.%20Same%20paths%20and%20fill;%20only%20the%20viewBox%20is%20cropped%20to%20the%20file's%20own%20clip%20rect%20so%20the%20bar%20sets%20the%20clear%20space.%20--%3e%3csvg%20width='480.615'%20height='97.3693'%20viewBox='97%2078%20480.615%2097.3693'%20fill='none'%20xmlns='http://www.w3.org/2000/svg'%3e%3cg%20clip-path='url(%23clip0_1709_25672)'%3e%3cpath%20d='M507.64%20157.772C497.197%20157.772%20490.907%20152.963%20490.907%20141.818V78H503.605V97.3565H530.187L547.394%20137.712L562.584%2097.3565H577.062L544.784%20175.369H530.424L540.63%20151.907L521.406%20108.501H503.605V139.941C503.605%20144.751%20505.741%20146.628%20510.725%20146.628H522.236V157.772H507.64Z'%20fill='%23121516'/%3e%3cpath%20d='M467.845%20157.772V97.3565H480.543V157.772H467.845ZM467.489%2091.139V78H480.899V91.139H467.489Z'%20fill='%23121516'/%3e%3cpath%20d='M404.356%20157.772L425.005%20126.215L405.899%2097.3566H421.089L433.549%20118.004L446.128%2097.3566H461.318L442.212%20126.215L462.861%20157.772H447.671L433.549%20134.779L419.546%20157.772H404.356Z'%20fill='%23121516'/%3e%3cpath%20d='M377.834%20159.532C359.915%20159.532%20348.285%20146.745%20348.285%20127.271C348.285%20108.501%20360.034%2095.5968%20377.122%2095.5968C394.329%2095.5968%20405.84%20107.445%20405.84%20125.16C405.84%20127.858%20405.722%20129.852%20405.366%20131.494H361.458C362.17%20142.874%20367.747%20148.974%20377.953%20148.974C385.31%20148.974%20390.176%20145.689%20391.956%20139.237H404.891C402.043%20151.907%20391.837%20159.532%20377.834%20159.532ZM361.458%20121.64H392.905C392.193%20111.317%20386.853%20106.155%20377.122%20106.155C367.984%20106.155%20362.288%20111.903%20361.458%20121.64Z'%20fill='%23121516'/%3e%3cpath%20d='M328.431%20157.772V78H341.128V157.772H328.431Z'%20fill='%23121516'/%3e%3cpath%20d='M263.88%20175.369V97.3565H276.578V103.457C280.375%2098.4123%20287.021%2095.5968%20294.853%2095.5968C311.111%2095.5968%20321.316%20107.915%20321.316%20127.506C321.316%20147.097%20311.111%20159.532%20294.853%20159.532C287.021%20159.532%20280.375%20156.717%20276.578%20151.672V175.369H263.88ZM292.123%20148.035C302.685%20148.035%20308.263%20140.997%20308.263%20127.506C308.263%20114.132%20302.685%20107.093%20292.123%20107.093C281.443%20107.093%20275.866%20114.132%20275.866%20127.506C275.866%20140.997%20281.443%20148.035%20292.123%20148.035Z'%20fill='%23121516'/%3e%3cpath%20d='M224.479%20157.772V97.3566H237.176V103.809C239.55%2099.7028%20244.178%2097.3566%20249.874%2097.3566H258.181V109.322H249.874C241.092%20109.322%20237.176%20113.546%20237.176%20122.931V157.772H224.479Z'%20fill='%23121516'/%3e%3cpath%20d='M188.82%20159.532C170.9%20159.532%20159.271%20146.745%20159.271%20127.271C159.271%20108.501%20171.019%2095.5968%20188.107%2095.5968C205.315%2095.5968%20216.825%20107.445%20216.825%20125.16C216.825%20127.858%20216.706%20129.852%20216.35%20131.494H172.443C173.155%20142.874%20178.733%20148.974%20188.938%20148.974C196.296%20148.974%20201.161%20145.689%20202.941%20139.237H215.876C213.028%20151.907%20202.823%20159.532%20188.82%20159.532ZM172.443%20121.64H203.891C203.179%20111.317%20197.838%20106.155%20188.107%20106.155C178.97%20106.155%20173.274%20111.903%20172.443%20121.64Z'%20fill='%23121516'/%3e%3cpath%20d='M97%20175.369V97.3565H109.698V103.457C113.495%2098.4123%20120.141%2095.5968%20127.973%2095.5968C144.231%2095.5968%20154.436%20107.915%20154.436%20127.506C154.436%20147.097%20144.231%20159.532%20127.973%20159.532C120.141%20159.532%20113.495%20156.717%20109.698%20151.672V175.369H97ZM125.244%20148.035C135.805%20148.035%20141.383%20140.997%20141.383%20127.506C141.383%20114.132%20135.805%20107.093%20125.244%20107.093C114.563%20107.093%20108.986%20114.132%20108.986%20127.506C108.986%20140.997%20114.563%20148.035%20125.244%20148.035Z'%20fill='%23121516'/%3e%3c/g%3e%3cdefs%3e%3cclipPath%20id='clip0_1709_25672'%3e%3crect%20width='480.615'%20height='97.3693'%20fill='white'%20transform='translate(97%2078)'/%3e%3c/clipPath%3e%3c/defs%3e%3c/svg%3e)
Cas d'usageRecherche
Comprendre les méthodologies d'évaluation de l'IA
Utilisez la Recherche approfondie pour décoder les benchmarks de LLM : leur fonctionnement, leurs limites et les métriques à privilégier pour l'évaluation des modèles
- Recherche
- Recherche approfondie
- Tech
Aperçu
Les chercheurs en IA, les ingénieurs et les responsables de produits peuvent utiliser Perplexity pour comprendre le paysage en rapid évolution de l'évaluation des LLM. La Recherche approfondie puise dans des documents académiques, la documentation des benchmarks et des analyses de l'industrie pour expliquer non seulement le fonctionnement des benchmarks, mais aussi leurs lacunes, aidant les équipes à prendre de meilleures décisions sur les métriques à privilégier lors de l'évaluation des modèles.
Astuce
- Partagez votre contexte de production spécifique (par exemple, pipeline RAG, assistant de codage, assistant de support) afin que Perplexity puisse associer ces benchmarks à la suite d'évaluation que vous souhaitez réellement créer
- Renseignez-vous sur les nouvelles approches d'évaluation en cours de développement pour remédier à ces limitations structurelles dans un contexte prospectif