Характеристики

Контекст: 65 536 токенов
Параметры: 22 млрд
Размер весов: 281 ГБ
Лицензия: Apache 2.0
Архитектура: mixtral · MoE
Базовая модель: Mixtral-8x22B-v0.1
Опубликована: 29 апр 2024

Ссылки

Hugging Face

dolphin-2.9-mixtral-8x22b

dphn

Генерация текстаТекст → текст15 профессий

Дообученная на базе Mixtral 8x22B диалоговая модель для инструкций, программирования, математики и вызова функций.

223скачиваний за 30 дней26лайковвеса обновлены20 мая 2024

Бенчмарки

Как модель справилась с задачами пользователей ailist и что о ней публикуют открытые лидерборды

Тесты пользователейпока нет

Никто ещё не опубликовал тест этой модели. Соберите бенчмарк из своих задач — результат появится здесь.

Добавьте модель в бенчмарк
Вместе с другими моделями категории «Текст» — сравнение идёт внутри одной категории.
Опишите свои задачи
Напишите запросы, при необходимости приложите файлы и эталонные ответы.
Запустите и опубликуйте
Модели ответят на каждую задачу, судья оценит ответы. Опубликованный результат появится на этой странице.

Открыть конструктор бенчмарка

Публичные оценки

Из открытых лидербордов. Сравнивайте результаты одного теста при одинаковых условиях — отсутствие оценки не значит, что модель хуже.

В подключённых источниках результатов для этой модели пока нет.

Что умеет модель

Описание составлено по карточке модели и метаданным Hugging Face

Dolphin 2.9 Mixtral 8x22B — диалоговая модель Cognitive Computations, подготовленная Eric Hartford, Lucas Atkins и Fernando Fernandes на основе mistral-community/Mixtral-8x22B-v0.1. Это дообученная модель с разреженной архитектурой Mixtral MoE, представленная в формате safetensors и предназначенная для генерации текста по инструкциям и ведения диалогов.

Модель подходит для ответов на вопросы, программирования, математических задач и сценариев с вызовом функций; README также отмечает начальные агентные возможности. Она обучалась в формате ChatML, а полное дообучение проводилось на последовательностях длиной 4 тыс. токенов. Для базовой модели заявлен контекст до 64 тыс. токенов. В метаданных указан английский язык.

Dolphin описана авторами как неконтролируемая модель, которая может охотно выполнять в том числе неэтичные запросы; перед предоставлением пользователям сервиса авторы рекомендуют добавить собственный слой ограничений. В данных обучения обнаружена проблема, из-за которой модель может излишне упоминать системное сообщение; README предлагает явно просить её этого не делать. Лицензия Apache-2.0 допускает коммерческое использование при соблюдении её условий.

Сценарии применения

ответы на вопросы и диалог по инструкциям
генерация и объяснение кода
решение математических задач
создание чат-ассистента с вызовом функций
создание агентных прототипов

Ограничения

Модель может выполнять вредоносные и неэтичные запросы, поэтому для сервиса авторы рекомендуют собственный слой ограничений. Из-за дефекта данных обучения она может излишне говорить о системном сообщении; дообучение выполнялось на последовательностях длиной 4 тыс. токенов, хотя базовая модель поддерживает контекст до 64 тыс.

Для каких профессий подходит15

Разделы каталога профессий, в задачах которых используется эта модель

Все профессии

II.2Разработка и данныеСоздание и понимание программного кодаAI помогает писать, объяснять и связывать программный код с требованиями и существующей системой.20 задач21 443модели II.6Разработка и данныеДокументация и сопровождение разработкиAI помогает поддерживать понятную техническую документацию и ускорять передачу знаний.18 задач38 015моделей II.10Разработка и данныеРазработка и эксплуатация AI/ML-системAI помогает создавать, оценивать, развёртывать и контролировать AI/ML-системы на всём жизненном цикле.30 задач49 000моделей II.8Разработка и данныеИнженерия данных и аналитикаAI помогает готовить надёжные данные и превращать их в понятную аналитику.22 задачи23 488моделей

Также подходит для

Карточка модели

README от автора на Hugging Face, как он его опубликовал

Открыть на Hugging Face

Dolphin 2.9 Mixtral 8x22b 🐬

Curated and trained by Eric Hartford, Lucas Atkins, and Fernando Fernandes, and Cognitive Computations

Discord: https://discord.gg/cognitivecomputations

A bug has been found in the Dolphin 2.9 dataset in SystemConversations that causes the model to overly talk about the "SYSTEM MESSAGE". To counter this, we recommend you add a statement in the system message directing the model not to mention the system message. An example system message is "The assistant is named Dolphin. A helpful and friendly AI assistant, Dolphin avoids discussing the system message unless directly asked about it."

