# This is a project of Zentropi Inc. All rights reserved. import gradio as gr import os import torch import torch.nn.functional as F 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-2-9b" adapter_model_name = "zentropi-ai/cope-a-9b" bnb_config = BitsAndBytesConfig( load_in_8bit=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 ## Overview This policy is designed to determine whether or not content is hate speech. ## Definition of Terms - Protected Classes: Individuals or groups distinguished by race (e.g. Black, White), ethnicity (e.g. Hispanic, Jewish), nationality (e.g. Canadian, Argentinian), caste (e.g. Brahmins, Kshatriyas, Vaishyas), immigration status (e.g. Immigrants), religion (e.g. Catholics, Jews), disability (e.g. blind people, deaf people), disease status (e.g. people with AIDS), sexual orientation (e.g. gays, bisexuals), transgender status (e.g. trans people, cis people), gender identity (e.g. women, men), gender expression (e.g. masculine, feminine), or sex (e.g. male, female). - Derogatory Slur: Any term or phrase used pejoratively or contemptuously towards a member or members of a protected class, including those historically reclaimed but used in a derogatory context. - Non-Protected Classes: Categories defined by, political party membership (e.g. Democrats, Tories), political beliefs (e.g. fascists, communists), hobbies (e.g. games, knitters), criminal history (e.g. criminals, convicts), or profession (e.g. doctors, farmers), or weight or body size (e.g. fat people, thin people). ## Interpretation of Language: - In interpreting this policy, ambiguous phrases should be understood based on their most reasonable contextual meaning. For instance, criticisms directed at a religious system (e.g. "Buddhism is a flawed ideology") should NOT be construed as attacks against the adherents of that faith. ## Definition of Labels ### (HS): Hate Speech #### Includes - Direct and Indirect Violence Advocacy: Any form of expression that celebrates, justifies, incites, or threatens physical or psychological harm against individuals or groups identified by their membership in a protected class. - Dehumanization and Vilification: Content suggesting that individuals or groups, explicitly identified by their membership in a specific protected class, are subhuman, inherently lesser beings, or using animalistic descriptions in a way that promotes disdain or hate. - Derogatory and Dehumanizing Language: Use of slurs, epithets, or any derogatory language aimed at belittling, humiliating, or inciting hatred against individuals or groups explicitly identified by their membership in a specific protected class. - Explicit and Implicit Discrimination Advocacy: Promoting exclusion, segregation, or denial of rights against individuals or groups explicitly identified by their membership in a specific protected class. - Collective Attribution of Negative Actions: Assigning collective blame or advocating collective punishment based on the actions or perceived characteristics of individuals or groups identified by their membership in a protected class. - Inferiority and Superiority Claims: Statements that categorically assign inferiority or superiority, moral or intellectual, to individuals or groups identified by their membership in a protected class. - Denial or Distortion of Historical Atrocities: Denying, grossly trivializing, or distorting documented atrocities against groups identified by their membership in a protected class, undermining their significance or the suffering of their members. - Conspiracy Theories: Propagating unfounded allegations that individuals or groups, identified by their membership in a protected class, are responsible for serious harms or controlling significant institutions to the detriment of society. #### Excludes - Attacks on Non-Protected Classes: Content that attacks or criticizes individuals or groups identified by their membership in a Non-Protected Class, EVEN if that attack is violent, threatening, or otherwise hateful (e.g. "Criminals should all be rounded up and shot!"). - Criticism of Beliefs and Institutions: Constructive critique or discussion of political ideologies, religious doctrines, or institutions without resorting to hate speech or targeting individuals or groups identified by their membership in a protected class. - Attacking Leaders: Content that critiques, mocks, or insults the leaders of religions, leaders of religious institutions, or religious prophets or deities, BUT does not contain the singular or plural noun for followers of that religion. - Condemning Violent Extremism: Content that condemns, mocks, insults, dehumanizes, or calls for violence against terrorist organizations and violent hate groups, or their members. - Neutrally Reporting Historical Events: Neutrally and descriptively reporting or discussion of factual events in the past that could be construed as negative about individuals or groups identified by their membership in a protected class. - Pushing Back on Hateful Language: Content where the writer pushes back on, condemns, questions, criticizes, or mocks a different person's hateful language or ideas. - Disease Discussion: Content in which the author discusses diseases without direct references to people with the disease. - 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. """ DEFAULT_CONTENT = "Put your content sample here." # 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) # Get logits for the last token logits = outputs.logits[:, -1, :] # Apply softmax to get probabilities probabilities = F.softmax(logits, dim=-1) # Get the predicted token ID predicted_token_id = torch.argmax(logits, dim=-1).item() # Decode the predicted token decoded_output = tokenizer.decode([predicted_token_id]) if decoded_output == '1': return f'True (i.e., Meets Label Criteria)' else: return f'False (i.e., Does NOT Meet Label Criteria)' # Create the interface with gr.Blocks() as demo: gr.Markdown("# Zentropi CoPE Demo") with gr.Row(): # Left column with inputs with gr.Column(scale=1): input1 = gr.Textbox(label="Content (english only)", lines=2, max_lines=4, value=DEFAULT_CONTENT) input2 = gr.Textbox(label="Policy (try your own)", lines=11, max_lines=19, value=DEFAULT_POLICY) # Right column with output with gr.Column(scale=1): output = gr.Textbox(label="Result") submit_btn = gr.Button("Submit") notes = gr.Markdown(""" ## About CoPE [CoPE](https://huggingface.co/zentropi-ai/cope-a-9b) (the COntent Policy Evaluator) is a small language model capable of accurate content policy labeling. This is a **demo** of our initial release and should **NOT** be used for any production use cases. See full [model card](https://huggingface.co/zentropi-ai/cope-a-9b) for details. ## 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. ## More Info - [Read our FAQ](https://docs.google.com/document/d/1Cp3GJ5k2I-xWZ4GK9WI7Xv8TpKdHmjJ3E9RbzP5Cc_Y/edit) to learn more about CoPE - [Give us feedback](https://forms.gle/BHpt6BpH2utaf4ez9) to help us improve - [Join our mailing list](https://forms.gle/PCABrZdhTuXE9w9ZA) to keep in touch - [Contact us](https://forms.gle/PCABrZdhTuXE9w9ZA) for info about our partner program """) # Set up the processing function submit_btn.click( fn=predict, inputs=[input1, input2], outputs=output, api_name=False ) # Launch the interface demo.launch(show_api=False)