BabyMistral Model Card

Model Overview

BabyMistral is a compact yet powerful language model designed for efficient text generation tasks. Built on the Mistral architecture, this model offers impressive performance despite its relatively small size.

Key Specifications

• Parameters: 1.5 billion
• Training Data: 1.5 trillion tokens
• Architecture: Based on Mistral
• Training Duration: 70 days
• Hardware: 4x NVIDIA A100 GPUs

Model Details

Architecture

BabyMistral utilizes the Mistral AI architecture, which is known for its efficiency and performance. The model scales this architecture to 1.5 billion parameters, striking a balance between capability and computational efficiency.

Training

• Dataset Size: 1.5 trillion tokens
• Training Approach: Trained from scratch
• Hardware: 4x NVIDIA A100 GPUs
• Duration: 70 days of continuous training

Capabilities

BabyMistral is designed for a wide range of natural language processing tasks, including:

• Text completion and generation
• Creative writing assistance
• Dialogue systems
• Question answering
• Language understanding tasks

Usage

Getting Started

To use BabyMistral with the Hugging Face Transformers library:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("OEvortex/BabyMistral")
tokenizer = AutoTokenizer.from_pretrained("OEvortex/BabyMistral")

# Define the chat input
chat = [
#     { "role": "system", "content": "You are BabyMistral" },
    { "role": "user", "content": "Hey there! How are you? 😊" }
]

inputs = tokenizer.apply_chat_template(
    chat,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)


# Generate text
outputs = model.generate(
    inputs,
    max_new_tokens=256,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
    eos_token_id=tokenizer.eos_token_id,

    
)

response = outputs[0][inputs.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))

#I am doing well! How can I assist you today? 😊

Ethical Considerations

While BabyMistral is a powerful tool, users should be aware of its limitations and potential biases:

• The model may reproduce biases present in its training data
• It should not be used as a sole source of factual information
• Generated content should be reviewed for accuracy and appropriateness

Limitations

• May struggle with very specialized or technical domains
• Lacks real-time knowledge beyond its training data
• Potential for generating plausible-sounding but incorrect information