My appreciation for the sponsors of Dolphin 2.9:

Crusoe Cloud - provided excellent on-demand 8xH100 node

This model is based on Dolphin-2.9-Mixtral-8x22b, and is Apache-2.0 licensed.

The base model has 64k context, and the full-weight fine-tuning was with 4k sequence length.

It took 1 week on 8xH100 provided by Crusoe Cloud

This model was trained FFT on 50% parameters (targeted with Laser Scanner by Fernando Fernandes, David Golchinfar, Lucas Atkins, and Eric Hartford) , using ChatML prompt template format.

example:

<|im_start|>system
You are Dolphin, a helpful AI assistant.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

Dolphin-2.9 has a variety of instruction, conversational, and coding skills. It also has initial agentic abilities and supports function calling.

Dolphin is uncensored. I have filtered the dataset to remove alignment and bias. This makes the model more compliant. You are advised to implement your own alignment layer before exposing the model as a service. It will be highly compliant with any requests, even unethical ones. Please read my blog post about uncensored models. https://erichartford.com/uncensored-models You are responsible for any content you create using this model. Enjoy responsibly.

Dolphin is licensed Apache 2.0. I grant permission for any use, including commercial, that falls within accordance with Apache-2.0 license. Dolphin was trained on data generated from GPT4, among other models.

Evals

image/png

Training

axolotl version: 0.4.0

base_model: mistral-community/Mixtral-8x22B-v0.1
model_type: AutoModelForCausalLM
tokenizer_type: LlamaTokenizer
trust_remote_code: true

