README.md

Base Small Language Model (SLM)

🚀 CPU-First Base Language Model

This is the base model before fine-tuning - a blazing-fast, CPU-optimized Small Language Model foundation:

⚡ Performance Highlights

• 164 tokens/sec on CPU (fast base performance)
• 45.2MB model size (base model)
• 3.7M parameters (tiny but powerful)
• General language understanding (pre-fine-tuning)

🎯 Training Speed

• 28 minutes for base training (4 epochs)
• Fast convergence with efficient architecture
• Ready for fine-tuning on any domain

🔧 Technical Specs

• Architecture: Transformer-lite with RMSNorm, SwiGLU, Rotary embeddings
• Optimization: CPU-first with memory mapping and efficient batching
• Framework: PyTorch (CPU optimized)
• Training: Trained on conversational data

📱 Deployment Ready

• CPU optimized: No GPU required
• Fast startup: Instant model loading
• Low memory: Efficient memory usage
• Fine-tuning ready: Perfect base for domain adaptation

Usage

Load and Use Base Model

import torch
import sys
sys.path.append('src')
from model import create_model_from_config
from tokenizer import BPETokenizer

# Load model
checkpoint = torch.load("checkpoints/model_latest.pt", map_location='cpu')
config = checkpoint['config']
model = create_model_from_config(config)
model.load_state_dict(checkpoint['model_state_dict'])

# Load tokenizer
tokenizer = BPETokenizer()
tokenizer.load("data/tokenizer.json")

# Generate
prompt = "Hello, how are you?"
input_ids = tokenizer.encode(prompt, add_special_tokens=True)
input_ids = torch.tensor([input_ids], dtype=torch.long)

model.eval()
with torch.no_grad():
    for _ in range(20):
        logits = model(input_ids)[0, -1, :]
        next_token = torch.argmax(logits, dim=-1).unsqueeze(0)
        input_ids = torch.cat([input_ids, next_token.unsqueeze(0)], dim=1)

response = tokenizer.decode(input_ids[0].tolist(), skip_special_tokens=True)
print(response)

Fine-tune on Your Data

# Use this base model for fine-tuning
python finetune_qa.py --base_model checkpoints/model_latest.pt --conversations your_data.json

Model Details

• Base Model: Trained on conversational data
• Architecture: Transformer-lite with modern optimizations
• Size: 45.2MB (base model)
• License: MIT

Performance

Metric	Value
Speed	164 tokens/sec
Size	45.2MB
Parameters	3.7M
Training Time	28 minutes

This base model provides an excellent foundation for fine-tuning on specific domains or tasks.