You can deploy this model and start earning money today!

Vardis/
medical-LM

Name: medical-LM
Availability: InStock

Views

No views yet

Greek mT5 Text Correction Model

Model Details

Model type: Transformer-based encoder-decoder (mT5-base)
Language: Greek (el)
License: Same as google/mt5-base
Task: Grammatical and orthographic text correction

Model Description

This model is a fine-tuned version of the multilingual T5 (mT5) model, specifically adapted for correcting noisy Greek text generated by ASR systems. The fine-tuning framed the task as a supervised sequence-to-sequence problem: the model takes a corrupted sentence as input and produces the corrected sentence as output.

To guide the model, an instruction-style prompting approach was employed. Input sentences were formatted as:

correct: "corrupted sentence"

Example:

Input	Model Output
correct:Ποσο θα διαρκέσει η αποκατάσταση μιτα από το χιρουργειο.	Πόσο θα διαρκέσει η αποκατάσταση μετά από το χειρουργείο;

The model captures long-range dependencies, allowing it to correct misspellings, punctuation errors, grammatical mistakes, and occasional missing or repeated words.

Training Data

Synthetic error-augmented dataset: 56,000 corrupted–clean sentence pairs
Data sources: Medical corpus and Wikipedia
Error types introduced: Vowel swaps, duplicated letters, punctuation mistakes, grammatical issues
Input-output format: Sequence-to-sequence (corrupted → correct)

Example corrupted–clean pairs:

Corrupted Sentence	Clean Sentence
Ιδανηκο για προληψοη αλλαα και θεραπεια των κατακλισεων.	Ιδανικό για πρόληψη αλλα και θεραπεία των κατακλίσεων.
H πιεση του αερα ρυθμιζετα ανυλιογα με το βαρος του ασθενη.	H πίεση του αέρα ρυθμίζεται ανάλογα με το βάρος του ασθενή.

How to Use

python

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("Vardis/greek-mt5-medical")
model = AutoModelForSeq2SeqLM.from_pretrained("Vardis/greek-mt5-medical")

input_text = "correct:Ποσο θα διαρκέσει η αποκατάσταση μιτα από το χιρουργειο."
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs)
corrected_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(corrected_text)

Vardis/medical-LM