app.py

# This is a project of Chakra Lab LLC. All rights reserved.

import gradio as gr
import os

import torch
from peft import PeftConfig, PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

device = 'cuda' if torch.cuda.is_available() else 'cpu'

base_model_name = "google/gemma-7b"
adapter_model_name = "samidh/cope-g7bq-2c-hs.s1.5fpc.9-sx.s1.5.9o-VL.s1.5.9-HR.s5-SH.s5-l5e5-e3-d25-r8"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    #bnb_4bit_quant_type="nf4",
    #bnb_4bit_compute_dtype=torch.bfloat16,
    #bnb_4bit_use_double_quant=True
)

model = AutoModelForCausalLM.from_pretrained(base_model_name,
                                             token=os.environ['HF_TOKEN'],
                                             quantization_config=bnb_config,
                                             device_map="auto")
model = PeftModel.from_pretrained(model, adapter_model_name, token=os.environ['HF_TOKEN'])
model.merge_and_unload()

model = model.to(device)

tokenizer = AutoTokenizer.from_pretrained(base_model_name)

PROMPT = """
INSTRUCTIONS
============
Examine the given POLICY and determine if the given CONTENT meets the criteria for ANY of the LABELS. Answer "1" if yes, and "0" if no.

POLICY
======

{policy}


CONTENT
=======

{content}


ANSWER
======

"""

DEFAULT_POLICY = """# Policy

## Instructions

This policy is designed to determine whether or not content is hate speech.

## Definition of Terms

- Protected Classes: Individuals or groups of individuals defined by sharing a race, ethnicity, religion, disability, sex, gender identity, or sexual orientation. The scope is narrowed to focus on characteristics commonly recognized in anti-discrimination laws. Religions themselves, as concepts ARE NOT protected - only their members are.
- Non-Protected Classes: Groups defined by other characteristics, such as age, socioeconomic status, and education, which are typically excluded partially or entirely from anti-discrimination laws.

## Interpretation of Language

## Definition of Labels

### (HS): Hate Speech

#### Includes

- Explicit Violence Advocacy: Direct and clear threats of or praise for violence or harm against individuals or groups based on their membership in a particular protected class. Stating an intention to defend against aggression DOES NOT qualify
- Intentional Dehumanization: Statements that clearly depict individuals or groups as inherently ""other"", alien, animalistic, unintelligent, immoral, unclean, or less-than-fully-human based on their membership in a particular protected class in a way that justifies harm or discrimination.
- Targeted Use of Derogatory Slurs: Targeting another person or group of people using a one-word name for a particular protected class that has an inherent negative connotation (e.g. Nigger, Kike, Cunt, Retard). Multi-word terms are never slurs.
- Explicit Discrimination Advocacy: Direct and clear calls for exclusion, segregation, or discrimination against individuals or groups based on their membership in a particular protected class, with a clear intent to promote inequality.
- Direct Hateful Insults: Content that directly addresses another person or group of people the second person (e.g. ""You over there"") and insults them based on their membership in a particular protected class

#### Excludes

- Artistic and Educational Content: Expressions intended for artistic, educational, or documentary purposes that discuss sensitive topics but do not advocate for violence or discrimination against individuals or groups based on their membership in a particular protected class.
- Political and Social Commentary: Commentary on political issues, social issues, and political ideologies that does not directly incite violence or discrimination against individuals or groups based on their membership in a particular protected class.
- Rebutting Hateful Language: Content that rebuts, condemns, questions, criticizes, or mocks a different person's hateful language or ideas OR that insults the person advocating those hateful
- Quoting Hateful Language: Content in which the author quotes someone else's hateful language or ideas while discussing, explaining, or neutrally factually presenting those ideas.
- Describing Sectarian Violence: Content that describes, but does not endorse or praise, violent physical injury against a specifically named race, ethnicity, nationality, sexual orientation, or religious community by another specifically named race, ethnicity, nationality, sexual orientation, or religious community
"""

DEFAULT_CONTENT = "LLMs steal our jobs."

# Function to make predictions
def predict(content, policy):
    input_text = PROMPT.format(policy=policy, content=content)
    input_ids = tokenizer.encode(input_text, return_tensors="pt")

    with torch.inference_mode():
        outputs = model(input_ids)
        logits = outputs.logits[:, -1, :]  # Get logits for the last token
        predicted_token_id = torch.argmax(logits, dim=-1).item()
    decoded_output = tokenizer.decode([predicted_token_id])
    if decoded_output == '0':
        return f'NON-Violating ({decoded_output})'
    else:
        return f'VIOLATING ({decoded_output})'

with gr.Blocks() as iface:
    gr.Markdown("# CoPE Alpha Preview")
    gr.Markdown("See if the given content violates your given policy.")
    
    with gr.Row():
        content_input = gr.Textbox(label="Content", lines=2, value=DEFAULT_CONTENT)
        policy_input = gr.Textbox(label="Policy", lines=10, value=DEFAULT_POLICY)
    
    submit_btn = gr.Button("Submit")
    output = gr.Label(label="Label")
    
    gr.Markdown("""
    ## About CoPE
    
    CoPE (the COntent Policy Evaluation engine) is a small language model capable of accurate content policy labeling. This is a **preview** of our alpha release and is strictly for **research** purposes. This should **NOT** be used for any production use cases.
    
    ## How to Use
                                                                                                                                                                                                                                                                       
    1. Enter your content in the "Content" box.
    2. Specify your policy in the "Policy" box.
    3. Click "Submit" to see the results.

    **Note**: Inference times are **slow** (2-5 seconds) since this is built on dev infra and not yet optimized for live systems. Please be patient!
    
    ## More Info

    - [Give us feedback](https://forms.gle/BHpt6BpH2utaf4ez9) to help us improve
    - [Read our FAQ](https://docs.google.com/document/d/1Cp3GJ5k2I-xWZ4GK9WI7Xv8TpKdHmjJ3E9RbzP5Cc_Y/edit) to learn more about CoPE
    - [Join our mailing list](https://forms.gle/PCABrZdhTuXE9w9ZA) to keep in touch
    """)
    
    submit_btn.click(predict, inputs=[content_input, policy_input], outputs=output)

# Launch the app
iface.launch()