| license: apache-2.0 | |
| base_model: theforecastingcompany/t0-alpha | |
| tags: | |
| - time-series | |
| - forecasting | |
| - onnx | |
| - onnxruntime | |
| - quantized | |
| - zero-shot | |
| # t0-alpha (int8 ONNX) | |
| int8 ONNX export of [theforecastingcompany/t0-alpha](https://huggingface.co/theforecastingcompany/t0-alpha) (Apache-2.0), quantized as: dynamic per-channel QInt8. | |
| int8 output verified within 15.2% of forecast spread (mean 3%) of the official `T0Forecaster.predict()`. Unofficial export, not | |
| affiliated with or endorsed by the model authors. | |
| ## Contract | |
| | input | dtype | shape | notes | | |
| |---|---|---|---| | |
| | `context` | float32 | `['batch', 512]` | NaN = missing; left-pad short series with NaN | | |
| Output `quantiles`: float32 `['batch', 64, 5]`. Quantile levels: `[0.1, 0.25, 0.5, 0.75, 0.9]`. | |
| One forward pass. Left-pad short series with NaN (the model is missing-aware; padding is exactly equivalent to the natural-length call). Median is level index 2. Univariate graph; for the multivariate `group_ids` variant run `scripts/export_t0_onnx.py --grouped`. | |
| See `manifest.json` for the machine-readable contract and the artifact's | |
| sha256. `LICENSE` is Apache-2.0; when the upstream repo ships a `NOTICE` | |
| file it is republished here unchanged, as the license requires. | |