bert-local: Smart Local Search for Edge AI and IoT

Overview

bert-local is a fine-tuned transformer model built on boltuix/bert-mini for intent classification, designed to interpret natural language queries and suggest 140 local business categories. With a compact ~20 MB size, it achieves 94.26% accuracy and is optimized for edge AI, IoT devices, and mobile applications. Perfect for local search, chatbots, and smart assistants, bert-local delivers real-time, offline-capable solutions.

bert-local redefines local search with precise intent understanding in a lightweight package.

Key Features

• Intent-Driven: Maps queries like “My dog is sick” to 🐾 pet stores or 🩺 clinics.
• High Accuracy: 94.26% test accuracy across 122 cases.
• 140 Categories: Covers businesses from 💼 accounting to 🦒 zoos.
• Edge-Optimized: Fast inference (<50ms) on low-resource devices.
• Offline Ready: No internet required.
• Extensible: Fine-tunable for other NLP tasks.

Use Cases

• Local Search Apps: Suggest 🐾 pet stores or 🩺 clinics for user queries.
• Chatbots: Enhance customer service with local recommendations.
• E-Commerce: Guide users to 💼 accounting firms or 📚 bookstores.
• Travel Apps: Recommend 🏨 hotels or 🗺️ attractions.
• Healthcare: Direct users to 🏥 hospitals or 💊 pharmacies.
• Smart Assistants: Enable hands-free search on IoT devices.

Supported Categories

bert-local supports 140 local business categories, including:

• 💼 Accounting Firm
• ✈️ Airport
• 🎢 Amusement Park
• 🐠 Aquarium
• 🖼️ Art Gallery
• 🥐 Bakery
• 🏦 Bank
• 🍻 Bar
• 💈 Barber Shop
• 🏖️ Beach
• ...and more (full list via Hugging Face)

Extract categories programmatically:

from transformers import AutoModelForSequenceClassification

# Load model
model = AutoModelForSequenceClassification.from_pretrained("boltuix/bert-local")

# Extract labels
label_mapping = model.config.id2label
supported_labels = sorted(label_mapping.values())

# Print categories
print("Supported Categories:", supported_labels)
                        Copy Code

Getting Started

Inference Example

from transformers import pipeline

# Load classifier
classifier = pipeline("text-classification", model="boltuix/bert-local")

# Predict intent
result = classifier("Where can I see ocean creatures behind glass?")
print(result)  # Output: [{'label': 'aquarium', 'score': 0.999}]
                        Copy Code

Installation

pip install transformers torch pandas scikit-learn tqdm
                        Copy Code

Requires Python 3.8+, ~20 MB for model.

Performance Metrics

Tested on 122 cases with 94.26% accuracy (115/122 correct):

Metric	Value
Accuracy	94.26%
F1 Score (Weighted)	~0.94 (estimated)
Processing Time	<50ms per query

Example results:

Query	Expected Category	Predicted Category	Confidence	Status
How do I catch the early ride to the runway?	✈️ Airport	✈️ Airport	0.997	✅
Are the roller coasters still running today?	🎢 Amusement Park	🎢 Amusement Park	0.997	✅
Where can I see ocean creatures behind glass?	🐠 Aquarium	🐠 Aquarium	1.000	✅

Fine-Tuning Guide

Fine-tune bert-local on custom datasets:

from transformers import AutoTokenizer, AutoModelForSequenceClassification, Trainer, TrainingArguments
from datasets import load_dataset
import torch

# Load dataset
dataset = load_dataset("csv", data_files="custom_intent_dataset.csv")

# Initialize tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("boltuix/bert-local")
model = AutoModelForSequenceClassification.from_pretrained("boltuix/bert-local")

# Tokenize dataset
def tokenize_function(examples):
    return tokenizer(examples["text"], padding="max_length", truncation=True, max_length=128)

tokenized_datasets = dataset.map(tokenize_function, batched=True)

# Training arguments
training_args = TrainingArguments(
    output_dir="./bert_local_finetuned",
    eval_strategy="epoch",
    learning_rate=2e-5,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    num_train_epochs=3,
    weight_decay=0.01,
    report_to="none"
)

# Initialize trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_datasets["train"],
    eval_dataset=tokenized_datasets["validation"]
)

# Train
trainer.train()

# Save model
model.save_pretrained("./bert_local_finetuned")
tokenizer.save_pretrained("./bert_local_finetuned")
                        Copy Code

Other Capabilities

While optimized for intent classification, bert-local’s bert-mini base supports fine-tuning for:

• Question Answering: Answer queries like “What’s the nearest hospital?”
• Sentiment Analysis: Detect positive/negative sentiment.
• Semantic Similarity: Cluster similar intents.
• Named Entity Recognition: Extract entities (e.g., locations).

Dataset Details

• Source: Open-source and synthetic queries (e.g., ChatGPT, Grok).
• Format: CSV with text and label columns.
• Categories: 140 local businesses.
• Size: Model footprint ~20 MB.

Comparison to Other Solutions

Solution	Categories	Accuracy	NLP Strength	Open Source
bert-local	140+	94.26%	Strong	Yes
Google Maps API	~100	~85%	Moderate	No
Yelp API	~80	~80%	Weak	No
OpenStreetMap	Varies	Varies	Weak	Yes

Frequently Asked Questions (FAQ)

License

Apache-2.0 License: Free to use. See LICENSE.

Support & Community

• Hugging Face Model Page
• Transformers Documentation
• Contact: boltuix@gmail.com

Conclusion

bert-local transforms local search with 94.26% accuracy across 140 categories, optimized for edge AI and IoT. From chatbots to travel apps, it’s your solution for intent-driven search in 2025. Explore it on Hugging Face!