Meyssa/coedit-large

CoEdIT Large (ONNX)

ONNX quantized version of Grammarly's CoEdIT-Large model for browser-based text editing and grammatical error correction with Transformers.js.

Original Model

• Source: grammarly/coedit-large
• Paper: CoEdIT: Text Editing by Task-Specific Instruction Tuning (EMNLP 2023)
• Architecture: Flan-T5-Large (encoder-decoder)
• Parameters: ~770M

Conversion Details

• Format: ONNX
• Quantization: INT8 (dynamic quantization)
• Size: ~780MB
• Converted by: Optimum CLI

How It Works

CoEdIT is a text-to-text model fine-tuned for various editing tasks. For grammar correction, use the prompt format:

Fix grammatical errors: <input text>

The model generates the corrected text directly.

Usage with Transformers.js

import { pipeline } from '@huggingface/transformers';

const corrector = await pipeline(
  'text2text-generation',
  'YOUR_USERNAME/coedit-large',
  { dtype: 'q8' }
);

const result = await corrector('Fix grammatical errors: He go to school yesterday.');
console.log(result[0].generated_text);
// "He went to school yesterday."

Supported Tasks

CoEdIT supports multiple editing tasks via different prompts:

• Fix grammatical errors: ... - Grammar correction
• Paraphrase: ... - Paraphrasing
• Simplify: ... - Text simplification
• Make this more formal: ... - Formality transfer
• Make this more coherent: ... - Coherence improvement

Performance

More flexible than GECToR (can rewrite sentences), but slower due to autoregressive generation.

License

Apache 2.0 (following original model license)