YoctoLM is an ultra-lightweight, high-performance, and resource-efficient Small Language Model (SLM) environment designed to run seamlessly across both Node.js and Modern Web Browsers.

By utilizing an elegant N-gram probabilistic framework backed by an intelligent, frequency-weighted semantic similarity fallback algorithm, YoctoLM provides fast intent-matching and text generation without the heavy memory overhead of massive modern LLMs.

✨ Features

• Isomorphic & Universal: One codebase that automatically detects its environment—running native fs/promises in Node.js or fetch() in the browser.
• Weighted N-Gram Architecture: Implements variable context windows to map linguistic dependencies efficiently.
• Intent-Matching Engine: Features an advanced token-similarity algorithm weighted by inverse term frequency (IDF) to dynamically prioritize high-value words over common structural words (e.g., what, is, the).
• Dynamic Fallback Loading: Safely attempts to fetch training datasets locally first; if the local environment is restricted (e.g., browser CORS), it smoothly falls back to a remote GitHub repository.
• Configurable Generation: Complete control over randomness with parameterized temperature logic during token selection.
• Customizability: You can train it on your own data.

📁 Project Architecture

To maintain clean code separation, scalability, and ease of deployment, the project is structured into four specialized modules:

• tokenizer.js: Manages word-to-ID mappings, dictionaries, and global token frequency counting during training.
• similarity.js: Houses mathematical operations including the weighted Jaccard/Bigram overlapping calculations and probability selection algorithms.
• yoctolm.js: The central brain orchestrating the custom sequence trainer, context windows, and token generation loops.
• main.js: The central entry point responsible for environment handling, dataset ingestion, performance benchmarking, and testing prompt outputs.
• index.html: A formatted web page is a chat for direct interaction with the form within the browser.

🚀 Getting Started

Download Source

• Click Code > Download ZIP.
• or using git :

git clone https://github.com/mohdmot/YoctoLM.git

• Unzip it.

1. Using Chat Webpage :

• Run local server in src folder:

# Using npx
npx serve . -l 8080

# Or using Python
python -m http.server 8080

• Go to http://localhost:8080

Now enjoy the chat experience!

2. Using Node.js

Just run it /src/test.js :

node test.js

3. Use it inside your project

import { YoctoLM } from './yoctolm.js';
import { loadTrainingData } from './data.js';

let trainingData = await loadTrainingData('training_data.json');

let yocto = new YoctoLM();
//yocto.temperature = 0.1;

yocto.train(trainingData);

const prompt = "hi, what is your name?"; 

let output = yocto.generate(prompt,1000)

console.log(output);

📄 Data

In the utils folder you will find everything you need. We used the dailydialog dataset for training and added an AI-generated dataset to train the model on it. In total 90,000 question, So .. How to add a new data ?

1. Generate

You can use any available AI (I recommend DeepSeek) you have to generate questions and answers in this format.

Hello __eou__ Hi there
How are you ? __eou__ I'm Fine.

Or format an entire conversation in one line

Matthew ? Hi ! __eou__ Steve ! Haven't seen you in ages ! How long's it been ? __eou__ It's got to be almost a year now . __eou__ How have you been ? __eou__ I'm pretty good . I've been looking for a place to live recently . My lease runs out next month and I don't want to renew it . __eou__ Yeah , I remember the neighborhood . Have you found a place yet ? __eou__ Not yet . But I am still looking through the classifieds . Wish me luck . __eou__ Well , maybe I can help . Remember my neighbor ? __eou__ Mrs . Thou ? __eou__ Yes , her daughter's having a baby , so she's moving in with her to help out . I think if you are interested , you can come over and have a look . __eou__ Great . It's a lovely neighborhood . And it would be nice to be neighbors again . It would be just like the old days ! __eou__ I'll ask Mrs . Thou when she's available to show the apartment and let you know . Has your number changed ? __eou__

Remember that each line is independent of the other.

2. Prepare

Now we need to convert it into a JSON file containing all the questions and answers in a unified format:

[
  {"q": "hi there", "a": "hello! how are you today?"},
  {"q": "how are you", "a": "i am doing great, thanks for asking."},
  {"q": "what is your name", "a": "i am a tiny language model."},
  {"q": "what time is it", "a": "i do not have a watch, sorry."},
  {"q": "are you a human", "a": "no, i am an ai assistant."}
]

Add your generated data to a text file in the utils folder, then add it to the datasets list in prepare_data.py :

datasets_files = ['dialogues_train.txt', 'dialogues_test.txt', 'dialogues_validation.txt','ai_generated.txt']
data_n = 1500000000 # All avaliable question 

Now you have training_data.json, move it to the required location.

📝 License

This project is open-source and available under the MIT License.