I

Ibert Roberta Base

Developed by kssteven
I-BERT is a pure integer quantized version of RoBERTa, storing parameters in INT8 format and using integer operations for inference, significantly improving inference speed.
Downloads 2,988
Release Time : 3/2/2022

Model Overview

I-BERT replaces floating-point operations with integer operations in the Transformer architecture, enabling efficient inference. Suitable for tasks requiring fast text processing.

Model Features

Pure integer operations
All parameters are stored in INT8 format, with inference performed entirely using integer operations, eliminating the need for floating-point operations.
Efficient inference
Tested on Nvidia T4 GPU, achieves up to 4x inference speedup compared to the floating-point version.
Quantization-aware training
Supports quantization-aware fine-tuning, optimizing quantized model performance through a three-stage process.

Model Capabilities

Text classification
Natural language understanding
Efficient inference

Use Cases

Text processing
Text classification
Tasks such as MRPC text classification
Maintains high accuracy through quantization-aware training
Featured Recommended AI Models
AIbase
Empowering the Future, Your AI Solution Knowledge Base
Š 2025AIbase