mithinsagar/exai-resumeintel-data

EXAI-ResumeIntel Datasets

Datasets supporting EXAI-ResumeIntel, an explainable artificial intelligence framework for automated resume analysis using Shapley values, LIME, and a hierarchical domain skill ontology.

Author: Mithin Sagar S · github.com/mithinsagar Institution: Vellore Institute of Technology (VIT), Vellore, Tamil Nadu, India Code repository: github.com/mithinsagar/EXAI-ResumeIntel Companion models: mithinsagar/exai-resumeintel-models

Overview

This repository contains the two datasets used to train and evaluate the EXAI-ResumeIntel framework. Together they provide both the candidate side (real resumes across 24 occupational categories) and the employer side (over one million job postings with extracted skill features) of the resume-to-role matching problem.

Files

File	Rows	Size	Description
clean_resume_data.csv	2,484	12.2 MB	Real-world resumes across 24 job categories
jobs_dataset_with_features.csv	1,000,000+	637.6 MB	Job postings with extracted skill feature strings

Dataset 1: clean_resume_data.csv

The primary training corpus for the semantic embedding engine and the supervised classification benchmark.

Schema

Column	Type	Description
ID	int64	Unique resume identifier
Category	string	Occupational category label (24 distinct values)
Feature	string	Preprocessed resume text

Categories (24)

INFORMATION-TECHNOLOGY, BUSINESS-DEVELOPMENT, FINANCE, ADVOCATE, ACCOUNTANT, ENGINEERING, CHEF, AVIATION, FITNESS, SALES, BANKING, HEALTHCARE, CONSULTANT, CONSTRUCTION, PUBLIC-RELATIONS, HR, DESIGNER, ARTS, TEACHER, APPAREL, DIGITAL-MEDIA, AGRICULTURE, AUTOMOBILE, BPO

Statistics

• Mean resume length: 587 words
• Median resume length: 549 words
• Range: 77 to 3,565 words
• Mean samples per category: 104
• Most categories contain 110–120 samples
• Under-represented categories: BPO (22 samples), AUTOMOBILE (36 samples)

Known issues

One NaN entry exists in the Feature column at row 656 (BUSINESS-DEVELOPMENT). The reference implementation fills this with an empty string before training.

Dataset 2: jobs_dataset_with_features.csv

Employer-side ground truth used to build weighted role profiles and to power semantic role matching in the interactive application.

Schema

Column	Type	Description
Role	string	Job title / role name
Features	string	Extracted skill and requirement text

A 50,000-row sample is used for exploratory analysis in the accompanying paper. The most frequent roles in that sample include Interaction Designer (646 postings), Network Administrator (502), and User Interface Designer (483), confirming the breadth of coverage across technical and non-technical occupations.

Usage

Download via the Hugging Face CLI

hf download mithinsagar/exai-resumeintel-data \
  --repo-type dataset --local-dir data/raw

Load with pandas

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="mithinsagar/exai-resumeintel-data",
    filename="clean_resume_data.csv",
    repo_type="dataset",
)
df = pd.read_csv(path)
df["Feature"] = df["Feature"].fillna("")
print(f"{len(df):,} resumes across {df['Category'].nunique()} categories")

Load the large job postings file in chunks

path = hf_hub_download(
    repo_id="mithinsagar/exai-resumeintel-data",
    filename="jobs_dataset_with_features.csv",
    repo_type="dataset",
)
for chunk in pd.read_csv(path, chunksize=50_000):
    process(chunk)

Benchmark Results

Supervised classification on clean_resume_data.csv using TfidfVectorizer (40,000 features, 1–3 grams, sublinear TF) with LinearSVC (C = 2.0) under 5-fold stratified cross-validation:

Metric	Value
Overall accuracy	70.73%
Macro precision	69.37%
Macro recall	66.45%
Macro F1-score	66.05%
Weighted F1-score	69.42%

ROC-AUC scores (one-vs-rest, calibrated SVM) range from 0.9313 (ADVOCATE) to 0.9925 (HR).

Off-diagonal confusions concentrate between semantically adjacent category pairs — Finance/Banking, Advocate/Consultant, Arts/Designer — reflecting genuine vocabulary overlap rather than model error.

Intended Use

These datasets are released to support reproducible research in explainable resume screening, skill ontology construction, and fair automated hiring systems. They are suitable for text classification, semantic similarity, feature attribution, and counterfactual explanation research.

Limitations and Ethical Considerations

Automated resume screening systems can encode and amplify hiring biases present in historical data. This corpus reflects the composition of its source and should not be treated as a representative sample of any labour market. Category support is uneven, with BPO and AUTOMOBILE substantially under-represented, which caps achievable per-category accuracy.

Any deployment of models trained on this data in a real hiring context should be accompanied by bias auditing, human review of all adverse decisions, and compliance with applicable employment and algorithmic transparency law. The EXAI framework was designed specifically so that every score can be audited and explained; that capability should be used, not bypassed.

Citation

@software{sagar2026exai,
  author    = {Mithin Sagar S},
  title     = {{EXAI-ResumeIntel: An Explainable Artificial Intelligence
               Framework for Automated Resume Analysis Using Shapley
               Values, LIME, and Domain Skill Ontology}},
  year      = {2026},
  publisher = {GitHub},
  url       = {https://github.com/mithinsagar/EXAI-ResumeIntel}
}

License

MIT