Proximity Grammar Corrector (T5-small)

A small fine-tuned T5-small model for English grammar correction. Part of the Proximity project — a lightweight background tool that fixes grammar via a global hotkey.

What this model does

Takes a sentence with grammar mistakes and outputs a corrected version. Examples:

Input	Output
She dont like going to the store.	She doesn't like going to the store.
He are moving here.	He is moving here.
I has went to the market yesterday.	I have gone to the market yesterday.
They was happy about they new house.	They were happy about their new house.

What this model does NOT do

• It is not a chat model. It will not hold a conversation or answer questions.
• It is not a style/tone rewriter. It targets grammatical correctness, not voice or wording improvements.
• It struggles with some irregular verb forms not well represented in training data (e.g. "drinked" instead of "drank" was not corrected in testing).
• It is trained on a small dataset (~6,000 sentence pairs), so coverage of rare or complex grammatical errors is limited.

Training details

• Base model: t5-small (~60M parameters)
• Dataset: Owishiboo/grammar-correction (~6,000 ungrammatical → grammatical sentence pairs)
• Epochs: 6
• Final train loss: 0.1685
• Final eval loss: ~0.186

Usage

from transformers import T5Tokenizer, T5ForConditionalGeneration

tokenizer = T5Tokenizer.from_pretrained("EnderAir/proximity")
model = T5ForConditionalGeneration.from_pretrained("EnderAir/proximity")

text = "grammar: He are moving here."
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_length=64, num_beams=5)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
# "He is moving here."

Note: always prefix input text with "grammar: " — this is the T5 task prefix convention used during training.

License

Apache 2.0