qwen2.5-gec-7b

Qwen2.5-GEC-7B: Russian Grammatical Error Correction

This model is a fine-tuned version of Qwen/Qwen2.5-7B for Russian grammatical error correction.
It is optimized for efficient inference using Unsloth and 4‑bit quantization, requiring only 4 GB of GPU VRAM.

Model Description

• Base model: Qwen2.5-7B
• Task: Correcting grammatical, orthographic, and stylistic errors in Russian sentences.
• Fine‑tuned on: p1746-lingua/ru-gec-v1.
• Quantization: 4‑bit (bitsandbytes) – enables running on consumer GPUs with 6–8 GB VRAM.
• Prompt format: Alpaca‑style (instruction, input, response).
• Max sequence length: 2048 tokens.

Intended Use

Use the model to automatically correct Russian text. It works best on short to medium sentences (up to 512 tokens). Example applications:

• Proofreading tools
• Language learning platforms
• Preprocessing for downstream NLP tasks

Quick Start

The following code loads the model in 4‑bit mode and runs inference on a single GPU:

import os
os.environ['HF_HUB_DOWNLOAD_TIMEOUT'] = '60'

max_seq_length = 2048 
dtype = None 
load_in_4bit = True
    
from unsloth import FastLanguageModel
    
model, tokenizer = FastLanguageModel.from_pretrained(
  model_name = "p1746-lingua/qwen2.5-gec-7b",
  max_seq_length = max_seq_length,
  dtype = dtype,
  load_in_4bit = load_in_4bit
)

FastLanguageModel.for_inference(model)

alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{}

### Input:
{}

### Response:
{}"""

inputs = tokenizer(
[
    alpaca_prompt.format(
        "Lec.",              # Instruction (task description)
        "Он шла по доороге", # Input text with errors
        "", # Keep empty for generation
    )
], return_tensors = "pt").to("cuda")

from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 128)

Expected output:

Она шла по дороге

Performance

Training Details

• Hardware: GPU NVIDIA GeForce RTX 4070 Ti SUPER
• Optimizer: AdamW (lr=2e‑4)
• Batch size: 4 (gradient accumulation steps = 4)
• Epochs: 3
• Precision: bfloat16 for base model, 4‑bit LoRA adapters
• Framework: Unsloth + PEFT + Transformers

Limitations

• The model may occasionally over‑correct or change stylistic choices that are not errors.
• Works best on standard Russian; dialects and highly informal language may degrade performance.
• Long sentences may cause truncation or memory issues on low‑VRAM GPUs.

Ethical Considerations

This model is intended for assisting human writers, not for fully automated grading or decision‑making. Always review corrections in critical contexts.

Citation

If you use this model in your research, please cite:

@misc{qwen2.5-gec-7b,
  author = {Plesovskikh A.},
  title = {Qwen2.5-GEC-7B: Russian Grammatical Error Correction},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/p1746-lingua/qwen2.5-gec-7b}
}

Contact

For questions or collaboration, open an issue on the model repository.