README.md

metadata

library_name: transformers
license: apache-2.0
base_model: bert-base-uncased
tags:
  - paraphrase-detection
  - sentence-pair-classification
  - glue
  - mrpc
metrics:
  - accuracy
  - f1
model-index:
  - name: bert_paraphrase
    results:
      - task:
          name: Paraphrase Detection
          type: text-classification
        dataset:
          name: GLUE MRPC
          type: glue
          config: mrpc
          split: validation
        metrics:
          - name: Accuracy
            type: accuracy
            value: 0.8676
          - name: F1
            type: f1
            value: 0.9078
language:
  - en

bert_paraphrase

This model is a fine-tuned version of bert-base-uncased on the Microsoft Research Paraphrase Corpus (MRPC), a subset of the GLUE benchmark.

It is trained to determine whether two sentences are semantically equivalent (paraphrases) or not.

📊 Evaluation Results

• Loss: 0.4042
• Accuracy: 0.8676
• F1: 0.9078

🧾 Model Description

• Model type: BERT-base (uncased)
• Task: Binary classification (paraphrase vs not paraphrase)
• Languages: English
• Labels:
  • 0 → Not paraphrase
  • 1 → Paraphrase

---

✅ Intended Uses & Limitations

Intended uses & limitations

Intended uses

• Detect if two sentences convey the same meaning.
• Useful for:
  • Duplicate question detection (e.g., Quora, FAQ bots).
  • Semantic similarity search.
  • Improving information retrieval systems.

Limitations

• Only trained on English (MRPC dataset).
• May not generalize well to other domains (e.g., legal, medical).
• Binary labels only (no "degree of similarity").

---

📚 How to Use

You can use this model with the Hugging Face pipeline for quick inference:

from transformers import pipeline

paraphrase_detector = pipeline(
    "text-classification",
    model="azherali/bert_paraphrase",
    tokenizer="azherali/bert_paraphrase"
)

single_pair = [
    {"text": "The car is red.", "text_pair": "The automobile is red."},
]
result = paraphrase_detector(single_pair)
print( result)
[{'label': 'paraphrase', 'score': 0.9801033139228821}]

# Test pairs
pairs = [
    {"text": "The car is red.", "text_pair": "The automobile is red."},
    {"text": "He enjoys playing football.", "text_pair": "She likes cooking."},
]
result = paraphrase_detector(pairs)
print( result)



[{'label': 'paraphrase', 'score': 0.9801033139228821}, {'label': 'not_paraphrase', 'score': 0.9302119016647339}]

Using AutoModel & AutoTokenizer:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("azherali/bert_paraphrase")
model = AutoModelForSequenceClassification.from_pretrained("azherali/bert_paraphrase")

# Example sentences
sent1 = "The quick brown fox jumps over the lazy dog."
sent2 = "A fast brown fox leaps over a lazy dog."

inputs = tokenizer(sent1, sent2, return_tensors="pt")
outputs = model(**inputs)

logits = outputs.logits
predicted_class = torch.argmax(logits, dim=1).item()

print("Prediction:", model.config.id2label[predicted_class])
Prediction: paraphrase

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1
No log	1.0	230	0.3894	0.8309	0.8836
No log	2.0	460	0.3511	0.8505	0.8964
0.4061	3.0	690	0.4042	0.8676	0.9078

Framework versions

• Transformers 4.55.2
• Pytorch 2.8.0+cu126
• Datasets 4.0.0
• Tokenizers 0.21.4