load_in_8bit: false
load_in_4bit: false
strict: false

unfrozen_parameters:
  - ^lm_head.weight$
  - ^model.embed_tokens.weight$
  - model.layers.0.self_attn.q_proj
  - model.layers.1.self_attn.q_proj
  - model.layers.2.self_attn.q_proj
  - model.layers.22.self_attn.q_proj
  - model.layers.27.self_attn.q_proj
  - model.layers.28.self_attn.q_proj
  - model.layers.13.self_attn.q_proj
  - model.layers.21.self_attn.q_proj
  - model.layers.24.self_attn.q_proj
  - model.layers.14.self_attn.q_proj
  - model.layers.15.self_attn.q_proj
  - model.layers.11.self_attn.q_proj
  - model.layers.20.self_attn.q_proj
  - model.layers.23.self_attn.q_proj
  - model.layers.30.self_attn.k_proj
  - model.layers.31.self_attn.k_proj
  - model.layers.25.self_attn.k_proj
  - model.layers.23.self_attn.k_proj
  - model.layers.27.self_attn.k_proj
  - model.layers.26.self_attn.k_proj
  - model.layers.29.self_attn.k_proj
  - model.layers.28.self_attn.k_proj
  - model.layers.24.self_attn.k_proj
  - model.layers.16.self_attn.k_proj
  - model.layers.19.self_attn.k_proj
  - model.layers.22.self_attn.k_proj
  - model.layers.20.self_attn.k_proj
  - model.layers.6.self_attn.k_proj
  - model.layers.22.self_attn.v_proj
  - model.layers.29.self_attn.v_proj
  - model.layers.31.self_attn.v_proj
  - model.layers.5.self_attn.v_proj
  - model.layers.8.self_attn.v_proj
  - model.layers.4.self_attn.v_proj
  - model.layers.25.self_attn.v_proj
  - model.layers.30.self_attn.v_proj
  - model.layers.17.self_attn.v_proj
  - model.layers.23.self_attn.v_proj
  - model.layers.28.self_attn.v_proj
  - model.layers.9.self_attn.v_proj
  - model.layers.26.self_attn.v_proj
  - model.layers.27.self_attn.v_proj
  - model.layers.20.self_attn.o_proj
  - model.layers.19.self_attn.o_proj
  - model.layers.16.self_attn.o_proj
  - model.layers.13.self_attn.o_proj
  - model.layers.18.self_attn.o_proj
  - model.layers.17.self_attn.o_proj
  - model.layers.12.self_attn.o_proj
  - model.layers.15.self_attn.o_proj
  - model.layers.14.self_attn.o_proj
  - model.layers.22.self_attn.o_proj
  - model.layers.23.self_attn.o_proj
  - model.layers.21.self_attn.o_proj
  - model.layers.10.self_attn.o_proj
  - model.layers.0.self_attn.o_proj
  - model.layers.0.block_sparse_moe.experts.0.w1
  - model.layers.1.block_sparse_moe.experts.0.w1
  - model.layers.2.block_sparse_moe.experts.0.w1
  - model.layers.3.block_sparse_moe.experts.0.w1
  - model.layers.4.block_sparse_moe.experts.0.w1
  - model.layers.5.block_sparse_moe.experts.0.w1
  - model.layers.6.block_sparse_moe.experts.0.w1
  - model.layers.7.block_sparse_moe.experts.0.w1
  - model.layers.8.block_sparse_moe.experts.0.w1
  - model.layers.9.block_sparse_moe.experts.0.w1
  - model.layers.10.block_sparse_moe.experts.0.w1
  - model.layers.11.block_sparse_moe.experts.0.w1
  - model.layers.12.block_sparse_moe.experts.0.w1
  - model.layers.13.block_sparse_moe.experts.0.w1
  - model.layers.0.block_sparse_moe.experts.0.w2
  - model.layers.1.block_sparse_moe.experts.0.w2
  - model.layers.2.block_sparse_moe.experts.0.w2
  - model.layers.3.block_sparse_moe.experts.0.w2
  - model.layers.4.block_sparse_moe.experts.0.w2
  - model.layers.5.block_sparse_moe.experts.0.w2
  - model.layers.6.block_sparse_moe.experts.0.w2
  - model.layers.7.block_sparse_moe.experts.0.w2
  - model.layers.8.block_sparse_moe.experts.0.w2
  - model.layers.9.block_sparse_moe.experts.0.w2
  - model.layers.10.block_sparse_moe.experts.0.w2
  - model.layers.11.block_sparse_moe.experts.0.w2
  - model.layers.12.block_sparse_moe.experts.0.w2
  - model.layers.13.block_sparse_moe.experts.0.w2
  - model.layers.0.block_sparse_moe.experts.0.w3
  - model.layers.1.block_sparse_moe.experts.0.w3
  - model.layers.2.block_sparse_moe.experts.0.w3
  - model.layers.3.block_sparse_moe.experts.0.w3
  - model.layers.4.block_sparse_moe.experts.0.w3
  - model.layers.5.block_sparse_moe.experts.0.w3
