File size: 5,462 Bytes
df53b4b
 
 
 
8b3404a
d905070
fccf350
df53b4b
 
 
0fbb17d
c0ec921
df53b4b
 
 
 
 
 
 
 
 
 
f5224a3
df53b4b
 
f5224a3
 
df53b4b
 
 
 
 
c0ec921
df53b4b
 
f5224a3
df53b4b
f5224a3
 
 
 
 
 
 
 
df53b4b
 
f5224a3
df53b4b
f5224a3
 
 
 
 
 
 
 
df53b4b
 
 
 
 
f5224a3
 
 
df53b4b
 
 
 
 
f5224a3
 
 
 
df53b4b
f5224a3
df53b4b
f5224a3
df53b4b
fccf350
aa36863
d905070
aa36863
 
 
df53b4b
d905070
f5224a3
d1f1b50
df53b4b
 
 
d1f1b50
 
f5224a3
d1f1b50
df53b4b
 
 
 
 
f5224a3
8b2be61
df53b4b
 
d905070
df53b4b
 
c741d40
c0ec921
df53b4b
c741d40
f5224a3
 
aa36863
f0fc1b5
df53b4b
c0ec921
d905070
bdb15a7
 
df53b4b
 
f5224a3
df53b4b
 
 
 
 
f5224a3
df53b4b
 
 
bdb15a7
8b3404a
df53b4b
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
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)