euhidaman/bitmar-attention-multimodal

BitMar 100M Token Model

This model was trained on exactly 100 million tokens for the BabyLM Challenge.

Key Features

BitNet Quantization: 1.58-bit quantized text encoder/decoder for efficient inference.
Vision Processing: Quantized Vision Transformer using pre-computed DiNOv2 features.
Episodic Memory: Human Brain inspired memory mechanism for learning, without fine-tuning.

Training Details

• Total tokens: 100,000,000
• Epochs completed: 10
• Tokens processed: 996,862,184
• Cross-modal similarity: 0.4552

Model Architecture

• Text encoder: 4 layers, 128 hidden size
• Vision encoder: DiNOv2 features compressed to 128
• Episodic memory: 32 slots

Usage

from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("euhidaman/bitmar-attention-multimodal")
tokenizer = AutoTokenizer.from_pretrained("euhidaman/bitmar-attention-multimodal")

Training Status

• Status: Completed
• Tokens Processed: 996,862,184
• Best Cross-modal Similarity: 0.4552