Introducing Fikra Nano 1B
A revolutionary edge model engineered for extreme low-resource environments. High-throughput inference without the heavy VRAM tax, fine-tuned locally for the African ecosystem.
// architecture details
Built for the Edge. Fully Transparent.
Built strictly as a solo operation, Fikra Nano 1B leverages a custom implementation of ternary weight layers. By quantizing standard matrix parameters down to -1, 0, and 1, we eliminate continuous floating-point operations.
-
01
Base Model Validation
Fine-tuned extensively on the TII Falcon-E-1B-Base utilizing Dolly-15k and GSM8K datasets.
-
02
Open Notebook Logic
We don't hide the pipeline. The exact LoRA attachments, data mapping, and evaluation code are public on Kaggle.
Frequently Asked Questions
What is Fikra Nano 1B?
Fikra Nano 1B is a small, instruction-tuned language model built on the Falcon-E-1B-Base. It is designed to be an ultra-lightweight assistant for edge deployment, ensuring inference runs smoothly in environments with strict hardware limitations.
Where is the training code hosted?
In the spirit of building in public, all data parsing routines, tokenizer mappings, and LoRA attachment scripts are available as a public executable script. You can review and fork it at our official Kaggle Notebook repository.
Is it fully Open Source?
Yes. Both the GGUF weights and standard safetensors are hosted publicly on Hugging Face under the Lacesseapp organization. It serves as a foundational component for developers integrating the FastAPI-based Fikra inference engine.