Incubator

This repo is the official implementation for Incubating Text Classifiers Following User Instruction with Nothing but LLM. We allow users to get a personalized classifier with only the instruction as input. The incubation is based on a llama-2-7b fine-tuned on Huggingface Meta Data and Self-Diversification.

Incubating Classifiers

You can use the script incubate.sh to incubate your own classifiers.

python incubate.py --n_epoch 16 \
    --batch_size 4 \
    --device 1 \
    --n_sample 16 \
    --max_new_tokens 64 \
    --instruction "Build a classifier that can categorize text messages by 'about food' and 'about movie'." \
    --incubator "KomeijiForce/Incubator-llama-2-7b" \
    --classifier "roberta-base" \
    --save_path "roberta-base-incubated"

By running the default incubation script, you can view the following output on default test cases:

Input: I love 'Spiderman 2'!
Predicted Label: about movie
Input: I ate a delicious pudding!
Predicted Label: about food