Weak supervision vs LLMs for intent classification

This package serves as basis for the paper "n a Few Words: Comparing Weak Supervision and LLMs for Short Query Intent Classification"

Link to the paper:

DOI of the paper: https://doi.org/10.1145/3726302.3730213

In-context learning:

Create conda environment:

$ conda create --name intent_classification_llms python==3.10

Activate the environment:

$ source activate intent_classification_llms

Use pip to install requirements:

(intent_classification_llms) $ pip install -r requirements.txt

Fine-tuning:

Fine-tuning experiments are in the folder fine-tuning (Jupyter notebooks)

Parameters used for fine-tuning up to 15000 examples:

learning_rate = 2e-5 
lora_rank = 8
batch_size = 8 
gradient_accumulation_steps = 8
weight_decay = 0.05 
max_grad_norm = 1.0
early_stop_patience = 2

Parameters used for fine-tuning from 30000 examples:

learning_rate = 1.5e-5  
lora_rank = 8  
batch_size = 16  
gradient_accumulation_steps = 8  
weight_decay = 0.1
max_grad_norm = 1.0 
early_stop_patience = 3