v0.3 — Colab T4 training, 21% MASE improvement, 6.5M params

v0.3 Release

Highlights

• Larger model: d_model=96, 8 layers, 6.5M params (up from 1.6M)
• Longer context: 512 timesteps (up from 256)
• More data: 6 real datasets + 10K synthetic records, 200 epochs
• 21% MASE improvement: Overall MASE 2.73 (was 3.45 in v0.2)
• 5/6 datasets improved: ETTh1 (-42%), ETTh2 (-26%), ETTm1 (-39%), electricity (-16%), traffic (-35%)
• Trained on Colab T4: 11.7h wall time, best epoch 147, val_loss 0.2230
• Best checkpoint pushed: https://huggingface.co/eulogik/nanoforecast-v03

Full Benchmarks

Dataset	MASE	sMAPE
ETTh1	1.95	12.06%
ETTh2	2.74	10.47%
ETTm1	2.17	10.70%
exchange_rate	7.44	1.72%
electricity	1.29	4.76%
traffic	0.81	24.00%

Links

• Model Hub: https://huggingface.co/eulogik/nanoforecast-v03
• Live Demo: https://huggingface.co/spaces/eulogik/nanoforecast
• Paper: deploy/paper.tex
• PyPI: https://pypi.org/project/nanoforecast/

What's Changed

• Colab training notebook: deploy/colab_training_v03.ipynb
• Dataset caching: retry on corrupted downloads
• Gradio 6.x placeholder fix