bert-base-uncased-finetuned-glue-mnli

Fine-Tuned BERT on MNLI (MultiNLI)

This repository contains a BERT-base model fine-tuned on the GLUE MultiNLI dataset.
The task is natural language inference (NLI): given a premise and a hypothesis, predict the relationship between them as one of:

• entailment
• contradiction
• neutral

---

Model Details

• Base Model: bert-base-uncased
• Fine-tuning Method: Full fine-tuning
• Task: Sequence Classification (3 classes)
• Frameworks: Hugging Face Transformers, PyTorch

---

Training Configuration

• Epochs: 4
• Batch Size: 64 (train) / 64 (eval)
• Learning Rate: 3e-5 (AdamW fused optimizer)
• Scheduler: Linear Warmup (10%)
• Other Techniques:
  • Gradient checkpointing
  • Gradient accumulation (steps = 2)
  • Mixed precision (FP16 when GPU available)

---

Dataset

• Name: GLUE MNLI
• Train Split: ~392k examples
• Evaluation Splits:
  • Matched (in-domain): premise/hypothesis from the same domain
  • Mismatched (cross-domain): premise/hypothesis from different domains

---

Results

Final Evaluation (Epoch 4)

---

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_name = "mia-project-2025/bert-base-uncased-finetuned-glue-mnli"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Example input
premise = "A soccer game with multiple males playing."
hypothesis = "Some men are playing a sport."
inputs = tokenizer(premise, hypothesis, return_tensors="pt", padding=True, truncation=True)

# Prediction
outputs = model(**inputs)
pred = outputs.logits.argmax(-1).item()
labels = ["entailment", "neutral", "contradiction"]
print("Prediction:", labels[pred])