{"id":114737,"date":"2024-04-03T00:25:44","date_gmt":"2024-04-02T22:25:47","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/databricks-otkryl-bolshuyu-yazykovuyu-model-dbrx-operezhayushhuyu-v-testah-gpt-3-5"},"modified":"2024-04-03T00:25:44","modified_gmt":"2024-04-02T22:25:47","slug":"databricks-otkryl-bolshuyu-yazykovuyu-model-dbrx-operezhayushhuyu-v-testah-gpt-3-5","status":"publish","type":"post","link":"https:\/\/prohoster.info\/da\/blog\/news\/databricks-otkryl-bolshuyu-yazykovuyu-model-dbrx-operezhayushhuyu-v-testah-gpt-3-5","title":{"rendered":"Databricks har \u00e5bnet den store sprogmodel DBRX, der overg\u00e5r GPT-3.5 i test","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"
Databricks har annonceret lanceringen af den store sprogmodel DBRX, der kan bruges til at skabe chatbots, der besvarer sp\u00f8rgsm\u00e5l p\u00e5 naturligt sprog, l\u00f8ser foresl\u00e5ede matematiske opgaver og kan generere indhold om et givet emne samt skabe kode p\u00e5 forskellige programmeringssprog. Modellen er udviklet af Mosaic ML, som blev k\u00f8bt af Databricks for 1,3 milliarder dollars. Til tr\u00e6ning blev der anvendt en klynge best\u00e5ende af 3072 NVIDIA H100 Tensor Core GPU'er. Det anbefales at have 320GB hukommelse til at k\u00f8re den f\u00e6rdige model. <\/p>\n
Ved tr\u00e6ningen af modellen blev arkitekturen MoE (Mixture of Experts) anvendt, hvilket muligg\u00f8r en mere pr\u00e6cis ekspertvurdering, samt en samling af tekster og kode med en st\u00f8rrelse p\u00e5 12 TB. Den m\u00e6ngde kontekst, som modellen DBRX tager i betragtning, er 32.000 tokens (antal tokens, som modellen kan bearbejde og huske under tekstgenerering). Til sammenligning er konteksten for modellerne Google Gemini og OpenAI GPT-4 32.000 tokens, Google Gemma \u2014 8.000, mens modellen GPT-4 Turbo har 128.000 tokens. <\/p>\n
Modellen d\u00e6kker 132 milliarder parametre og er opdelt i 16 ekspert-netv\u00e6rk, hvoraf der maksimalt kan anvendes 4 (d\u00e6kker op til 36 milliarder parametre pr. token) ved behandling af en foresp\u00f8rgsel. Til sammenligning ansl\u00e5s modellen GPT-4 at inkludere 1,76 trillioner parametre, den nyligt lancerede X\/Twitter-model Grok (X\/Twitter) \u2014 314 milliarder, GPT-3.5 \u2014 175 milliarder, YaLM (Yandex) \u2014 100 milliarder, LLaMA (Meta) \u2014 65 milliarder, GigaChat (Sber) \u2014 29 milliarder, Gemma (Google) \u2014 7 milliarder. <\/p>\n
Modellen og de tilknyttede komponenter distribueres under Databricks Open Model License, som tillader brug, reproduktion, kopiering, \u00e6ndring og oprettelse af derivative produkter, men med visse begr\u00e6nsninger. For eksempel forbyder licensen brugen af DBRX, derivative modeller og enhver output baseret p\u00e5 dem til at forbedre andre sprogmodeller, der ikke er DBRX. Licensen forbyder ogs\u00e5 brugen af modellen inden for omr\u00e5der, der overtr\u00e6der love og reguleringer. Derivative modeller skal distribueres under den samme licens. Ved brug i produkter og tjenester, der benyttes af over 700 millioner brugere om m\u00e5neden, kr\u00e6ves der en s\u00e6rskilt tilladelse. <\/p>\n
If the creators of the model are to be believed, the characteristics and capabilities of DBRX surpass those of OpenAI's GPT-3.5 and Twitter's Grok-1, and it can compete with the Gemini 1.0 Pro model when testing language comprehension, coding abilities, and solving mathematical problems. In some applications, such as generating SQL queries, DBRX approaches the efficiency of the GPT-4 Turbo model, which leads the market. Furthermore, the model distinguishes itself from competing services with its extremely fast operation and allows for nearly instantaneous responses. In particular, DBRX can generate text at a speed of up to 150 tokens per second per user, which is about twice as fast as the LLaMA2-70B model.
<\/center>
<\/center> <\/p>\n
Additionally, it is worth noting the publication of the technical description of the open large language model InternLM2, which is distributed under the Apache 2.0 license and is available in versions with 20, 7, and 1.8 billion parameters. The model is being developed by a Shanghai AI laboratory with the participation of several Chinese universities and is noteworthy for its ability to handle up to 200K tokens of context and support not only English but also Chinese. In many tests, the model is close to GPT-4.
<\/center>
<\/center>
<\/center> <\/p>\n
Moreover, there are reports of the development of 84 new matrix multiplication cores for the llamafile toolkit, developed by Mozilla, which allows the creation of universal executable files for running large machine learning language models (LLM). The changes have significantly accelerated model performance in llamafile when executed on CPU. For example, model execution using llamafile is now faster than using llama.cpp by 30% to 500%, depending on the environment, and compared to the MKL library, matrix operations that fit into the L2 cache are executed twice as fast in the new implementation. \n Kilde: opennet.ru<\/a> <\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"
![\"Databricks](\"\/wp-content\/uploads\/2024\/04\/d46e64df9ccd0933e38912f561c6d137.png\")
<\/center> ![\"Databricks](\"\/wp-content\/uploads\/2024\/04\/ab1bfbed106ba989fc5732d1424d30c5.png\")
<\/center> <\/p>\n![\"Databricks](\"\/wp-content\/uploads\/2024\/04\/a8e99d24da8c97b82707305694061bb6.png\")
<\/center> ![\"Databricks](\"\/wp-content\/uploads\/2024\/04\/a4b4d1a46b99ac7872a4dfd7bcf117bc.png\")
<\/center> ![\"Databricks](\"\/wp-content\/uploads\/2024\/04\/1ee229046c62e435a286b871f08e3de4.png\")
<\/center> <\/p>\n