Upload trained Offline Delivery Assistant SLM
Browse files- README.md +52 -0
- config.json +16 -0
- pytorch_model.bin +3 -0
README.md
ADDED
|
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- en
|
| 4 |
+
license: mit
|
| 5 |
+
tags:
|
| 6 |
+
- small-language-model
|
| 7 |
+
- slm
|
| 8 |
+
- pytorch
|
| 9 |
+
- edge-ai
|
| 10 |
+
- logistics
|
| 11 |
+
- customer-support
|
| 12 |
+
- delivery-assistant
|
| 13 |
+
pipeline_tag: text-generation
|
| 14 |
+
widget:
|
| 15 |
+
- text: "<|user|> Customer is asking to change delivery address after order pickup. <|assistant|>"
|
| 16 |
+
---
|
| 17 |
+
|
| 18 |
+
# Delivery-Rider-Assistant-SLM (30.0M Parameters)
|
| 19 |
+
|
| 20 |
+
## Overview
|
| 21 |
+
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).
|
| 22 |
+
|
| 23 |
+
## Intended Use Case
|
| 24 |
+
- Offline edge execution in areas with poor connectivity.
|
| 25 |
+
- Quick responses for address changes, payment status checks, COD handling, and order issues.
|
| 26 |
+
|
| 27 |
+
## Model Details & Architecture
|
| 28 |
+
- Architecture: Decoder-only causal transformer (nanoGPT-style, flash attention)
|
| 29 |
+
- Parameters: ~30.0 million
|
| 30 |
+
- Layers: 6
|
| 31 |
+
- Attention heads: 8
|
| 32 |
+
- Embedding dimension: 384
|
| 33 |
+
- Context length: 256 tokens
|
| 34 |
+
- Tokenizer: BPE (GPT-2 encoding via tiktoken)
|
| 35 |
+
|
| 36 |
+
## Training Setup
|
| 37 |
+
- Hardware: Single NVIDIA Tesla T4 GPU (Google Colab Free Tier)
|
| 38 |
+
- Training time: ~23 minutes
|
| 39 |
+
- Precision: Mixed precision (float16 AMP)
|
| 40 |
+
- Optimizer: AdamW (lr=6e-4, weight_decay=0.1, cosine schedule with warmup)
|
| 41 |
+
- Best validation loss: 0.7974
|
| 42 |
+
- Data: Bitext customer-support instruction dataset
|
| 43 |
+
|
| 44 |
+
## How to Load in PyTorch
|
| 45 |
+
```python
|
| 46 |
+
import torch
|
| 47 |
+
from huggingface_hub import hf_hub_download
|
| 48 |
+
|
| 49 |
+
weights_path = hf_hub_download(repo_id="learner-subodh/Delivery-Rider-Assistant-SLM", filename="pytorch_model.bin")
|
| 50 |
+
# Re-create the GPT class and GPTConfig from the training notebook, then:
|
| 51 |
+
# model.load_state_dict(torch.load(weights_path, map_location="cpu"))
|
| 52 |
+
```
|
config.json
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"GPT"
|
| 4 |
+
],
|
| 5 |
+
"model_type": "nanogpt-custom",
|
| 6 |
+
"vocab_size": 50257,
|
| 7 |
+
"block_size": 256,
|
| 8 |
+
"n_embd": 384,
|
| 9 |
+
"n_layer": 6,
|
| 10 |
+
"n_head": 8,
|
| 11 |
+
"dropout": 0.1,
|
| 12 |
+
"bias": true,
|
| 13 |
+
"tokenizer": "tiktoken gpt2",
|
| 14 |
+
"task": "causal-lm",
|
| 15 |
+
"domain": "Logistics & Offline Delivery Partner Assistance"
|
| 16 |
+
}
|
pytorch_model.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9192739d562b053780906c8d33727a9e3dd321f1335fd833bbab861c8d188984
|
| 3 |
+
size 120207595
|