yug / README.md

docs(README) : add GIFT-Eval results (CRPS, MASE) to model-index metadata

ec53758 verified 17 days ago

5.64 kB

---
license: cc-by-nc-4.0
tags:
  - time series
  - forecasting
  - foundation models
  - pretrained models
  - safetensors
datasets:
  - autogluon/chronos_datasets
  - Salesforce/GiftEvalPretrain
pipeline_tag: time-series-forecasting
library_name: yug
model-index:
  - name: Yug
    results:
      - task:
          type: time-series-forecasting
        dataset:
          name: GIFT-Eval
          type: GIFT-Eval
        metrics:
          - type: CRPS
            value: 0.529
            name: CRPS
          - type: MASE
            value: 0.744
            name: MASE
        source:
          url: https://huggingface.co/spaces/Salesforce/GIFT-Eval
          name: GIFT-Eval Time Series Forecasting Leaderboard
---
# Yug

**Yug** is a 271.8M-parameter time series foundation model developed by Birla AI Labs for **zero-shot probabilistic forecasting**. A single checkpoint forecasts a previously unseen series without task-specific training or per-series tuning: given a historical context and a forecast horizon, it returns a predictive distribution summarised by quantile levels. It supports **univariate** and **covariate-informed** forecasting within a single architecture.


## Links

- 💻 [GitHub](https://github.com/birla-ai-labs/yug)
- 📓 [Quickstart notebook](https://github.com/birla-ai-labs/yug/blob/main/notebooks/quickstart.ipynb)
- 🐛 [Issues and questions](https://github.com/birla-ai-labs/yug/issues)
<!-- TODO(release): add a technical report / paper link here once public. -->

## Overview

| Capability | Yug |
|---|---|
| Univariate forecasting | ✅ |
| Covariates | ✅ |
| Probabilistic (quantile) output | ✅ |
| Max context length | 2048 |
| Quantile levels | 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9 |


## Model architecture

| | |
|---|---|
| Parameters | 271,845,260 |
| Model dimension | 960 |
| Decoder layers | 12 |
| Attention heads | 8 |
| MLP expansion | 1.25 |
| Patch length | 16 |
| Block size (points per forward pass) | 64 |
| Precision | float32 |
| Format | `safetensors` |

## Running the model locally

Install the package:

```bash
pip install yug
```

Optional extras:

```bash
pip install 'yug[pandas]'      # needed for forecast.to_dataframe()
pip install 'yug[gluonts]'     # GluonTS predictor for benchmark harnesses
```

Check the machine can run it before downloading weights:

```bash
yug-check
```

Make zero-shot predictions:

```python
import numpy as np
from yug import YugPipeline
pipeline = YugPipeline.from_pretrained(
    "birlaailabs/yug", device_map="cuda"
)
# Synthetic signal: trend + 64-step season + 12-step season + noise
CONTEXT, HORIZON, PERIOD = 2000, 256, 64
rng = np.random.default_rng([20240902, 315595964, 0])
t = np.arange(CONTEXT + HORIZON)
trend = np.linspace(0.0, rng.uniform(50, 150), len(t))
phase1, amp1 = rng.uniform(0, 6), rng.uniform(10, 30)
phase2, amp2 = rng.uniform(0, 6), rng.uniform(2, 8)
signal = (
    trend                                                 # trend
    + amp1 * np.sin(2 * np.pi * t / PERIOD + phase1)      # 64-step season
    + amp2 * np.sin(2 * np.pi * t / 12 + phase2)          # 12-step season
    + rng.normal(0, 1.0, len(t))                          # noise
).astype(np.float32)
context, truth = signal[:CONTEXT], signal[CONTEXT:]
forecast = pipeline.predict(
    context,
    prediction_length=HORIZON,
    freq="D",            # a model input, not metadata: pandas offset alias
    seed=0,              # makes the call reproducible
)
print(forecast.median)                # (256,)     point forecast
print(forecast.quantile(0.9))         # (256,)
print(forecast.interval())            # {"lower": (256,), "upper": (256,)}
print(forecast.to_dataframe().head())
```

Several series at once (lengths and frequencies may differ):

```python
forecast = pipeline.predict(
    [series_a, series_b, series_c],
    prediction_length=48,
    freq=["D", "D", "H"],
    item_ids=["a", "b", "c"],
)
forecast.values          # (3, n_quantiles, 48)
```

With covariates — stack channels into a 2-D array, target first:

```python
multivariate = np.stack([target, covariate_1, covariate_2])
forecast = pipeline.predict([multivariate], prediction_length=48)
```

Covariate forecasts degrade faster with horizon than univariate ones. For
long horizons the univariate path is often the better answer — measure both.

