Lab_hr / app.py
Tushar14's picture
Update app.py
c0ec921 verified
Raw History Blame Contribute Delete
5.46 kB
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 β€” 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 β€” 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 = "πŸ¦™ Llama-3.2-1B" if HF_TOKEN else "πŸ“‹ Keyword rules"
demo = gr.Interface(
fn=analyze_file,
inputs=gr.File(label="Upload Job Posts Excel (.xlsx)"),
outputs=[
gr.Dataframe(label=f"Results β€” {backend_label}"),
gr.File(label="Download results.xlsx"),
],
title="πŸ” AI Job Post Discrimination Checker",
description=(
"Upload the job posts Excel file. "
"Detects: **Gender Β· Age Β· Origin Β· Mobility Β· Salary Β· Standing Β· PPE Β· Role clarity**\n\n"
f"Active backend: `{backend_label}`"
),
allow_flagging="never",
)
demo.launch(server_name="0.0.0.0", server_port=7860)