Tinymistral 6x248M Instruct

🚀 Model Card for M4-ai/TinyMistral-6x248M-Instruct

M4-ai/TinyMistral-6x248M-Instruct is a fine - tuned language model. It offers high - quality text generation capabilities across a wide range of topics, suitable for various instructive tasks.

✨ Features

• Mixture of Experts (MoE) Architecture: An ensemble of different models combined via the LazyMergekit framework.
• Domain - Specific Expertise: Each expert in the model specializes in different domains, from software development to multilingual text generation.
• Enhanced Instructive Performance: Fine - tuned on the hercules - v1.0 dataset to improve performance on instructive tasks.

📦 Installation

There is no specific installation section in the original README. However, for usage, you need to install the following Python libraries:

!pip install -qU transformers bitsandbytes accelerate

💻 Usage Examples

Basic Usage

from transformers import AutoTokenizer
import transformers
import torch

model = "M4-ai/TinyMistral-6x248M-Instruct"

tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    model_kwargs={"torch_dtype": torch.float16, "load_in_4bit": True},
)

messages = [{"role": "user", "content": "Explain what a Mixture of Experts is in less than 100 words."}]
prompt = pipeline.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])

📚 Documentation

Model Details

Model Description

M4-ai/TinyMistral-6x248M-Instruct is a fine - tuned language model based on a Mixture of Experts (MoE) architecture. It is an ensemble of various models expertly combined using the LazyMergekit framework. The base pre - trained mixture model includes several versions of the TinyMistral model, with each expert specialized in different domains from technical software development to multilingual text generation. This fine - tuned version aims to enhance the model's performance on instructive tasks by leveraging the hercules - v1.0 dataset. The model is intended for applications requiring guidance, explanations, and analysis across a wide array of topics.

Intended Use

This model is designed for developers and researchers who need a sophisticated language model capable of understanding and generating text in response to instructive prompts. It is suitable for various tasks, including but not limited to technical explanations, educational content, policy analysis, and problem - solving across disciplines such as computer science, history, and natural sciences. Users should be aware of the model's limitations and potential biases, especially when dealing with sensitive topics.

Training Data

The model was fine - tuned using the hercules - v1.0 dataset, an augmented version of the teknium/openhermes dataset. Hercules - v1.0 includes updated data sources like ise - uiuc/Magicoder - Evol - Instruct - 110K, jondurbin/airoboros - 3.2, and WizardLM/WizardLM_evol_instruct_V2_196k, as well as specialized datasets in mathematics, chemistry, physics, and biology. The dataset has been cleaned to remove RLHF refusals and potentially toxic content from airoboros - 3.2. However, users should note that a small portion of the data might still contain sensitive content. You can use the ChatML prompt format for this model.

Limitations and Bias

While efforts have been made to clean the training data, the potential for biases and harmful content remains, as with any large language model. Users should exercise caution and discretion when using the model, especially in applications that might amplify existing biases or expose users to sensitive content. The model is not recommended for scenarios requiring strict content moderation or for users without the ability to filter or assess the model's outputs critically.

Evaluation

Performance degradation has been observed when using the Inference API; thus, the model is not recommended for this usage. Instead, users should follow the recommended inference parameters provided in the base model card to optimize performance.

Disclaimer

The model has been fine - tuned on the hercules - v1.0 dataset, which contains content from sources with known issues of toxic examples. Users of M4 - ai/TinyMistral - 6x248M - Instruct must acknowledge and agree to the following:

• The dataset may include "toxic"/"harmful" content, profanity, and sensitive material.
• The content does not necessarily reflect the beliefs or opinions of the developers.
• Users must comply with local laws regarding free speech and content use.
• Users assume full responsibility for the download and utilization of the dataset and model, indemnifying the developers from all liabilities.

Contributions

Thanks to @jtatman, @aloobun, @Felladrin, and @Locutusque for their contributions to this model.

📄 License

The model is licensed under the Apache - 2.0 license.