Fine-tuned Qwen2.5-1.5B and Qwen2.5-3B models for automated text quality assessment. Predicts quality scores on a 0-1 scale focusing on educational value and mathematical intelligence.

Model Details

Base Models: Qwen2.5-1.5B / Qwen2.5-3B
Task: Text Quality Regression
Languages: English, Chinese
Training Data: OpenSQZ/Classifiers-Data
Loss Function: MSE Loss

Performance

Model	Test MSE Loss
Qwen2.5-1.5B	0.00226
Qwen2.5-3B	0.00209

Quick Start

Installation

pip install transformers torch

Usage

python

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

# Load model and tokenizer
model_name = "OpenSQZ/Qwen2.5-1.5B-Classifier"  # or Qwen2.5-3B-Quality-Classifier
model = AutoModelForSequenceClassification.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Predict quality score
text = "Linear algebra is fundamental to understanding vector spaces and matrix operations in mathematics."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=8192)

with torch.no_grad():
    outputs = model(**inputs)
    score = torch.sigmoid(outputs.logits).item()

print(f"Quality Score: {score:.3f}")  # Output: Quality Score: 0.847

Quality Score Interpretation

Score Range	Quality Level	Use Case
0.8 - 1.0	Excellent	Premium training data
0.6 - 0.8	Good	Standard training data
0.4 - 0.6	Average	Conditional use
0.0 - 0.4	Poor	Filter out

Model Selection

1.5B Model: Faster inference, good for real-time applications
3B Model: Higher accuracy, better for batch processing

Limitations

Optimized for educational and mathematical content
May not generalize well to creative or subjective content
Scores should be used as guidance, not absolute judgments

Citation

bibtex

@model{qwen25_quality_classifier_2025,
  title={Qwen2.5 Text Quality Classifier},
  author={Chao Li, Yifan Zhang},
  year={2025},
  publisher={OpenSQZ}
}

License

Apache 2.0

OpenSQZ/Qwen2.5-3B-classifier

Qwen2.5 Text Quality Classifier