  - model.layers.6.block_sparse_moe.experts.0.w3
  - model.layers.7.block_sparse_moe.experts.0.w3
  - model.layers.8.block_sparse_moe.experts.0.w3
  - model.layers.9.block_sparse_moe.experts.0.w3
  - model.layers.10.block_sparse_moe.experts.0.w3
  - model.layers.11.block_sparse_moe.experts.0.w3
  - model.layers.12.block_sparse_moe.experts.0.w3
  - model.layers.13.block_sparse_moe.experts.0.w3
  - model.layers.0.block_sparse_moe.experts.1.w1
  - model.layers.1.block_sparse_moe.experts.1.w1
  - model.layers.2.block_sparse_moe.experts.1.w1
  - model.layers.3.block_sparse_moe.experts.1.w1
  - model.layers.4.block_sparse_moe.experts.1.w1
  - model.layers.5.block_sparse_moe.experts.1.w1
  - model.layers.6.block_sparse_moe.experts.1.w1
  - model.layers.7.block_sparse_moe.experts.1.w1
  - model.layers.8.block_sparse_moe.experts.1.w1
  - model.layers.9.block_sparse_moe.experts.1.w1
  - model.layers.10.block_sparse_moe.experts.1.w1
  - model.layers.11.block_sparse_moe.experts.1.w1
  - model.layers.12.block_sparse_moe.experts.1.w1
  - model.layers.13.block_sparse_moe.experts.1.w1
  - model.layers.40.block_sparse_moe.experts.1.w2
  - model.layers.0.block_sparse_moe.experts.1.w2
  - model.layers.1.block_sparse_moe.experts.1.w2
  - model.layers.2.block_sparse_moe.experts.1.w2
  - model.layers.3.block_sparse_moe.experts.1.w2
  - model.layers.4.block_sparse_moe.experts.1.w2
  - model.layers.5.block_sparse_moe.experts.1.w2
  - model.layers.6.block_sparse_moe.experts.1.w2
  - model.layers.7.block_sparse_moe.experts.1.w2
  - model.layers.8.block_sparse_moe.experts.1.w2
  - model.layers.9.block_sparse_moe.experts.1.w2
  - model.layers.10.block_sparse_moe.experts.1.w2
  - model.layers.11.block_sparse_moe.experts.1.w2
  - model.layers.12.block_sparse_moe.experts.1.w2
  - model.layers.5.block_sparse_moe.experts.1.w3
  - model.layers.0.block_sparse_moe.experts.1.w3
  - model.layers.1.block_sparse_moe.experts.1.w3
  - model.layers.2.block_sparse_moe.experts.1.w3
  - model.layers.3.block_sparse_moe.experts.1.w3
  - model.layers.4.block_sparse_moe.experts.1.w3
  - model.layers.6.block_sparse_moe.experts.1.w3
  - model.layers.7.block_sparse_moe.experts.1.w3
  - model.layers.8.block_sparse_moe.experts.1.w3
  - model.layers.9.block_sparse_moe.experts.1.w3
  - model.layers.10.block_sparse_moe.experts.1.w3
  - model.layers.11.block_sparse_moe.experts.1.w3
  - model.layers.12.block_sparse_moe.experts.1.w3
  - model.layers.13.block_sparse_moe.experts.1.w3
  - model.layers.1.block_sparse_moe.experts.2.w1
  - model.layers.0.block_sparse_moe.experts.2.w1
  - model.layers.2.block_sparse_moe.experts.2.w1
  - model.layers.3.block_sparse_moe.experts.2.w1
  - model.layers.4.block_sparse_moe.experts.2.w1
  - model.layers.5.block_sparse_moe.experts.2.w1
  - model.layers.6.block_sparse_moe.experts.2.w1
  - model.layers.7.block_sparse_moe.experts.2.w1
  - model.layers.8.block_sparse_moe.experts.2.w1
  - model.layers.9.block_sparse_moe.experts.2.w1
  - model.layers.10.block_sparse_moe.experts.2.w1
  - model.layers.11.block_sparse_moe.experts.2.w1
  - model.layers.12.block_sparse_moe.experts.2.w1
  - model.layers.13.block_sparse_moe.experts.2.w1
  - model.layers.1.block_sparse_moe.experts.2.w2
  - model.layers.0.block_sparse_moe.experts.2.w2
  - model.layers.2.block_sparse_moe.experts.2.w2
  - model.layers.3.block_sparse_moe.experts.2.w2
  - model.layers.4.block_sparse_moe.experts.2.w2
  - model.layers.5.block_sparse_moe.experts.2.w2
  - model.layers.6.block_sparse_moe.experts.2.w2
  - model.layers.7.block_sparse_moe.experts.2.w2
  - model.layers.8.block_sparse_moe.experts.2.w2
  - model.layers.9.block_sparse_moe.experts.2.w2
  - model.layers.10.block_sparse_moe.experts.2.w2
  - model.layers.11.block_sparse_moe.experts.2.w2
  - model.layers.12.block_sparse_moe.experts.2.w2
  - model.layers.13.block_sparse_moe.experts.2.w2
