Sentence-bert finetuning

Mizz54112 · November 13, 2025, 6:07am

github.com/huggingface/sentence-transformers

examples/sentence_transformer/training/nli/training_nli.py

master

"""
The system trains BERT (or any other transformer model like RoBERTa, DistilBERT etc.) on the SNLI + MultiNLI (AllNLI) dataset
with softmax loss function. At every 1000 training steps, the model is evaluated on the
STS benchmark dataset

Usage:
python training_nli.py

OR
python training_nli.py pretrained_transformer_model_name
"""

import logging
import sys
import traceback
from datetime import datetime

from datasets import load_dataset

from sentence_transformers import SentenceTransformer, losses

This file has been truncated. show original

This code on GitHub trains Sentence-BERT on an NLI dataset. The evaluation section currently uses STS. Is it possible to change this to accuracy (the correct answer rate for the NLI data)? If so, I would appreciate specific guidance on how to implement this change.

John6666 · November 13, 2025, 11:05am

Probably something like this…?

Mizz54112 · November 14, 2025, 3:57am

Thank you for explaining it so thoroughly. It’s very easy to understand.

Topic		Replies	Views
Info regarding sentence-transformers Models	8	1732	August 26, 2020
Fine-tuning multilingual BERT for sequence classification with Trainer API Beginners	7	715	December 12, 2023
Fine-tuning lm with nsp 🤗Transformers	0	1203	January 19, 2021
Fine Tuning A sentence transformer model with my own data Intermediate	2	3440	April 17, 2024
Inference of bert-base with sst2 dataset Beginners	0	433	July 11, 2023

Sentence-bert finetuning

Related topics