biorecall-PubMedBERT-NLI

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

biorecall-PubMedBERT-NLI

This model is a fine-tuned version of microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext on synthetic biomedical Natural Language Inference (NLI) dataset available at [https://huggingface.co/datasets/sharraks97/PubMedQA-synthetic-nli] to classify whether a hypothesis sentence is entailed by, contradicted by, or neutral with respect to a biomedical context. It achieves the following results on the evaluation set:

• Loss: 0.2108
• Accuracy: 0.9274
• F1 Macro: 0.9054

Model description

• Base model: microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext
• Task: Natural Language Inference (3-class classification)
• Labels: entailment, neutral, contradiction
• Architecture: cross-encoder classification model

Intended uses & limitations

This model can be used to:

• Verify generated biomedical answers against retrieved context
• Detect hallucinated sentences
• Score factual consistency in RAG pipelines
• Support biomedical entailment research

Example usage:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("sharraks97/biorecall-PubMedBERT-NLI")
model = AutoModelForSequenceClassification.from_pretrained("sharraks97/biorecall-PubMedBERT-NLI")

premise = "Biomedical context text"
hypothesis = "Claim sentence"

inputs = tokenizer(premise, hypothesis, return_tensors="pt", truncation=True)
outputs = model(**inputs)
prediction = torch.argmax(outputs.logits, dim=-1)

Training and evaluation

• Accuracy
• Macro F1-score

Best model checkpoint selected using validation Macro F1-score.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Use adamw_torch_fused with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 4

Training results

Framework versions

• Transformers 4.57.1
• Pytorch 2.9.1+cu128
• Datasets 4.4.1
• Tokenizers 0.22.1

Limitations

• Trained on synthetic data
• May not generalize beyond biomedical QA
• Sentence-level reasoning only
• Not intended for clinical use

Author

Rakshit Sharma Project: BioRecall (Reasoning Enabled and Contradiction Augmented LLM RAG)