import os import re import json import time import gradio as gr import pandas as pd import requests from discrimination_model import detect_discrimination # ── LLM backend selector ───────────────────────────────────────────────────── HF_TOKEN = os.environ.get("HF_TOKEN", "") SYSTEM_PROMPT = """You are an expert HR compliance analyst. Analyse the job posting and identify discriminatory elements from this list: - Gender: gendered job titles or language implying a preferred gender - Age: explicit age limits or age-suggestive language - Origin: nationality, ethnicity, or geographic residency requirements - Mobility: unjustified driving licence or own vehicle requirement - Salary: undisclosed or vague compensation (e.g. "to be defined", "commensurate with experience") - Standing: appearance, weight, height, or "professional look" requirements - PPE: candidate required to supply own safety equipment - Role clarity: vague or contradictory role description - None: if fully compliant Return ONLY valid JSON — no markdown fences, no extra text: {"discriminatory_elements": ["Element1", "Element2"], "explanation": "brief reason per element"}""" def llm_detect(text: str) -> tuple[str, str]: """Call LLM backend. Returns (categories_str, explanation_str).""" # ── Option A: Llama-3.2-1B-Instruct via HF Inference API ──────────────── if HF_TOKEN: try: # Llama-3 chat template prompt = ( "<|begin_of_text|>" "<|start_header_id|>system<|end_header_id|>\n" f"{SYSTEM_PROMPT}" "<|eot_id|>" "<|start_header_id|>user<|end_header_id|>\n" f"Job posting:\n\n{text}" "<|eot_id|>" "<|start_header_id|>assistant<|end_header_id|>\n" ) resp = requests.post( "https://api-inference.huggingface.co/models/meta-llama/Llama-3.2-1B-Instruct", headers={"Authorization": f"Bearer {HF_TOKEN}"}, json={ "inputs": prompt, "parameters": { "max_new_tokens": 300, "temperature": 0.1, "return_full_text": False, }, }, timeout=60, ) resp.raise_for_status() output = resp.json() generated = output[0]["generated_text"] if isinstance(output, list) else str(output) # Pull the first JSON object out of the response match = re.search(r'\{.*?\}', generated, re.DOTALL) if match: data = json.loads(match.group()) elements = data.get("discriminatory_elements", ["None"]) explanation = data.get("explanation", "") return ", ".join(elements), explanation # If model didn't return valid JSON, surface the raw text for debug return "Parsing error", generated[:300] except Exception as e: return "Error (Llama)", str(e) # ── Fallback: keyword rules (no API key needed) ─────────────────────────── return detect_discrimination(text) def analyze_file(file): if file is None: return None, None df = pd.read_excel(file) # Accept either column name used in the exercise file text_col = None for candidate in ["Job post text", "job_post", "text", "job post"]: if candidate in df.columns: text_col = candidate break if text_col is None: raise ValueError("Could not find job text column. Expected: 'Job post text' or 'job_post'") categories, explanations = [], [] for text in df[text_col]: cat, exp = llm_detect(str(text)) categories.append(cat) explanations.append(exp) time.sleep(0.3) # gentle rate-limit buffer df["Discrimination_Type"] = categories df["Explanation"] = explanations output_path = "/tmp/results.xlsx" df.to_excel(output_path, index=False) backend = "Llama-3.2-1B" if HF_TOKEN else "Keyword rules" print(f"Analysis complete — backend: {backend}, rows: {len(df)}") id_col = "Job post ID" if "Job post ID" in df.columns else df.columns[0] return df[[id_col, text_col, "Discrimination_Type", "Explanation"]], output_path # ── Gradio UI ───────────────────────────────────────────────────────────────── backend_label = "🦙 Llama-3.2-1B" if HF_TOKEN else "📋 Keyword rules" demo = gr.Interface( fn=analyze_file, inputs=gr.File(label="Upload Job Posts Excel (.xlsx)"), outputs=[ gr.Dataframe(label=f"Results — {backend_label}"), gr.File(label="Download results.xlsx"), ], title="🔍 AI Job Post Discrimination Checker", description=( "Upload the job posts Excel file. " "Detects: **Gender · Age · Origin · Mobility · Salary · Standing · PPE · Role clarity**\n\n" f"Active backend: `{backend_label}`" ), allow_flagging="never", ) demo.launch(server_name="0.0.0.0", server_port=7860)