README.md

metadata

license: mit
language:
  - en
tags:
  - resume-screening
  - ats
  - recruitment-ai
  - nlp
  - bert
  - text-classification
  - job-matching
  - skill-extraction
datasets:
  - HRPBloom/hf-hrpbloom
metrics:
  - accuracy
  - f1
  - precision
  - recall
library_name: transformers
base_model:
  - Qwen/Qwen3.5-9B
new_version: zai-org/GLM-5
title: Papan-Pemuka
sdk: gradio
emoji: 📉
colorFrom: yellow
colorTo: red
short_description: Papan-Pemuka
sdk_version: 6.13.0

Resume Screening & Job Matching Model (hf-hrpbloom)

This model fine‑tunes Qwen/Qwen3.5-9B – a powerful 9‑billion parameter large language model – for HR‑specific tasks. It is designed to automate and enhance resume screening, candidate‑job matching, and skill extraction.

Model Description

• Developed by: HRPBloom
• Model type: Causal language model fine‑tuned for text classification / sequence classification
• Base model: Qwen/Qwen3.5-9B
• Language: English (en)
• License: MIT
• Fine‑tuned on: Custom HR dataset HRPBloom/hf-hrpbloom (see below)

Intended Uses & Limitations

Primary Use Cases

• Resume‑job description matching – predict a compatibility score or class (e.g., “high match”, “low match”).
• Skill extraction – identify technical and soft skills from free‑text resumes.
• Job title classification – infer the most appropriate job role from a resume.
• ATS (Applicant Tracking System) enhancement – reduce manual screening time.

Out‑of‑Scope

• Making final hiring decisions – always involve human judgment.
• Evaluating non‑textual factors (e.g., cultural fit, interpersonal skills not reflected in writing).
• Cross‑language support (currently English only).

Training Data

The model was fine‑tuned on the dataset HRPBloom/hf-hrpbloom. This dataset comprises:

• Resumes (anonymized) from various industries (IT, finance, healthcare, etc.).
• Corresponding job descriptions with annotated match labels.
• Skill annotations (both technical and soft skills) for named entity recognition tasks.

The dataset was split 80/10/10 for training, validation, and testing.

Training Procedure

Preprocessing

• Resumes and job descriptions were converted to plain text (PDF parsing via pypdf).
• Text was truncated to 2048 tokens (Qwen’s context window supports up to 32k, but we limited for efficiency).
• For classification tasks, we used the model’s hidden state corresponding to the [CLS] token (or the last token) fed into a linear classifier head.

Hyperparameters

• Learning rate: 1e‑5
• Batch size: 8 (per device) with gradient accumulation steps = 4
• Epochs: 3
• Optimizer: AdamW (β1=0.9, β2=0.999)
• Weight decay: 0.01
• Warmup steps: 500
• LR scheduler: linear decay

Hardware

• GPU: 4 × NVIDIA A100 (80 GB)
• Training time: ~8 hours
• Framework: PyTorch 2.2 + Transformers 4.40 + DeepSpeed (ZeRO‑3)

Evaluation Results

The model was evaluated on the held‑out test set. Below are the macro‑averaged metrics for the resume‑job matching classification task (3 classes: low/medium/high match).

Metric	Value
Accuracy	0.91
F1 (macro)	0.89
Precision	0.90
Recall	0.88

Per‑class performance

Match Level	Precision	Recall	F1
Low	0.88	0.85	0.86
Medium	0.90	0.92	0.91
High	0.92	0.88	0.90

For skill extraction (treated as token classification), we achieved:

• Entity‑level F1: 0.87
• Precision: 0.89
• Recall: 0.86

How to Use

### With the `transformers` pipeline

from transformers import pipeline

# Load the fine‑tuned model (auto‑detects task)
classifier = pipeline("text-classification", model="HRPBloom/hf-hrpbloom")

# Example: match a resume snippet with a job description
resume = "5 years of Python development, team leadership, agile methodology."
job_desc = "Looking for a Senior Python Developer with leadership experience."

# Combine inputs (the model expects a single text; you can format as needed)
input_text = f"Resume: {resume}\nJob: {job_desc}"
result = classifier(input_text)
print(result)
# [{'label': 'HIGH_MATCH', 'score': 0.92}]

Direct model inference

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("HRPBloom/hf-hrpbloom")
model = AutoModelForSequenceClassification.from_pretrained("HRPBloom/hf-hrpbloom")

inputs = tokenizer(input_text, return_tensors="pt", truncation=True, max_length=2048)
with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.softmax(logits, dim=-1)

Using the Hugging Face Inference API

import requests

API_URL = "https://api-inference.huggingface.co/models/HRPBloom/hf-hrpbloom"
headers = {"Authorization": "Bearer YOUR_HF_TOKEN"}

def query(payload):
    response = requests.post(API_URL, headers=headers, json=payload)
    return response.json()

output = query({"inputs": input_text})

Environmental Impact

• Carbon emissions: estimated 15 kg CO₂ (based on ML CO2 Impact using 4 × A100 for 8 hours).
• Compute region: Europe (Paris) – low‑carbon energy mix.

Technical Specifications

• Framework: PyTorch 2.2, Transformers 4.40, DeepSpeed
• Hardware: 4 × NVIDIA A100 (80 GB)
• Model size: ~18 GB (9B parameters, half‑precision)
• Inference speed: ~0.5 seconds per example on a T4 GPU

Citation

If you use this model in your research or product, please cite:

@misc{hrpbloom2025,
  author = {HRPBloom},
  title = {Resume Screening and Job Matching Model (hf-hrpbloom)},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face Hub},
  howpublished = {\url{https://huggingface.co/HRPBloom/hf-hrpbloom}}
}

Additional Information

• Demo: Try it live on our Hugging Face Space (coming soon).
• Contact: For questions or collaborations, reach out via GitHub Issues.
• New version: An even more efficient model (zai-org/GLM-5) is available for comparison.
• Contributions: We welcome feedback and contributions – open an issue or PR!

---

Last updated: March 2026

An example chatbot using Gradio, huggingface_hub, and the Hugging Face Inference API.