juan-reina33/mrpc-bert-uncased-finetuned

BERT Base Uncased Fine-Tuned on MRPC

This model is a fine-tuned version of bert-base-uncased on the GLUE MRPC (Microsoft Research Paraphrase Corpus) dataset.
It determines whether two English sentences are paraphrases (have the same meaning).
The model was trained using the Hugging Face Transformers library and PyTorch on an Apple M2 machine.

Model Details

Model Description

This is a BERT-based sequence classification model fine-tuned for semantic similarity and paraphrase detection.
It outputs a binary label:

• 1 → the two sentences mean the same thing (paraphrase)

• 0 → they don’t mean the same thing (non-paraphrase)

• Developed by: Juan Sebastián Reina García

• Model type: Transformer-based encoder (BERT)

• Language(s): English

• License: Apache-2.0 (inherits from BERT base)

• Finetuned from model: bert-base-uncased

Model Sources

• Repository: https://huggingface.co/juan-reina33/mrpc-bert-uncased-finetuned
• Paper: GLUE Benchmark (Wang et al., 2018)
• Demo: Coming soon

---

Uses

Direct Use

Use this model for sentence-pair classification or semantic similarity detection tasks.

Example use cases:

• Detecting duplicate questions or answers
• Identifying paraphrased customer support tickets
• Measuring semantic equivalence in English text

Example code:

clf = pipeline("text-classification", model="juan-reina33/mrpc-bert-uncased-finetuned")
result = clf({
    "text": "The company released a new product.",
    "text_pair": "A new product was launched by the company."
})
print(result)
# [{'label': 'paraphrase', 'score': 0.89}]

Downstream Use

You can fine-tune this model further for:

Natural Language Inference (NLI)

Question–Answer entailment

Duplicate detection in domain-specific datasets

Out-of-Scope Use

Not intended for:

Non-English text

Factual reasoning or high-stakes decision systems (e.g., legal, medical)

Bias, Risks, and Limitations

The model may reflect biases from BERT’s original pretraining data.

Performance may drop on noisy, informal, or domain-specific English.

Not suitable for multilingual text.

Recommendations

Always test the model on your target dataset before deployment. Use human validation for sensitive use cases.

How to Get Started

python Copy code from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("juan-reina33/mrpc-bert-uncased-finetuned") model = AutoModelForSequenceClassification.from_pretrained("juan-reina33/mrpc-bert-uncased-finetuned")

Training Details

Training Data

Dataset: MRPC (Microsoft Research Paraphrase Corpus) from the GLUE benchmark

3,668 training pairs

408 validation pairs

Binary labels: 1 = paraphrase, 0 = not

Training Procedure

Optimizer: AdamW

Learning rate: 5e-5

Batch size: 8

Epochs: 3

Scheduler: Linear decay, no warm-up

Precision: fp32

Loss: CrossEntropyLoss

Speeds, Sizes, Times

Training time: ~25 min total (3 epochs on Apple M2)

Model size: ~420 MB

Frameworks: PyTorch 2.9.0, Transformers 4.57.1

Evaluation

Metric Score

Accuracy 84.8% F1 Score 89.2%

Environmental Impact

Hardware: Apple MacBook Air M2, 16 GB RAM

Training duration: ~0.5 hrs

Cloud Provider: None (local)

Estimated CO₂: < 0.01 kg

Citation

BibTeX

bibtex Copy code

  title = {BERT Base Uncased Fine-Tuned on MRPC},
  author = {Juan Sebastián Reina García},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/juan-reina33/mrpc-bert-uncased-finetuned}}
}

APA

Reina García, J. S. (2025). BERT Base Uncased Fine-Tuned on MRPC [Computer software]. Hugging Face. https://huggingface.co/juan-reina33/mrpc-bert-uncased-finetuned

Model Card Author

Juan Sebastián Reina García

Contact

Hugging Face: @juan-reina33