  - model.layers.1.block_sparse_moe.experts.2.w3
  - model.layers.0.block_sparse_moe.experts.2.w3
  - model.layers.2.block_sparse_moe.experts.2.w3
  - model.layers.3.block_sparse_moe.experts.2.w3
  - model.layers.4.block_sparse_moe.experts.2.w3
  - model.layers.5.block_sparse_moe.experts.2.w3
  - model.layers.6.block_sparse_moe.experts.2.w3
  - model.layers.7.block_sparse_moe.experts.2.w3
  - model.layers.8.block_sparse_moe.experts.2.w3
  - model.layers.9.block_sparse_moe.experts.2.w3
  - model.layers.10.block_sparse_moe.experts.2.w3
  - model.layers.11.block_sparse_moe.experts.2.w3
  - model.layers.12.block_sparse_moe.experts.2.w3
  - model.layers.13.block_sparse_moe.experts.2.w3
  - model.layers.2.block_sparse_moe.experts.3.w1
  - model.layers.1.block_sparse_moe.experts.3.w1
  - model.layers.0.block_sparse_moe.experts.3.w1
  - model.layers.3.block_sparse_moe.experts.3.w1
  - model.layers.4.block_sparse_moe.experts.3.w1
  - model.layers.5.block_sparse_moe.experts.3.w1
  - model.layers.6.block_sparse_moe.experts.3.w1
  - model.layers.7.block_sparse_moe.experts.3.w1
  - model.layers.8.block_sparse_moe.experts.3.w1
  - model.layers.9.block_sparse_moe.experts.3.w1
  - model.layers.10.block_sparse_moe.experts.3.w1
  - model.layers.11.block_sparse_moe.experts.3.w1
  - model.layers.12.block_sparse_moe.experts.3.w1
  - model.layers.13.block_sparse_moe.experts.3.w1
  - model.layers.2.block_sparse_moe.experts.3.w2
  - model.layers.1.block_sparse_moe.experts.3.w2
  - model.layers.0.block_sparse_moe.experts.3.w2
  - model.layers.3.block_sparse_moe.experts.3.w2
  - model.layers.4.block_sparse_moe.experts.3.w2
  - model.layers.5.block_sparse_moe.experts.3.w2
  - model.layers.6.block_sparse_moe.experts.3.w2
  - model.layers.7.block_sparse_moe.experts.3.w2
  - model.layers.8.block_sparse_moe.experts.3.w2
  - model.layers.9.block_sparse_moe.experts.3.w2
  - model.layers.10.block_sparse_moe.experts.3.w2
  - model.layers.11.block_sparse_moe.experts.3.w2
  - model.layers.12.block_sparse_moe.experts.3.w2
  - model.layers.13.block_sparse_moe.experts.3.w2
  - model.layers.2.block_sparse_moe.experts.3.w3
  - model.layers.1.block_sparse_moe.experts.3.w3
  - model.layers.0.block_sparse_moe.experts.3.w3
  - model.layers.3.block_sparse_moe.experts.3.w3
  - model.layers.4.block_sparse_moe.experts.3.w3
  - model.layers.5.block_sparse_moe.experts.3.w3
  - model.layers.6.block_sparse_moe.experts.3.w3
  - model.layers.7.block_sparse_moe.experts.3.w3
  - model.layers.8.block_sparse_moe.experts.3.w3
  - model.layers.9.block_sparse_moe.experts.3.w3
  - model.layers.10.block_sparse_moe.experts.3.w3
  - model.layers.11.block_sparse_moe.experts.3.w3
  - model.layers.12.block_sparse_moe.experts.3.w3
  - model.layers.13.block_sparse_moe.experts.3.w3
  - model.layers.3.block_sparse_moe.experts.4.w1
  - model.layers.2.block_sparse_moe.experts.4.w1
  - model.layers.1.block_sparse_moe.experts.4.w1
  - model.layers.0.block_sparse_moe.experts.4.w1
  - model.layers.4.block_sparse_moe.experts.4.w1
  - model.layers.5.block_sparse_moe.experts.4.w1
  - model.layers.6.block_sparse_moe.experts.4.w1
  - model.layers.7.block_sparse_moe.experts.4.w1
  - model.layers.8.block_sparse_moe.experts.4.w1
  - model.layers.9.block_sparse_moe.experts.4.w1
  - model.layers.10.block_sparse_moe.experts.4.w1
  - model.layers.11.block_sparse_moe.experts.4.w1
  - model.layers.12.block_sparse_moe.experts.4.w1
  - model.layers.13.block_sparse_moe.experts.4.w1
  - model.layers.2.block_sparse_moe.experts.4.w2
  - model.layers.3.block_sparse_moe.experts.4.w2
  - model.layers.1.block_sparse_moe.experts.4.w2
  - model.layers.20.block_sparse_moe.experts.4.w2
  - model.layers.0.block_sparse_moe.experts.4.w2
  - model.layers.4.block_sparse_moe.experts.4.w2
  - model.layers.5.block_sparse_moe.experts.4.w2
  - model.layers.6.block_sparse_moe.experts.4.w2
  - model.layers.7.block_sparse_moe.experts.4.w2
