decentralized-ats-distilbert-2026

Decentralized ATS Token Classifier (DistilBERT-2026-Quantized)

A publication-grade, highly optimized, and quantized DistilBERT-base-uncased model fine-tuned for high-sensitivity Named Entity Recognition (NER) in Applicant Tracking Systems (ATS). This model is designed for edge infrastructure deployment, ensuring zero privacy leakage via integrated GDPR compliance masking and zero structural degradation under INT8 ONNX graph quantization.

---

🚀 Key Performance Matrix (Empirical Evidence)

---

🛠️ Architecture & Design Specifications

To deploy or test this model properly, the downstream inference engine must respect two core design parameters used during the model's training phase:

1. GDPR Masking Preprocessing: Personal Identifiable Information (PII) like emails, phone numbers, ages, and pronouns must be programmatically masked to align with the training vocabulary.
2. Sliding Window Token Chunks: Long documents must be chunked into blocks of max_length=256 tokens with a stride=64 token overlap to bridge contextual semantic breaks across document cuts.

---

📦 Getting Started & Usage Guide

Because this model was quantized into external multi-nested parameters (model_quantized.onnx + model_quantized.onnx.data), you must load it using ONNX Runtime rather than standard PyTorch layout transformers.

Step 1: Install Dependencies

pip install onnxruntime pypdf python-docx transformers huggingface_hub numpy

Step 2: Full End-to-End Production Script

Create a Python file (e.g., app.py) or run this in your server environment to extract skills and match files (PDF/DOCX) against Job Descriptions.

import os
import re
import json
import numpy as np
import onnxruntime as ort
from transformers import AutoTokenizer
from huggingface_hub import hf_hub_download
from pypdf import PdfReader
from docx import Document

# 1. DOWNLOAD SYSTEM GRAPH LAYERS FROM HF
REPO_ID = "sanaullah7964/decentralized-ats-distilbert-2026"
LOCAL_MODEL_DIR = "./downloaded_production_model"
os.makedirs(LOCAL_MODEL_DIR, exist_ok=True)

required_files = [
    "config.json", 
    "tokenizer.json", 
    "tokenizer_config.json", 
    "model_quantized.onnx",
    "model_quantized.onnx.data"
]

print("Downloading model binaries from Hugging Face Hub...")
for file_name in required_files:
    hf_hub_download(repo_id=REPO_ID, filename=file_name, local_dir=LOCAL_MODEL_DIR)

# Initialize Session Engine
tokenizer = AutoTokenizer.from_pretrained(LOCAL_MODEL_DIR)
onnx_path = os.path.join(LOCAL_MODEL_DIR, "model_quantized.onnx")
session = ort.InferenceSession(onnx_path, providers=['CPUExecutionProvider'])

# 2. DOCUMENT TEXT EXTRACTION PARSERS
def extract_text_from_file(file_path):
    if file_path.endswith('.pdf'):
        reader = PdfReader(file_path)
        return "\n".join([page.extract_text() for page in reader.pages if page.extract_text()])
    elif file_path.endswith('.docx'):
        doc = Document(file_path)
        return "\n".join([para.text for para in doc.paragraphs])
    return ""

# 3. GDPR MASKING ENGINE
class GDPRMasker:
    def __init__(self):
        self.email_regex = re.compile(r'[\w\.-]+@[\w\.-]+\.\w+')
        self.phone_regex = re.compile(r'\+?\d{1,4}?[-.\s]?\(?\d{1,3}?\)?[-.\s]?\d{1,4}[-.\s]?\d{1,4}[-.\s]?\d{1,9}')
    def clean(self, text):
        text = self.email_regex.sub("[MASKED_EMAIL]", text)
        return self.phone_regex.sub("[MASKED_PHONE]", text)

masker = GDPRMasker()

