A video reviewing lmsys chatbot arena (voice in Bulgarian) and analyzing the results from various models: why smaller models sometimes are comparable or better with the big ones?
https://youtu.be/goJoxzTda-M
LMSys Chatbot Arena: страхотна находка за безплатно ползване на най-големите езикови модели, които иначе са платени.
LMSys Chatbot Arena: a great find for free to use the biggest language models that are otherwise paid. A variety of very big and smaller ones to try out with real job and compare the results.
| E.g. (Source: https://chat.lmsys.org/ - many o f the models come in various versions such as Claude, GPT4-Turbo, LLAMA-3 Command-R and Command-R-Plus, Reka-Flash ... The ranking page lists 82 at the moment and they are probably updated on the fly. Edit 20.4.2024: the latest LLAMA 3 70B has just appeared:  https://chat.lmsys.org/?leaderboard | ||
|---|---|---|
| Claude: Claude by Anthropic | Mixtral of experts: A Mixture-of-Experts model by Mistral AI | Reka Flash: Multimodal model by Reka |
| Command-R-Plus: Command-R Plus by Cohere | Command-R: Command-R by Cohere | Zephyr 141B-A35B: ORPO fine-tuned of Mixtral-8x22B-v0.1 |
| Gemma: Gemma by Google | Qwen 1.5: A large language model by Alibaba Cloud | DBRX Instruct: DBRX by Databricks Mosaic AI |
| Starling-LM-7B: An open model trained using RLAIF by Berkeley | Gemini: Gemini by Google | GPT-4-Turbo: GPT-4-Turbo by OpenAI |
| GPT-3.5: GPT-3.5-Turbo by OpenAI | Llama 2: Open foundation and fine-tuned chat models by Meta | OLMo-7B: OLMo by Allen AI |
| Vicuna: A chat assistant fine-tuned on user-shared conversations by LMSYS | Yi-Chat: A large language model by 01 AI | Code Llama: Open foundation models for code by Meta |
| DeepSeek LLM: An advanced language model by DeepSeek | OpenHermes-2.5-Mistral-7B: A mistral-based model fine-tuned on 1M GPT-4 outputs |
0:00 Introduction: Why is INSAIT's BGGPT really comparable in quality to Bulgarian GPT4 for some tasks, as claimed by INSAIT? Who are the ten most... athletes, musicians, ... 0:30 What is LMSYS Chatbot Arena? 2:00 Why are models widely thought of as GPT4 and why do they produce similar results? ... GPT4-turbo-2024-04-09 vs qwen1.5-32B -chat 3:40 Suggestions. You are ... German ... born ... what do you think about ... 7:40 Reflections: fruits and vegetables. Reasoning. 9:37 Command-R vs. Command-R-plus: 35B parameters, 104B; RAG 11:01 An example of intelligence on large models with 35 and 104 billion parameters and a context of 128 thousand tokens 12:00 You are German... 13:44 "new language" test 14:55 Reflections: claude-3-sonnet-20240229 vs GPT4-turbo-2024-04-0 9 ... Claude-3: 200k letters of context! 16:55 Complex systems: non-monolithic systems, Mixture of Experts ... Pedro Domingos: Ensemble learning, Five tribes in Machine Learning 17:50 The most unknown athletes, musicians, truck drivers in Bulgaria ... 19:10 Scandals in Bulgarian politics with Boyko Borisov, Kiril Petkov, Asen Vasilev, Hristo Ivanov, GERB, PP-DB, Democratic Bulgaria... 2 22:00 Announcement for the deepfake movie "Lord of Bulgaria 2". The first part: https://youtu.be/VPj9L61R_Ak 23:09 Can it be programmed? ... 25:40 Summary Watch the other videos on working with BGGPT and GPT2 in Google Colab and the upcoming ones where we continue with other models, tests and more complex processing. Be sure to support "THE SACRED COUNTER" if you add: subscribe, share, comment, like! We also accept donations of computer equipment, disk memory, video cards. thanks! Изпращане на отзиви Странични панели Налице са резултати за превод
0:00 Въведение: Защо наистина BGGPT на INSAIT е сравним по качество с GPT4 на български за някои задачи, както твърдят от INSAIT? Кои са десетте най-... спортисти, музиканти, ... 0:30 Какво е LMSYS Chatbot Arena? 2:00 Защо моделите масово се мислят за GPT4 и защо пораждат подобни резултати? ... GPT4-turbo-2024-04-09 vs qwen1.5-32B -chat 3:40 Внушения. Ти си ... германец ... роден ... какво мислиш за ... 7:40 Разсъждения: плодове и зеленчуци. Reasoning. 9:37 Command-R vs Command-R-plus: 35B params, 104B; RAG 11:01 Пример за интелигентността на големи модели с 35 и 104 милиарда параметъра и контекст от 128 хил. токена 12:00 Ти си немец... 13:44 Тест на "нов език" 14:55 Разсъждения: claude-3-sonnet-20240229 vs GPT4-turbo-2024-04-0 9 ... Клод-3: 200 хил. буквачета контекст! 16:55 Сложни системи: немонолитни системи, Mixture of Experts ... Педро Домингос (Pedro Domingos): Ensemble learning, Five tribes in Machine Learning 17:50 Най-неизвестни спортисти, музканти, шофьори на камиони в България ... 19:10 Скандали в българската политика с Бойко Борисов, Кирил Петков, Асен Василев, Христо Иванов, ГЕРБ, ПП-ДБ, Демоктратична България ... 2 22:00 Анонс за дийпфейк филма "Властелинът на България 2". Първата част: https://youtu.be/VPj9L61R_Ak 23:09 Дали могат да програмират? ... 25:40 Обобщение Гледайте и другите клипове за работа с BGGPT и GPT2 в Google Colab и предстоящите, в които продължаваме с други модели, тестове и по-сложни обработки. Не забравяйте да подкрепяте "СВЕЩЕНИЯТ СМЕТАЧ", ако ви допада: абонирайте, споделяйте, коментирайте, харесвайте! Приемаме дарения и на компютърна техника, дискова памет, видеокарти. Благодаря! https://github.com/twenkid/ https://artificial-mind.blogspot.com/ https://github.com/Twenkid/Vsy-Jack-Of-All-Trades-AGI-Bulgarian-Internet-Archive-And-Search-Engine
0 коментара:
Post a Comment