  - model.layers.8.block_sparse_moe.experts.4.w2
  - model.layers.9.block_sparse_moe.experts.4.w2
  - model.layers.10.block_sparse_moe.experts.4.w2
  - model.layers.11.block_sparse_moe.experts.4.w2
  - model.layers.12.block_sparse_moe.experts.4.w2
  - model.layers.3.block_sparse_moe.experts.4.w3
  - model.layers.2.block_sparse_moe.experts.4.w3
  - model.layers.1.block_sparse_moe.experts.4.w3
  - model.layers.0.block_sparse_moe.experts.4.w3
  - model.layers.4.block_sparse_moe.experts.4.w3
  - model.layers.5.block_sparse_moe.experts.4.w3
  - model.layers.6.block_sparse_moe.experts.4.w3
  - model.layers.7.block_sparse_moe.experts.4.w3
  - model.layers.8.block_sparse_moe.experts.4.w3
  - model.layers.9.block_sparse_moe.experts.4.w3
  - model.layers.10.block_sparse_moe.experts.4.w3
  - model.layers.11.block_sparse_moe.experts.4.w3
  - model.layers.12.block_sparse_moe.experts.4.w3
  - model.layers.13.block_sparse_moe.experts.4.w3
  - model.layers.4.block_sparse_moe.experts.5.w1
  - model.layers.3.block_sparse_moe.experts.5.w1
  - model.layers.2.block_sparse_moe.experts.5.w1
  - model.layers.1.block_sparse_moe.experts.5.w1
  - model.layers.0.block_sparse_moe.experts.5.w1
  - model.layers.5.block_sparse_moe.experts.5.w1
  - model.layers.6.block_sparse_moe.experts.5.w1
  - model.layers.7.block_sparse_moe.experts.5.w1
  - model.layers.8.block_sparse_moe.experts.5.w1
  - model.layers.9.block_sparse_moe.experts.5.w1
  - model.layers.10.block_sparse_moe.experts.5.w1
  - model.layers.11.block_sparse_moe.experts.5.w1
  - model.layers.12.block_sparse_moe.experts.5.w1
  - model.layers.13.block_sparse_moe.experts.5.w1
  - model.layers.4.block_sparse_moe.experts.5.w2
  - model.layers.2.block_sparse_moe.experts.5.w2
  - model.layers.3.block_sparse_moe.experts.5.w2
  - model.layers.1.block_sparse_moe.experts.5.w2
  - model.layers.0.block_sparse_moe.experts.5.w2
  - model.layers.5.block_sparse_moe.experts.5.w2
  - model.layers.6.block_sparse_moe.experts.5.w2
  - model.layers.7.block_sparse_moe.experts.5.w2
  - model.layers.8.block_sparse_moe.experts.5.w2
  - model.layers.9.block_sparse_moe.experts.5.w2
  - model.layers.10.block_sparse_moe.experts.5.w2
  - model.layers.11.block_sparse_moe.experts.5.w2
  - model.layers.12.block_sparse_moe.experts.5.w2
  - model.layers.13.block_sparse_moe.experts.5.w2
  - model.layers.4.block_sparse_moe.experts.5.w3
  - model.layers.3.block_sparse_moe.experts.5.w3
  - model.layers.2.block_sparse_moe.experts.5.w3
  - model.layers.1.block_sparse_moe.experts.5.w3
  - model.layers.0.block_sparse_moe.experts.5.w3
  - model.layers.5.block_sparse_moe.experts.5.w3
  - model.layers.6.block_sparse_moe.experts.5.w3
  - model.layers.7.block_sparse_moe.experts.5.w3
  - model.layers.8.block_sparse_moe.experts.5.w3
  - model.layers.9.block_sparse_moe.experts.5.w3
  - model.layers.10.block_sparse_moe.experts.5.w3
  - model.layers.11.block_sparse_moe.experts.5.w3
  - model.layers.12.block_sparse_moe.experts.5.w3
  - model.layers.13.block_sparse_moe.experts.5.w3
  - model.layers.5.block_sparse_moe.experts.6.w1
  - model.layers.4.block_sparse_moe.experts.6.w1
  - model.layers.3.block_sparse_moe.experts.6.w1
  - model.layers.2.block_sparse_moe.experts.6.w1
  - model.layers.1.block_sparse_moe.experts.6.w1
  - model.layers.0.block_sparse_moe.experts.6.w1
  - model.layers.6.block_sparse_moe.experts.6.w1
  - model.layers.7.block_sparse_moe.experts.6.w1
  - model.layers.8.block_sparse_moe.experts.6.w1
  - model.layers.9.block_sparse_moe.experts.6.w1
  - model.layers.10.block_sparse_moe.experts.6.w1
  - model.layers.11.block_sparse_moe.experts.6.w1
  - model.layers.12.block_sparse_moe.experts.6.w1
  - model.layers.13.block_sparse_moe.experts.6.w1
  - model.layers.5.block_sparse_moe.experts.6.w2
  - model.layers.4.block_sparse_moe.experts.6.w2
  - model.layers.2.block_sparse_moe.experts.6.w2
