Iris314/finetuned_model

Model Card for Cuisine Classification (Fine-Tuned DistilBERT)

This model predicts the cuisine type of Pittsburgh restaurants based on review text.
It was fine-tuned from distilbert-base-uncased on the dataset ecopus/pgh_restaurants.

It achieves the following results:

• Evaluation (Augmented split): Accuracy 0.969, F1 0.957, Precision 0.948, Recall 0.969
• External Validation (Original split): Accuracy 0.94, F1 0.92

---

Model Details

• Developed by: Xinxuan Tang (CMU)
• Dataset curated by: Emily Copus (CMU)
• Base model: DistilBERT (distilbert-base-uncased)
• Library: Transformers
• Language(s): English
• License: apache-2.0 (dataset + model card)

---

Uses

Direct Use

• Educational practice in text classification.
• Experimenting with fine-tuning compact transformers.

Downstream Use

• Could be adapted for restaurant recommendation demos.
• Teaching NLP pipelines for classification tasks.

Out-of-Scope Use

• Not suitable for production deployment.
• Not intended for sentiment analysis or tasks outside cuisine prediction.

---

Training Procedure

Training Hyperparameters

• learning_rate: 2e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: AdamW (betas=(0.9,0.999), eps=1e-08)
• lr_scheduler_type: linear
• num_epochs: 5

Training Results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1	Precision	Recall
2.6677	1.0	80	2.4746	0.3563	0.2142	0.1662	0.3563
1.7201	2.0	160	1.5893	0.7750	0.6895	0.6644	0.7750
1.1994	3.0	240	1.1417	0.8938	0.8503	0.8180	0.8938
1.0890	4.0	320	0.9315	0.9250	0.8959	0.8784	0.9250
0.7052	5.0	400	0.8675	0.9688	0.9570	0.9480	0.9688

---

Evaluation

Testing Data

• Augmented split: 1000 reviews (synthetic augmentation)
• Original split: 100 reviews (external validation)

Metrics

• Accuracy, weighted F1, Precision, Recall
• Confusion matrix used for external validation

---

Framework Versions

• Transformers: 4.56.1
• PyTorch: 2.8.0+cu126
• Datasets: 4.0.0
• Tokenizers: 0.22.0

---

Bias, Risks, and Limitations

• Small dataset: only 100 original reviews.
• Synthetic augmentation: may introduce artifacts.
• Geographic bias: limited to Pittsburgh restaurants.

Recommendations

Treat results as proof-of-concept, not production-ready.

---

How to Get Started with the Model

You can load this fine-tuned DistilBERT model directly from the Hugging Face Hub and run inference:

from transformers import pipeline

# Load the fine-tuned model
classifier = pipeline("text-classification", model="YOUR_USERNAME/finetuned_model")

# Example input
sample = "This cozy little spot serves delicious tacos with great service!"
print(classifier(sample))