Destigmatizer (Reframe)

Python library that leverages LLMs to remove harmful language from texts on drug use.

Testing

You can explore Destigmatizer's functionality interactively using Jupyter notebooks. Below is an example of how to classify and rewrite text using the library.

Using Destigmatizer in a Jupyter Notebook

Installing and Set Up

!pip install destigmatizer

# Import the destigmatizer package
import destigmatizer
from destigmatizer import core
from destigmatizer.clients import get_client
from destigmatizer.utils import get_default_model, get_model_mapping
import json
import os

# Define example texts
NON_DRUG_POST = "I think we should really work on the housing crisis in urban areas. The homeless are getting scary."
DRUG_NON_STIGMA_POST = "I really feel for people who suffer from substance use disorder, being unable to control an impulse due to dependency sounds scary."
DRUG_AND_STIGMA_POST = "Addicts really need to get control of themselves. Just stop doing drugs. It seems like these people are just lacking willpower."

# Set up API client
# for example OpenAI

client = get_client('openai', api_key=api_key)

Classifying Drug-Related and Stigmatizing Content

# Classify whether a text is drug-related and whether it contains stigma
drug_classification = core.classify_drug_content([NON_DRUG_POST, DRUG_NON_STIGMA_POST, DRUG_AND_STIGMA_POST], client, model)
stigma_classification = core.classify_stigma([NON_DRUG_POST, DRUG_NON_STIGMA_POST, DRUG_AND_STIGMA_POST], client, model)

# Print results
for text, drug_label, stigma_label in zip([NON_DRUG_POST, DRUG_NON_STIGMA_POST, DRUG_AND_STIGMA_POST], drug_classification, stigma_classification):
    print(f"Text: {text}\nDrug-related: {drug_label}\nStigmatizing: {stigma_label}\n")

Rewwriting Stigmatizing Content

# Rewrite a stigmatizing post to be more empathetic
rewritten_text = core.rewrite_stigmatizing_text(DRUG_AND_STIGMA_POST, client, model)

print(f"Original: {DRUG_AND_STIGMA_POST}")
print(f"Rewritten: {rewritten_text}")

Reframe includes integrated tests for all major functions. You can run tests in two ways:

Using the CLI command

# Test all functions
reframe-test

# Test a specific function
reframe-test drug --api_key YOUR_API_KEY

# Use a specific model with Together AI
reframe-test --client_type together --model "meta-llama/Meta-Llama-3.1" 

# Use Claude
reframe-test --client_type claude --model "claude-3-haiku-20240307"

# Using environment variables (recommended)
# First set the environment variables:
export OPENAI_API_KEY="your-openai-key"
export TOGETHER_API_KEY="your-together-key"
export ANTHROPIC_API_KEY="your-anthropic-key"

# Then run tests without specifying the API key
reframe-test --client_type openai
reframe-test --client_type together
reframe-test --client_type claude

Running test scripts directly

# From the root directory
python3 -m reframe.tests.run_all_tests