pewcubes/punctuation-restoration

Punctuation Restoration Model

This is a demo of a fine-tuned Qwen3-0.6B model for automatic punctuation restoration. The model takes unpunctuated text (like ASR transcripts) and restores proper punctuation and capitalization.

Model Details

• —Base Model: Qwen/Qwen3-0.6B-Base
• —Task: Punctuation Restoration
• —Training: Fine-tuned on combined datasets (AMI, SWBD, Earnings, CHiME, GTN, SPGI)
• —Checkpoint: 200,000 steps

Usage

Simply paste or type unpunctuated text into the input box, and the model will restore punctuation and capitalization.

Examples

Input: hello world this is a test Output: Hello, world! This is a test.

Input: i went to the store yesterday and bought some milk Output: I went to the store yesterday and bought some milk.

How It Works

The model uses a causal language modeling approach where it learns to transform normalized (unpunctuated) text into unnormalized (properly punctuated) text. It was trained using a prompt-response format:

### Input
[normalized text]

### Output
[unnormalized text with punctuation]

Performance

The model achieves competitive Character Error Rate (CER) scores across multiple test datasets, making it suitable for real-world ASR post-processing applications.

Built as part of a Final Year Project (FYP) 2024-2025.