CASE(S) Sentence classification (e.g., NER) Decoder only = Autoregressive model Ex: GPT, BLOOM PRE-TRAINING OBJECTIVE To predict the next token based on the previous sequence of tokens (= Causal Language Modeling) OUTPUT Next token USE CASES Text generation Encoder-Decoder = Seq-to-seq model Ex: T5, BART PRE-TRAINING OBJECTIVE Vary from model to model (e.g., Span corruption like T5) OUTPUT Sentinel token + predicted tokens USE CASES Translation, Q&A, summarization CONFIGURATION SETTINGS Parameters to set at inference time Max new tokens Maximum number of tokens generated during completion Decoding strategy 1 Greedy Decoding The word/token with the highest probability is selected from the final probability distribution (prone to repetition) 2 Random Sampling The model chooses an output word at random using the probability distribution to weigh the selection (could be too creative) TECHNIQUES TO CONTROL RANDOM SAMPLING – Top K The next token is drawn from the k tokens with the highest probabilities – Top P The next token is drawn from the tokens with the highest probabilities, whose combined probabilities exceed p Temperature Influence the shape of the probability distribution through a scaling factor in the softmax layer © 2024 Dataiku Mehr Informationen Weniger Informationen Schließen Geben Sie zum Öffnen der PDF-Datei deren Passwort ein. Abbrechen OK Dateiname: - Dateigröße: - Titel: - Autor: - Thema: - Stichwörter: - Erstelldatum: - Bearbeitungsdatum: - Anwendung: - PDF erstellt mit: - PDF-Version: - Seitenzahl: - Seitengröße: - Schnelle Webanzeige: - Schließen Dokument wird für Drucken vorbereitet… 0 % Abbrechen Next Next Cheatsheet: LLM-Powered Applications In this cheatsheet, discover insights on model optimization for deployment, LLM-integrated applications, LLM reasoning, program-aided language and more. LinkedIn LinkTwitter LinkFacebook LinkEmail LinkLike Button EXPLORE USE CASE LIBRARY pdf:Cheatsheet: Introduction to LLMs pdf:Cheatsheet: LLM-Powered Applications pdf:Cheatsheet: LLM Compute Challenges and Scaling Laws pdf:Cheatsheet: Parameter Efficient Fine-Tuning (PEFT) Methods pdf:Cheatsheet: LLM-Instruction Fine-Tuning & Evaluation pdf:Cheatsheet: LLM Preference Fine-Tuning (Part 1) pdf:Cheatsheet: LLM Preference Fine-Tuning (Part 2) GET MORE CONTENT FROM DATAIKU! Sign up for our newsletter for exclusive updates on just-released content, Dataiku product announcements, and more