albert-paraphrase-detector

ALBERT-base-v2 Fine-tuned for Semantic Similarity (QQP/MRPC)

Model Details

Model Description

This is a fine-tuned version of albert-base-v2 on paraphrase detection tasks such as GLUE-QQP (Quora Question Pairs) and MRPC (Microsoft Research Paraphrase Corpus).
It can be used to determine whether two sentences are paraphrases (semantically similar) or not.

• Developed by: Peeyush
• Model type: Sentence-pair classification (binary: paraphrase vs not paraphrase)
• Language(s): English
• License: Apache-2.0
• Finetuned from model: albert-base-v2

Model Sources [optional]

• Repository: your-username/albert-paraphrase-similarity
• Paper (base model): ALBERT: A Lite BERT for Self-supervised Learning of Language Representations

Uses

Direct Use

• Paraphrase detection: Check if two sentences mean the same thing.
• Semantic textual similarity: Determine closeness of meaning between two texts.

Downstream Use

• Duplicate question detection (e.g., Q&A forums like Quora or StackOverflow).
• Information retrieval (ranking by semantic similarity).
• Chatbots / Virtual assistants (detecting intent rephrasing).

Out-of-Scope Use

• Not a generative model → cannot rewrite or generate paraphrases.
• Not trained on multilingual data → limited to English.

---

Bias, Risks, and Limitations

• The model inherits biases from QQP/MRPC (e.g., common question styles, certain domains).
• May not generalize to informal text, code-mixed text, or specialized domains (e.g., medical, legal).
• Can misclassify edge cases where semantic similarity is subtle.

Recommendations

• Always evaluate on your target domain before deployment.
• For production, consider threshold-tuning (instead of raw classification).

---

How to Get Started with the Model

Example usage:

model  = AutoModelForSequenceClassification.from_pretrained('peeyush01/albert-paraphrase-detector')
tokenizer = AutoTokenizer.from_pretrained('peeyush01/albert-paraphrase-detector-tokenizer')

def predict_paraphrase(sentence1, sentence2):
    inputs = tokenizer(sentence1, sentence2, return_tensors="pt", padding=True, truncation=True)
    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits
        probs = torch.softmax(logits, dim=1)
        paraphrase_prob = probs[0][1].item()
    return {"Paraphrase": paraphrase_prob, "Not Paraphrase": 1 - paraphrase_prob}

import torch

pairs = [
    ("The movie was fantastic!", "The film was amazing!"),
    ("He is playing cricket.", "She is reading a book."),
]

for s1, s2 in pairs:
    result = predict_paraphrase(s1, s2)
    print(f"Sentence 1: {s1}")
    print(f"Sentence 2: {s2}")
    print(f"Result: {result}\n")

Training Details

Training Data

• Dataset: GLUE MRPC
• Description: The Microsoft Research Paraphrase Corpus (MRPC) contains pairs of sentences automatically extracted from online news sources, with human annotations indicating whether each pair captures a paraphrase/semantic equivalence relationship.
• Size: ~3,700 training pairs, 408 validation pairs, 1,725 test pairs.
• Labels:
  • 1 → Paraphrase (semantically equivalent)
  • 0 → Not paraphrase

Training Procedure

Preprocessing

• Both sentences were tokenized using AlbertTokenizer with truncation and padding (max_length).
• Columns sentence1, sentence2, and idx were dropped.
• The label column was renamed from label → labels.
• Dataset was set in PyTorch format.

Training Hyperparameters

• Base model: albert-base-v2
• Epochs: 3
• Batch size: 16 (train and eval)
• Optimizer: AdamW (via Hugging Face Trainer)
• Warmup steps: 600
• Weight decay: 0.01
• Evaluation strategy: Per epoch
• Precision regime: FP32

Speeds, Sizes, Times

• Training performed with Hugging Face Trainer.
• Training time: ~20–30 mins on a single GPU (Tesla T4); longer on CPU.
• Final checkpoint size: ~47 MB.

---

Evaluation

Testing Data, Factors & Metrics

Testing Data

• Evaluation performed on the GLUE MRPC validation set (~408 examples).

Factors

• Sentence pairs vary in length, syntactic complexity, and semantic overlap.
• Evaluation primarily captures semantic similarity in short news-style English text.

Metrics

• Accuracy: percentage of correctly classified sentence pairs.
• F1 Score: harmonic mean of precision and recall, important due to class imbalance.

Results

(Expected range for ALBERT-base on MRPC — please replace with your actual run metrics if available)

• Accuracy: ~86–88%
• F1 Score: ~89–91%

Summary

The fine-tuned ALBERT model achieves strong performance on the MRPC benchmark, demonstrating effectiveness at capturing semantic similarity and paraphrase relationships between sentence pairs.