---
language: en
tags:
  - gpt
  - weather
  - character-level
  - pytorch
  - nano
license: mit
---

# nanoWeatherGPT

A character-level GPT (~813K parameters) trained from scratch on synthetic weather forecast text.

## Download & Use

```python
from huggingface_hub import snapshot_download
import torch, pickle, sys

# Download all files
path = snapshot_download("agoel7029/nano-weather-gpt-model")
sys.path.insert(0, path)

from model import GPT

ckpt = torch.load(f"{path}/model.pt", map_location="cpu", weights_only=False)
meta = pickle.load(open(f"{path}/meta.pkl", "rb"))
stoi, itos = meta["stoi"], meta["itos"]

model = GPT(ckpt["config"])
model.load_state_dict(ckpt["model_state_dict"])
model.eval()

prompt = "Today the weather is"
ctx = torch.tensor([[stoi[c] for c in prompt]], dtype=torch.long)
with torch.no_grad():
    out = model.generate(ctx, max_new_tokens=200)
print("".join(itos[i] for i in out[0].tolist()))
```

## Fine-tune on your own data

Edit `train.py` — replace `generate_weather_corpus()` with your own text, then run:

```bash
python train.py
```

## Model details

| Property | Value |
|---|---|
| Architecture | Character-level GPT (Transformer) |
| Parameters | ~813K |
| Layers | 4 |
| Attention heads | 4 |
| Embedding dim | 128 |
| Context length | 128 characters |
| Trained on | Synthetic Swiss city weather reports |