Using LLMsUtiliser les GML
Choosing and comparing models, what chat products offer, prompting, and checking what comes back.Choisir et comparer les modèles, ce qu'offrent les agents conversationnels, formuler ses requêtes et vérifier ce qui en ressort.
- What to expectÀ quoi s'attendre
This module is the practical one. It explains which model to open for a given task and how to reach it, how to read the leaderboards that rank models, and what a chat product adds around the model. It then covers how to write a prompt that gets followed, and how to work with long documents, corrections and standing instructions. It ends with how to check an answer, instructions hidden in a document the model reads for you, and where your text goes when you send it.This module is the practical one. It explains which model to open for a given task and how to reach it, how to read the leaderboards that rank models, and what a chat product adds around the model. It then covers how to write a prompt that gets followed, and how to work with long documents, corrections and standing instructions. It ends with how to check an answer, instructions hidden in a document the model reads for you, and where your text goes when you send it.
- PrerequisitesPréalables
The LLM lifecycle module, especially its chapters on inference and on in-context learning: this module applies what those explain and does not repeat it. The chapters on the jagged frontier and on reasoning models in the capabilities module help with choosing a model.The LLM lifecycle module, especially its chapters on inference and on in-context learning: this module applies what those explain and does not repeat it. The chapters on the jagged frontier and on reasoning models in the capabilities module help with choosing a model.
- StructureStructure
The first chapter is about choosing: model families and tiers, how much effort, the three ways of reaching a model, how to match a task to a model, and how to keep a small test set of your own. The second explains how to read the leaderboards that rank models. The third is about what chat products add around the model: search, files and memory, code execution and connectors, deep research and agents, shared documents, images, voice and study modes. The fourth is about prompting: what still works, examples, the language of the prompt, long documents, correcting or starting over, and standing instructions and memory. The last is about risks and safeguards: checking an answer, instructions hidden in a document the model reads for you, and where your own text goes.The first chapter is about choosing: model families and tiers, how much effort, the three ways of reaching a model, how to match a task to a model, and how to keep a small test set of your own. The second explains how to read the leaderboards that rank models. The third is about what chat products add around the model: search, files and memory, code execution and connectors, deep research and agents, shared documents, images, voice and study modes. The fourth is about prompting: what still works, examples, the language of the prompt, long documents, correcting or starting over, and standing instructions and memory. The last is about risks and safeguards: checking an answer, instructions hidden in a document the model reads for you, and where your own text goes.
Choosing a modelChoisir un modèle
- Families, generations and tiersFamilles, générations et niveaux TRA6008 S3 · in classS3 · en classe
- Execution “effort”L'« effort » d'exécution TRA6008 S3 · in classS3 · en classe
- Chat app, API or your own computerApplication, API ou votre propre ordinateur TRA6008 S3 · in classS3 · en classe
- Which model for which taskQuel modèle pour quelle tâche TRA6008 S3 · in classS3 · en classe
- Testing it yourselfFaire vos propres tests TRA6008 OptionalFacultatif
- RelatedEn lien
- LLM leaderboardsLes classements de GML · InferenceInférence · The jagged frontierLa frontière irrégulière · Reasoning modelsModèles de raisonnement · Tools and agentsOutils et agents
3 concepts introducedconcepts introduits
application programming interfaceinterface de programmation d'application open-weight modelmodèle à poids ouverts reasoning efforteffort de raisonnement
LLM leaderboardsLes classements de GML
- Reading a leaderboardLire un classement
- Human preference rankingsLes classements fondés sur les préférences humaines
- Benchmarks in French and other languagesLes tests en français et dans d'autres langues
- Benchmarks for translationLes tests pour la traduction
- Builds onPréalables
- Choosing a modelChoisir un modèle
- RelatedEn lien
- Evaluating machine translationÉvaluation de la traduction automatique · The information ecosystemL'écosystème informationnel · Evaluation foundationsFondements de l'évaluation
4 concepts introducedconcepts introduits
human preference evaluationévaluation fondée sur les préférences humaines benchmarktest de performances data contaminationcontamination des données shared tasktâche partagée
What chat products offerCe qu'offrent les agents conversationnels
- What a chat product adds to the modelCe qu'un agent conversationnel ajoute au modèle TRA6008 OptionalFacultatif
- Code execution and connectorsL'exécution de code et les connecteurs TRA6008 OptionalFacultatif
- Deep researchLa recherche approfondie TRA6008 OptionalFacultatif
- AgentsLes agents TRA6008 OptionalFacultatif
- Documents, images, voice and study modesDocuments, images, voix et modes d'étude TRA6008 OptionalFacultatif
- Builds onPréalables
- Choosing a modelChoisir un modèle
- RelatedEn lien
- PromptingFormulation de requêtes · Risks and safeguardsRisques et précautions · Tools and agentsOutils et agents
PromptingFormulation de requêtes
- Prompts and prompt engineeringRequêtes et rédactique TRA6008 S3 · in classS3 · en classe
- Be specific and ask for enoughÊtre précis et en demander assez TRA6008 S3 · in classS3 · en classe
- Split the taskDécouper la tâche TRA6008 S3 · in classS3 · en classe
- Structure the promptStructurer la requête TRA6008 S3 · in classS3 · en classe
- Giving examplesDonner des exemples TRA6008 S3 · in classS3 · en classe
- The language of the promptLa langue de la requête TRA6008 S3 · in classS3 · en classe
- Working with long documentsTravailler avec de longs documents TRA6008 OptionalFacultatif
- Correcting the model or starting overCorriger le modèle ou recommencer TRA6008 S3 · in classS3 · en classe
- Improving a promptAméliorer une requête TRA6008 OptionalFacultatif
- Custom instructions, memory, projects and skillsInstructions personnalisées, mémoire, projets et compétences TRA6008 OptionalFacultatif
- Builds onPréalables
- In-context learningApprentissage en contexte · InferenceInférence
- RelatedEn lien
- Choosing a modelChoisir un modèle · Risks and safeguardsRisques et précautions
8 concepts introducedconcepts introduits
promptrequête prompt engineeringrédactique zero-shot promptingformulation de requête sans exemple few-shot promptingformulation de requête avec quelques exemples structured outputsortie structurée automatic prompt optimizationoptimisation automatique de requêtes custom instructionsinstructions personnalisées persistent memorymémoire persistante
Risks and safeguardsRisques et précautions
- Checking the answerVérifier la réponse TRA6008 S3 · in classS3 · en classe
- Prompt injectionL'infiltration de requête TRA6008 OptionalFacultatif
- Training, retention and privacy settingsEntraînement, conservation et paramètres de confidentialité TRA6008 S3 · in classS3 · en classe
- RelatedEn lien
- HallucinationsLes fabulations de l'IA · AlignmentAlignement · Tools and agentsOutils et agents · Privacy and confidentialityVie privée et confidentialité · PromptingFormulation de requêtes