TypoFormer (Stage 2 INT8)

TypoFormer is a 100% offline, character-aware hybrid Transformer designed for ultra-low latency (<10ms on CPU) next-word autocomplete and non-autoregressive grammatical error correction (GEC).

Companion Windows desktop app and training source code available on GitHub.

Model Architecture

• Trunk: 10 Transformer blocks (768 hidden dimension, 2048 FFN dimension)
• Attention: Grouped Query Attention (12 Query heads, 4 Key-Value heads) + Rotary Position Embeddings (RoPE)
• Non-Linearity: SwiGLU + Pre-RMSNorm
• Hybrid Input: Gated fusion of 16,000 Byte-Level BPE embeddings with Multi-Scale Character 1D-Convolutions (kernel sizes 2, 3, 5, 8)
• Quantization: Symmetric INT8 ONNX graph (130 MB footprint)

Files in this Repository

• lm_stage2_int8.onnx: Causal language model for next-word suggestion (~138 MB)
• gec_stage2_int8.onnx: Non-autoregressive edit-tagger for grammar & typo correction (~130 MB)
• tokenizer.json, vocab.json, merges.txt: 16k Byte-Level BPE tokenizer
• edit_maps.json: High-frequency confusion dictionary for atomic replacements & appends
• infer.py: Self-contained Python runner using ONNX Runtime

Quickstart

pip install onnxruntime tokenizers numpy
git clone https://huggingface.co/fahiiim/TypoFormer
cd TypoFormer
python infer.py

Python Example

from infer import TypoFormer

model = TypoFormer(".")

# 1. Autocomplete (Causal LM)
suggestions = model.autocomplete("Could you please review the configuration file before the ", top_k=3)
print("Autocomplete suggestions:", suggestions)

# 2. Grammar & Typo Fix (GEC)
clean = model.correct_grammar("This is teh wrong way to write.")
print("Fixed text:", clean)

License

MIT