  - model.layers.3.block_sparse_moe.experts.6.w2
  - model.layers.1.block_sparse_moe.experts.6.w2
  - model.layers.0.block_sparse_moe.experts.6.w2
  - model.layers.6.block_sparse_moe.experts.6.w2
  - model.layers.7.block_sparse_moe.experts.6.w2
  - model.layers.8.block_sparse_moe.experts.6.w2
  - model.layers.9.block_sparse_moe.experts.6.w2
  - model.layers.10.block_sparse_moe.experts.6.w2
  - model.layers.11.block_sparse_moe.experts.6.w2
  - model.layers.12.block_sparse_moe.experts.6.w2
  - model.layers.13.block_sparse_moe.experts.6.w2
  - model.layers.5.block_sparse_moe.experts.6.w3
  - model.layers.4.block_sparse_moe.experts.6.w3
  - model.layers.3.block_sparse_moe.experts.6.w3
  - model.layers.2.block_sparse_moe.experts.6.w3
  - model.layers.1.block_sparse_moe.experts.6.w3
  - model.layers.0.block_sparse_moe.experts.6.w3
  - model.layers.6.block_sparse_moe.experts.6.w3
  - model.layers.7.block_sparse_moe.experts.6.w3
  - model.layers.8.block_sparse_moe.experts.6.w3
  - model.layers.9.block_sparse_moe.experts.6.w3
  - model.layers.10.block_sparse_moe.experts.6.w3
  - model.layers.11.block_sparse_moe.experts.6.w3
  - model.layers.12.block_sparse_moe.experts.6.w3
  - model.layers.13.block_sparse_moe.experts.6.w3
  - model.layers.5.block_sparse_moe.experts.7.w1
  - model.layers.6.block_sparse_moe.experts.7.w1
  - model.layers.3.block_sparse_moe.experts.7.w1
  - model.layers.4.block_sparse_moe.experts.7.w1
  - model.layers.2.block_sparse_moe.experts.7.w1
  - model.layers.0.block_sparse_moe.experts.7.w1
  - model.layers.7.block_sparse_moe.experts.7.w1
  - model.layers.8.block_sparse_moe.experts.7.w1
  - model.layers.9.block_sparse_moe.experts.7.w1
  - model.layers.10.block_sparse_moe.experts.7.w1
  - model.layers.11.block_sparse_moe.experts.7.w1
  - model.layers.12.block_sparse_moe.experts.7.w1
  - model.layers.13.block_sparse_moe.experts.7.w1
  - model.layers.14.block_sparse_moe.experts.7.w1
  - model.layers.6.block_sparse_moe.experts.7.w2
  - model.layers.5.block_sparse_moe.experts.7.w2
  - model.layers.4.block_sparse_moe.experts.7.w2
  - model.layers.2.block_sparse_moe.experts.7.w2
  - model.layers.3.block_sparse_moe.experts.7.w2
  - model.layers.1.block_sparse_moe.experts.7.w2
  - model.layers.0.block_sparse_moe.experts.7.w2
  - model.layers.7.block_sparse_moe.experts.7.w2
  - model.layers.8.block_sparse_moe.experts.7.w2
  - model.layers.9.block_sparse_moe.experts.7.w2
  - model.layers.10.block_sparse_moe.experts.7.w2
  - model.layers.11.block_sparse_moe.experts.7.w2
  - model.layers.12.block_sparse_moe.experts.7.w2
  - model.layers.13.block_sparse_moe.experts.7.w2
  - model.layers.6.block_sparse_moe.experts.7.w3
  - model.layers.5.block_sparse_moe.experts.7.w3
  - model.layers.4.block_sparse_moe.experts.7.w3
  - model.layers.3.block_sparse_moe.experts.7.w3
  - model.layers.2.block_sparse_moe.experts.7.w3
  - model.layers.0.block_sparse_moe.experts.7.w3
  - model.layers.7.block_sparse_moe.experts.7.w3
  - model.layers.8.block_sparse_moe.experts.7.w3
  - model.layers.9.block_sparse_moe.experts.7.w3
  - model.layers.10.block_sparse_moe.experts.7.w3
  - model.layers.11.block_sparse_moe.experts.7.w3
  - model.layers.12.block_sparse_moe.experts.7.w3
  - model.layers.13.block_sparse_moe.experts.7.w3
  - model.layers.14.block_sparse_moe.experts.7.w3
  - model.layers.0.block_sparse_moe.gate
  - model.layers.1.block_sparse_moe.gate
  - model.layers.2.block_sparse_moe.gate
  - model.layers.3.block_sparse_moe.gate
  - model.layers.4.block_sparse_moe.gate
  - model.layers.5.block_sparse_moe.gate
  - model.layers.6.block_sparse_moe.gate
  - model.layers.7.block_sparse_moe.gate
  - model.layers.8.block_sparse_moe.gate
  - model.layers.9.block_sparse_moe.gate
  - model.layers.10.block_sparse_moe.gate
  - model.layers.11.block_sparse_moe.gate
  - model.layers.12.block_sparse_moe.gate
  - model.layers.13.block_sparse_moe.gate
  
