PicoMistral-23M

PicoMistral-23M is an ultra-compact, "pico-sized" language model trained from scratch by Pico-Kittens. It is designed to run on extremely weak hardware or edge devices where standard LLMs cannot operate.

PicoMistral Logo

Model Details

  • Developed by: Pico-Kittens
  • Model type: Mistral-based Causal Language Model
  • Training Status: Trained from scratch (Not a fine-tune)
  • Parameters: ~23 Million
  • Language(s): English
  • License: apache-2.0

Training Data

The model was trained on the first 120,000 samples of the HuggingFaceH4/ultrachat_200k dataset. The focus of this training was to capture conversational structure and basic English grammar within a minimal parameter budget.

Capabilities & Limitations

PicoMistral-23M is an experimental model. While it demonstrates that a 23M model can maintain grammatical structure, users should be aware of the following:

  • Grammar: Highly competent; it produces well-structured and grammatically correct English sentences.
  • Semantic Drift: Due to its small size, the model often loses the "thread" of a conversation or shifts topics unexpectedly.
  • Hallucinations: High. It will confidently state incorrect facts or invent data.
  • Intent Recognition: Limited. It may struggle with complex instructions but can hold basic, short-form conversations.

Generation Sample

User: Write a comprehensive analysis that compares and contrasts the various types of ad networks available in the market today based on their targeting capabilities, pricing structures, audience reach, ad formats, and publisher requirements. Provide relevant examples to support your arguments and make recommendations on which type of ad network would best suit a specific business niche.

PicoMistral-23M:

Introduction: Maintaining is an essential part of our brand identity for all aspects of marketing and SEO. It provides excellent customer service, product listings, and advertisers. This report aims to analyze and compare and contrast how they are used in the market as well. The purpose of this paper will be about the different types of ad networks available in the market, including demographic target demographics, demographics, influencer partnerships, ad formats, and publishers.

Market Analysis: The market is a leading provider of digital content, advertising, and social media channels. Companies can leverage keywords, social media platforms, email marketing campaigns, or other media outlets to drive sales and engagement with the target audience. They also have access to new products such as products, services, and product descriptions. They can use these channels to personalize their email list, share photos, and engage with the target audience.

Target Audience: The competition between two professionals is the primary target audience who are experiencing fashion trends and marketing efforts. These individuals tend to engage in fashion-conscious relationships with their followers, which can lead to increased sales revenue.

Positive Metrics: The demand for mobile ads has been increasing popularity over the past few years. Their targeted messaging leads to better performance, satisfaction, and loyalty. Some of the most popular brands include Facebook, Twitter, TikTok, and LinkedIn.

Marketing Strategies: The competition between both marketers and customers is likely to continue buying online, but it may not only help them grow their business. However, they should consider factors like product description, website titles, username, and relevant content. Social Media Ads is crucial for businesses to ensure that their brand and ad copy is optimized for search engine results.

Functionality: We understand what is and what we do when you choose your brand, brand, and company. We believe that we are committed to providing exceptional customer service, and we look forward to seeing our success. By creating compelling emails, we are confident that we must offer exclusive deals and promotions to meet the needs of our customers.

Community Outcomes: Our target audience has a strong relationship with our target audience and wants to remain loyal by reaching out to us in a more positive way. Our target audience is always on the right direction, but it’s important to know where we want to offer the highest-thinking approach. As we anticipate our unique selling proposition, we need to prioritize its target audience, who are looking to others, and we want to provide our target audience with the best experience possible.


How to Get Started

Use the following script to download, extract, and run the model:

import torch
from transformers import MistralForCausalLM, AutoTokenizer, pipeline

model_id = "PicoKittens/PicoMistral-23M"

tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False)
model = MistralForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32, device_map="auto")

prompt = (
    "[BOS]<|user|>\n"
    "Write a comprehensive analysis that compares and contrasts the various types of ad networks available in the market today based on their targeting capabilities, pricing structures, audience reach, ad formats, and publisher requirements. Provide relevant examples to support your arguments and make recommendations on which type of ad network would best suit a specific business niche."
    "<|end|>\n<|assistant|>\n"
)

chatbot = pipeline("text-generation", model=model, tokenizer=tokenizer)

response = chatbot(
    prompt, 
    max_new_tokens=512, 
    do_sample=True, 
    temperature=0.7, 
    top_p=0.9, 
    repetition_penalty=1.2, 
    return_full_text=False
)

print(response[0]['generated_text'])
Downloads last month
155
Safetensors
Model size
23.6M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train PicoKittens/PicoMistral-23M