Ephraimmm/Pidgin_0.1

Nigerian Pidgin Assistant

Overview

This Space is a chat demo for a Nigerian Pidgin English conversational assistant. It loads a 4-bit quantized gemma-3-4b-it base model via Unsloth and applies a LoRA adapter (Ephraimmm/PIDGIN_gemma-3) fine-tuned to respond in Pidgin English, then serves the model through a simple Gradio chat interface with streaming text generation.

Features

• —Pidgin English chat — a system prompt instructs the model to reply only in Nigerian Pidgin (e.g. "I de o, how you de").
• —Streaming responses — uses TextIteratorStreamer with a background generation thread so tokens can be produced incrementally.
• —Conversation history — chat turns are kept in-session and displayed in a gr.Chatbot component.
• —Save & download transcript — a "Save Conversation" button writes the chat history to a timestamped file that can be downloaded from the UI.

Tech Stack

• —Gradio (Blocks UI) for the web interface
• —Unsloth (FastLanguageModel) for fast, memory-efficient loading of the 4-bit quantized base model
• —Base model: unsloth/gemma-3-4b-it-unsloth-bnb-4bit
• —Fine-tuned adapter: Ephraimmm/PIDGIN_gemma-3 (PEFT/LoRA), loaded with peft.PeftModel
• —Transformers for tokenization, chat templating, and streaming generation
• —PyTorch for inference (CUDA if available, falling back to CPU)

This Space is an early ("0.1") demo build: the app is a single script with no batching, authentication, or conversation persistence beyond a local text file, and generation parameters are hard-coded. It's best understood as a functional proof of concept for the underlying Pidgin fine-tune rather than a production interface.

Author

Built by Ephraimmm