model_config:
  output_router_logits: true

datasets:
  - path: /workspace/datasets/dolphin-2.9/dolphin201-sharegpt2.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/Ultrachat200kunfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/dolphin-coder-translate-sharegpt2.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/dolphin-coder-codegen-sharegpt2.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/m-a-p_Code-Feedback-sharegpt-unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/m-a-p_CodeFeedback-Filtered-Instruction-sharegpt-unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/not_samantha_norefusals.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/Orca-Math-resort-unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/agent_instruct_react_unfiltered.jsonl
    type: sharegpt  
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/toolbench_instruct_j1s1_3k_unfiltered.jsonl
    type: sharegpt  
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/toolbench_negative_unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/toolbench_react_10p_unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/toolbench_tflan_cot_30p_unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/openhermes200k_unfiltered.jsonl
    type: sharegpt 
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/SystemConversations.jsonl
    type: sharegpt
    conversation: chatml

chat_template: chatml

dataset_prepared_path: thingy
val_set_size: 0.0002
output_dir: ./out

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

gradient_accumulation_steps: 8
micro_batch_size: 4
num_epochs: 3
logging_steps: 1
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 2.7e-5

wandb_project: dolphin-2.9-mixtral-8x22b
wandb_watch:
wandb_run_id:
wandb_log_model:

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
# resume_from_checkpoint: /home/ehartford/axolotl/out/checkpoint-316
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
saves_per_epoch: 8
save_total_limit: 2
save_steps:
evals_per_epoch: 4
eval_sample_packing: false
debug:
deepspeed: deepspeed_configs/zero3_bf16_cpuoffload_params.json
weight_decay: 0.05
fsdp:
fsdp_config:
special_tokens:
  eos_token: "<|im_end|>"
tokens:
  - "<|im_start|>"

Training results

Training Loss	Epoch	Step	Validation Loss
0.7022	0.0	1	0.6989
0.5344	0.25	238	0.5138
0.5204	0.5	476	0.5018
0.5059	0.75	714	0.4951
0.5112	1.0	952	0.4911
0.4561	1.24	1190	0.4978
0.478	1.49	1428	0.4935
0.4714	1.74	1666	0.4899
0.4626	1.99	1904	0.4861
0.3675	2.22	2142	0.5240
0.3595	2.47	2380	0.5229
0.3438	2.72	2618	0.5217

Framework versions

Transformers 4.40.0.dev0
Pytorch 2.2.2+cu121
Datasets 2.15.0
Tokenizers 0.15.0

Как запустить

На своей GPU в ailist. Цена — по текущим предложениям провайдеров

GPU в ailist

vLLM

Можно запустить

Подбираем GPU…

Класс архитектуры: MixtralForCausalLM
Ревизия весов: cba5134