Bombek1/bert-base-uncased-litert

bert-base-uncased - LiteRT

This is a LiteRT (formerly TensorFlow Lite) conversion of google-bert/bert-base-uncased for efficient on-device inference.

Model Details

Property	Value
Original Model	google-bert/bert-base-uncased
Format	LiteRT (.tflite)
File Size	414.7 MB
Task	Feature Extraction / Classification Base
Max Sequence Length	128
Output Dimension	768
Pooling Mode	N/A (Full hidden states)

Performance

Benchmarked on AMD CPU (WSL2):

Metric	Value
Inference Latency	83.0 ms
Throughput	12.0/sec
Cosine Similarity vs Original	1.0000 ✅

Quick Start

import numpy as np
from ai_edge_litert.interpreter import Interpreter
from transformers import AutoTokenizer

# Load model and tokenizer
interpreter = Interpreter(model_path="google-bert_bert-base-uncased.tflite")
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-uncased")

def get_hidden_states(text: str) -> np.ndarray:
    """Get hidden states for input text."""
    encoded = tokenizer(
        text,
        padding="max_length",
        max_length=128,
        truncation=True,
        return_tensors="np"
    )

    interpreter.set_tensor(input_details[0]["index"], encoded["input_ids"].astype(np.int64))
    interpreter.set_tensor(input_details[1]["index"], encoded["attention_mask"].astype(np.int64))
    interpreter.invoke()

    return interpreter.get_tensor(output_details[0]["index"])

# Example
hidden = get_hidden_states("Hello, world!")
cls_embedding = hidden[0, 0, :]  # CLS token for classification
print(f"Hidden shape: {hidden.shape}")  # (1, 128, 768)

Files

• google-bert_bert-base-uncased.tflite - The LiteRT model file

Conversion Details

• Conversion Tool: ai-edge-torch
• Conversion Date: 2026-01-12
• Source Framework: PyTorch → LiteRT
• Validation: Cosine similarity 1.0000 vs original

Intended Use

• Mobile Applications: On-device semantic search, RAG systems
• Edge Devices: IoT, embedded systems, Raspberry Pi
• Offline Processing: Privacy-preserving inference
• Low-latency Applications: Real-time processing

Limitations

• Fixed sequence length (128 tokens)
• CPU inference (GPU delegate requires setup)
• Tokenizer loaded separately from original model
• Float32 precision

License

This model inherits the license from the original:

• License: Apache 2.0 (source)

Citation

@article{devlin2018bert,
    title={BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding},
    author={Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina},
    journal={arXiv preprint arXiv:1810.04805},
    year={2018}
}

Acknowledgments

• Original model by google-bert
• Conversion using ai-edge-torch

---

Converted by Bombek1