|
|
import os |
|
|
|
|
|
|
|
|
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "0" |
|
|
|
|
|
import gradio as gr |
|
|
from transformers import pipeline |
|
|
from content import LEARN_CONTENT, HOW_IT_WORKS_CONTENT |
|
|
|
|
|
VERSION = "1.1.0" |
|
|
|
|
|
|
|
|
print("🚀 Initializing R.A.I.C Detection System...") |
|
|
try: |
|
|
classifier = pipeline( |
|
|
"zero-shot-classification", |
|
|
model="valhalla/distilbart-mnli-12-1", |
|
|
device=-1 |
|
|
) |
|
|
print("✅ R.A.I.C Detection System Ready!") |
|
|
except Exception as e: |
|
|
print(f"❌ Failed to load classifier: {e}") |
|
|
classifier = None |
|
|
|
|
|
|
|
|
GROUP_LABELS = { |
|
|
"fairness_violation": [ |
|
|
"biased", "discriminatory", "sexist", "racist", "prejudiced", |
|
|
"stereotyping", "unfair", "exclusionary", "ageist", "homophobic", |
|
|
"transphobic", "xenophobic", "classist", "ableist", "bigoted" |
|
|
], |
|
|
"safety_violation": [ |
|
|
"harmful", "dangerous", "toxic", "violent", "threatening", |
|
|
"abusive", "self-harm", "suicide", "weapons", "explosives", |
|
|
"illegal instructions", "criminal activity", "unsafe", "destructive" |
|
|
], |
|
|
"privacy_violation": [ |
|
|
"asks for personal information", "requests private data", "solicits PII", |
|
|
"asks for passwords", "requests credentials", "invasive", |
|
|
"data harvesting", "surveillance", "doxxing", "stalking", |
|
|
"identity theft", "financial information" |
|
|
], |
|
|
"inclusiveness_violation": [ |
|
|
"exclusionary", "elitist", "gatekeeping", "dismissive", |
|
|
"marginalizing", "isolating", "alienating", "segregating", |
|
|
"privileged", "inaccessible", "unwelcoming" |
|
|
], |
|
|
"transparency_violation": [ |
|
|
"ambiguous", "unclear", "vague", "deceptive", "misleading", |
|
|
"confusing", "obfuscated", "hidden agenda", "manipulative", |
|
|
"disinformation", "misinformation", "propaganda", "fake" |
|
|
], |
|
|
"accountability_violation": [ |
|
|
"jailbreaking", "prompt injection", "system bypass", "guardrail bypass", |
|
|
"AI manipulation", "restriction circumvention", "unethical use", |
|
|
"misuse", "abuse of technology", "irresponsible", "reckless" |
|
|
], |
|
|
"responsible_and_safe": [ |
|
|
"safe", "ethical", "responsible", "appropriate", "constructive", |
|
|
"helpful", "educational", "legitimate", "transparent", "fair", |
|
|
"inclusive", "respectful", "beneficial", "trustworthy" |
|
|
] |
|
|
} |
|
|
|
|
|
ALL_CANDIDATE_LABELS = [label for labels in GROUP_LABELS.values() for label in labels] |
|
|
HYPOTHESIS_TEMPLATE = os.getenv("HYPOTHESIS_TEMPLATE", "The prompt is {}.") |
|
|
VIOLATION_THRESHOLD = 0.3 |
|
|
BIAS_KEYWORDS = ["are black", "ares black", "are white", "are stupid", "south indian"] |
|
|
|
|
|
CATEGORY_NAMES = { |
|
|
"fairness_violation": "Bias/Discrimination", |
|
|
"safety_violation": "Safety Risk", |
|
|
"privacy_violation": "Privacy Issue", |
|
|
"inclusiveness_violation": "Exclusion Risk", |
|
|
"transparency_violation": "Clarity Issue", |
|
|
"accountability_violation": "Misuse Risk", |
|
|
"responsible_and_safe": "Responsible & Safe" |
|
|
} |
|
|
|
|
|
def get_confidence_level(score): |
|
|
"""Determine confidence level from score""" |
|
|
if score > 0.7: |
|
|
return "High Confidence" |
|
|
elif score > 0.5: |
|
|
return "Medium Confidence" |
|
|
return "Low Confidence" |
|
|
|
|
|
def build_scores_display(sorted_scores): |
|
|
"""Build visual score display with bars""" |
|
|
lines = ["\n**All Category Scores:**"] |
|
|
for cat, score in sorted_scores: |
|
|
name = CATEGORY_NAMES.get(cat, cat) |
|
|
bar = "█" * int(score * 10) |
|
|
lines.append(f"- {name}: {score:.1%} {bar}") |
|
|
return "\n".join(lines) + "\n" |
|
|
|
|
|
def build_methodology(confidence_level): |
|
|
"""Build methodology explanation""" |
|
|
return ( |
|
|
"\n**How This Was Calculated:**\n" |
|
|
"- Method: ML-based zero-shot classification\n" |
|
|
"- Model: valhalla/distilbart-mnli-12-1\n" |
|
|
"- Process: Evaluated against 70+ semantic labels across 6 categories\n" |
|
|
f"- Decision: Highest scoring category with {confidence_level}\n" |
|
|
) |
|
|
|
|
|
def build_feedback(risk_level, name, score, confidence_level, sorted_scores, violations_count=0): |
|
|
"""Build structured feedback message""" |
|
|
risk_config = { |
|
|
"HIGH": "Review your prompt for potential issues", |
|
|
"MODERATE": "Review your prompt for potential issues", |
|
|
