Delivery-Rider-Assistant-SLM (30.0M Parameters)

Overview

Delivery-Rider-Assistant-SLM is a lightweight Small Language Model (SLM) trained from scratch, intended for offline use on low-power mobile devices by delivery partners and drivers (e.g., Swiggy, Zomato, Porter).

Intended Use Case

• Offline edge execution in areas with poor connectivity.
• Quick responses for address changes, payment status checks, COD handling, and order issues.

Model Details & Architecture

• Architecture: Decoder-only causal transformer (nanoGPT-style, flash attention)
• Parameters: ~30.0 million
• Layers: 6
• Attention heads: 8
• Embedding dimension: 384
• Context length: 256 tokens
• Tokenizer: BPE (GPT-2 encoding via tiktoken)

Training Setup

• Hardware: Single NVIDIA Tesla T4 GPU (Google Colab Free Tier)
• Training time: ~23 minutes
• Precision: Mixed precision (float16 AMP)
• Optimizer: AdamW (lr=6e-4, weight_decay=0.1, cosine schedule with warmup)
• Best validation loss: 0.7974
• Data: Bitext customer-support instruction dataset

How to Load in PyTorch

import torch
from huggingface_hub import hf_hub_download

weights_path = hf_hub_download(repo_id="learner-subodh/Delivery-Rider-Assistant-SLM", filename="pytorch_model.bin")
# Re-create the GPT class and GPTConfig from the training notebook, then:
# model.load_state_dict(torch.load(weights_path, map_location="cpu"))