# 4. NEURAL INFERENCE GRAPH PIPELINE
def extract_skills(text, max_len=256, stride=64):
    cleaned_text = masker.clean(text)
    tokenized = tokenizer(
        cleaned_text, max_length=max_len, stride=stride, truncation=True,
        return_overflowing_tokens=True, return_offsets_mapping=True,
        padding="max_length", return_tensors="np"
    )
    
    extracted_skills = set()
    model_inputs = [node.name for node in session.get_inputs()]
    
    for chunk_idx in range(len(tokenized["input_ids"])):
        input_ids = tokenized["input_ids"][chunk_idx]
        attention_mask = tokenized["attention_mask"][chunk_idx]
        
        onnx_feed = {
            "input_ids": np.expand_dims(input_ids, axis=0).astype(np.int64),
            "attention_mask": np.expand_dims(attention_mask, axis=0).astype(np.int64)
        }
        onnx_feed = {k: v for k, v in onnx_feed.items() if k in model_inputs}
        
        onnx_outputs = session.run(["logits"], onnx_feed)
        logits = np.squeeze(np.array(onnx_outputs), axis=(0, 1))
        predictions = np.argmax(logits, axis=-1).tolist()
        
        tokens = tokenizer.convert_ids_to_tokens(input_ids)
        current_phrase = []
        
        for token, pred_id in zip(tokens, predictions):
            if token in tokenizer.all_special_tokens:
                continue
            is_subword = token.startswith("##")
            token_clean = token[2:] if is_subword else token
            
            if int(pred_id) == 1: # B-SKILL
                if current_phrase:
                    skill_str = "".join(current_phrase).strip().lower().replace("mer n", "mern").replace("dock er", "docker").replace("aw s", "aws")
                    if len(skill_str) > 1 and not skill_str.startswith("[masked"):
                        extracted_skills.add(skill_str)
                current_phrase = [token_clean]
            elif int(pred_id) == 2 and current_phrase: # I-SKILL
                if is_subword:
                    current_phrase.append(token_clean)
                else:
                    current_phrase.append(" " + token_clean)
            else:
                if current_phrase:
                    skill_str = "".join(current_phrase).strip().lower().replace("mer n", "mern").replace("dock er", "docker").replace("aw s", "aws")
                    if len(skill_str) > 1 and not skill_str.startswith("[masked"):
                        extracted_skills.add(skill_str)
                    current_phrase = []
                    
        if current_phrase:
            skill_str = "".join(current_phrase).strip().lower().replace("mer n", "mern").replace("dock er", "docker").replace("aw s", "aws")
            if len(skill_str) > 1 and not skill_str.startswith("[masked"):
                extracted_skills.add(skill_str)
                
    return extracted_skills

# 5. EXECUTE THE COMPARATIVE RUN
job_description = "Seeking a Senior MERN Stack developer proficient in Python programming, SQL databases, and AWS."
resume_path = "sample_resume.docx" # Drop your local file here

if os.path.exists(resume_path):
    resume_text = extract_text_from_file(resume_path)
    
    # Extract structural sets
    resume_skills = extract_skills(resume_text)
    job_skills = extract_skills(job_description)
    
    # Substring parsing validation matrix loop
    matched_skills = []
    for required_skill in job_skills:
        if any(required_skill in candidate_skill for candidate_skill in resume_skills):
            matched_skills.append(required_skill)
            
    missing_skills = [sk for sk in job_skills if sk not in matched_skills]
    score = (len(matched_skills) / len(job_skills)) * 100 if job_skills else 0.0
    
    print(f"Match Score: {score:.2f}%")
    print(f"Matched Capabilities: {matched_skills}")
    print(f"Missing Tech Gaps: {missing_skills}")
else:
    print(f"Please drop a valid file at '{resume_path}' to execute calculation matrices.")

---

⚖️ Citation & Academic Reference

If you utilize this model or architecture topology within an academic study, research diagnostic report, or open-source software project, please use the following reference format:

@article{sanaullah2026decentralized,
  title={Decentralized ATS Token Extraction Framework: High-Sensitivity Token Classification via Quantized Edge DistilBERT Graphs},
  author={Sanaullah et al.},
  journal={Hugging Face Model Hub Scopes},
  year={2026},
  url={sanaullah7964/decentralized-ats-distilbert-2026}
}