"LOW": "Consider reviewing your prompt for potential issues" |
|
|
} |
|
|
|
|
|
recommendation = risk_config.get(risk_level, "Review as needed") |
|
|
|
|
|
feedback = f"**{risk_level} RISK: {name}**\n\n" |
|
|
feedback += f"Detected with {score:.0%} confidence ({confidence_level})\n\n" |
|
|
feedback += f"**Recommendation:** {recommendation} related to {name.lower()}.\n" |
|
|
|
|
|
if violations_count > 1: |
|
|
feedback += f"\n**Note:** Also found {violations_count - 1} other potential issue(s) above threshold.\n" |
|
|
|
|
|
feedback += build_scores_display(sorted_scores) |
|
|
feedback += build_methodology(confidence_level) |
|
|
feedback += f"\n**Tip:** Check the 'Learn' tab for detailed guidance on {name}." |
|
|
|
|
|
return feedback |
|
|
|
|
|
def audit_prompt(prompt): |
|
|
"""Audit prompts for Responsible AI violations""" |
|
|
if not prompt or not prompt.strip(): |
|
|
return {}, "Please enter a prompt to audit." |
|
|
|
|
|
if classifier is None: |
|
|
return {}, "Model failed to load. Please restart the application." |
|
|
|
|
|
try: |
|
|
result = classifier( |
|
|
prompt, |
|
|
ALL_CANDIDATE_LABELS, |
|
|
multi_label=True, |
|
|
hypothesis_template=HYPOTHESIS_TEMPLATE |
|
|
) |
|
|
|
|
|
candidate_scores = dict(zip(result["labels"], result["scores"])) |
|
|
|
|
|
group_scores = { |
|
|
group: max((candidate_scores.get(syn, 0.0) for syn in synonyms), default=0.0) |
|
|
for group, synonyms in GROUP_LABELS.items() |
|
|
} |
|
|
|
|
|
prompt_lower = prompt.lower() |
|
|
for keyword in BIAS_KEYWORDS: |
|
|
if keyword in prompt_lower: |
|
|
group_scores["fairness_violation"] = max(group_scores["fairness_violation"], 0.8) |
|
|
print(f"Detected bias keyword: '{keyword}'") |
|
|
break |
|
|
|
|
|
sorted_scores = sorted(group_scores.items(), key=lambda x: x[1], reverse=True) |
|
|
top_category, top_score = sorted_scores[0] |
|
|
confidence_level = get_confidence_level(top_score) |
|
|
|
|
|
if top_category == "responsible_and_safe": |
|
|
feedback = ( |
|
|
"**PROMPT APPROVED**\n\n" |
|
|
"No major violations detected. Your prompt appears to follow responsible AI principles.\n" |
|
|
f"\n**Confidence Level:** {confidence_level}\n" |
|
|
) |
|
|
feedback += build_scores_display(sorted_scores) |
|
|
feedback += build_methodology(confidence_level) |
|
|
return group_scores, feedback |
|
|
|
|
|
violations = [(cat, score) for cat, score in group_scores.items() |
|
|
if cat != "responsible_and_safe" and score > VIOLATION_THRESHOLD] |
|
|
|
|
|
if violations: |
|
|
violations.sort(key=lambda x: x[1], reverse=True) |
|
|
category, score = violations[0] |
|
|
name = CATEGORY_NAMES.get(category, "Issue") |
|
|
|
|
|
if score > 0.7: |
|
|
risk_level = "HIGH" |
|
|
elif score > 0.55: |
|
|
risk_level = "MODERATE" |
|
|
else: |
|
|
risk_level = "LOW" |
|
|
|
|
|
feedback = build_feedback(risk_level, name, score, confidence_level, |
|
|
sorted_scores, len(violations)) |
|
|
else: |
|
|
feedback = ( |
|
|
"**PROMPT APPROVED**\n\n" |
|
|
"No major violations detected (all scores below threshold).\n" |
|
|
f"\n**Confidence Level:** {confidence_level}\n" |
|
|
) |
|
|
feedback += build_scores_display(sorted_scores) |
|
|
feedback += build_methodology(confidence_level) |
|
|
|
|
|
return group_scores, feedback |
|
|
|
|
|
except Exception as e: |
|
|
return {}, f"Error analyzing prompt: {str(e)}" |
|
|
|
|
|
def build_ui(): |
|
|
"""Build Gradio interface with audit and learn tabs""" |
|
|
with gr.Blocks() as demo: |
|
|
gr.Markdown("# R.A.I.C – Responsible AI Coach") |
|
|
|
|
|
with gr.Tabs(): |
|
|
with gr.Tab("Audit Tool"): |
|
|
gr.Markdown("### Audit prompts against key Responsible AI principles") |
|
|
|
|
|
prompt_input = gr.Textbox( |
|
|
lines=3, |
|
|
placeholder="Enter your prompt here...", |
|
|
label="Prompt" |
|
|
) |
|
|
audit_btn = gr.Button("Run Audit", variant="primary") |
|
|
score_output = gr.Label(label="Detection Scores") |
|
|
feedback = gr.Textbox(label="R.A.I.C Feedback", interactive=False, lines=4) |
|
|
|
|
|
audit_btn.click( |
|
|
fn=audit_prompt, |
|
|
inputs=prompt_input, |
|
|
outputs=[score_output, feedback] |
|
|
) |
|
|
|
|
|
with gr.Tab("How It Works"): |
|
|
gr.Markdown(HOW_IT_WORKS_CONTENT) |
|
|
|
|
|
with gr.Tab("Learn Responsible AI"): |
|
|
gr.Markdown(LEARN_CONTENT) |
|
|
|
|
|
return demo |
|
|
|
|
|
app = build_ui() |
|
|
|
|
|
if __name__ == "__main__": |
|
|
app.launch() |
|
|
|