In a GluonTS benchmark:

```python
from yug.gluonts_adapter import YugPredictor
predictor = YugPredictor.from_pretrained(
    "birlaailabs/yug", freq="H", prediction_length=48
)
forecasts = list(predictor.predict(dataset))
```

## Training data

- Subset of [Chronos Datasets](https://huggingface.co/datasets/autogluon/chronos_datasets)
- Subset of [GIFT-Eval Pretrain](https://huggingface.co/datasets/Salesforce/GiftEvalPretrain)
- Proprietary synthetic data

## Citation

If you find Yug useful for your research, please consider citing it:

```bibtex
@software{yug_2026,
  title  = {Yug: a foundation model for zero-shot probabilistic time-series forecasting},
  author = {Aaditya Jain* and Debdeep Sanyal* and Aaryan Nagpal and Dhruv Kumar and Murari Mandal and Saurabh Deshpande},
  year   = {2026},
  organization = {Birla AI Labs},
  url    = {https://github.com/birla-ai-labs/yug}
}
```
*Equal contribution.
## Acknowledgements
We thank Karthik Sridhar, Deepak Lenka, Akash Kokare, Manya Pandey, and Varya Srivastava, as well as the extended Birla AI Labs team for their support.
## License
Model weights : [CC-BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)
Code : licensed separately — see [GitHub repository](https://github.com/birla-ai-labs/yug) for its license.

	---
	license: cc-by-nc-4.0
	tags:
	- time series
	- forecasting
	- foundation models
	- pretrained models
	- safetensors
	datasets:
	- autogluon/chronos_datasets
	- Salesforce/GiftEvalPretrain
	pipeline_tag: time-series-forecasting
	library_name: yug
	model-index:
	- name: Yug
	results:
	- task:
	type: time-series-forecasting
	dataset:
	name: GIFT-Eval
	type: GIFT-Eval
	metrics:
	- type: CRPS
	value: 0.529
	name: CRPS
	- type: MASE
	value: 0.744
	name: MASE
	source:
	url: https://huggingface.co/spaces/Salesforce/GIFT-Eval
	name: GIFT-Eval Time Series Forecasting Leaderboard
	---

	# Yug

	Yug is a 271.8M-parameter time series foundation model developed by Birla AI Labs for zero-shot probabilistic forecasting. A single checkpoint forecasts a previously unseen series without task-specific training or per-series tuning: given a historical context and a forecast horizon, it returns a predictive distribution summarised by quantile levels. It supports univariate and covariate-informed forecasting within a single architecture.


	## Links

	- 💻 [GitHub](https://github.com/birla-ai-labs/yug)
	- 📓 [Quickstart notebook](https://github.com/birla-ai-labs/yug/blob/main/notebooks/quickstart.ipynb)
	- 🐛 [Issues and questions](https://github.com/birla-ai-labs/yug/issues)
	<!-- TODO(release): add a technical report / paper link here once public. -->

	## Overview

	\| Capability \| Yug \|
	\|---\|---\|
	\| Univariate forecasting \| ✅ \|
	\| Covariates \| ✅ \|
	\| Probabilistic (quantile) output \| ✅ \|
	\| Max context length \| 2048 \|
	\| Quantile levels \| 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9 \|


	## Model architecture

	\| \| \|
	\|---\|---\|
	\| Parameters \| 271,845,260 \|
	\| Model dimension \| 960 \|
	\| Decoder layers \| 12 \|
	\| Attention heads \| 8 \|
	\| MLP expansion \| 1.25 \|
	\| Patch length \| 16 \|
	\| Block size (points per forward pass) \| 64 \|
	\| Precision \| float32 \|
	\| Format \| `safetensors` \|

	## Running the model locally

	Install the package:

	```bash
	pip install yug
	```

	Optional extras:

	```bash
	pip install 'yug[pandas]' # needed for forecast.to_dataframe()
	pip install 'yug[gluonts]' # GluonTS predictor for benchmark harnesses
	```

	Check the machine can run it before downloading weights:

	```bash
	yug-check
	```

	Make zero-shot predictions:

	```python
	import numpy as np
	from yug import YugPipeline

	pipeline = YugPipeline.from_pretrained(
	"birlaailabs/yug", device_map="cuda"
	)

	# Synthetic signal: trend + 64-step season + 12-step season + noise
	CONTEXT, HORIZON, PERIOD = 2000, 256, 64
	rng = np.random.default_rng([20240902, 315595964, 0])
	t = np.arange(CONTEXT + HORIZON)
	trend = np.linspace(0.0, rng.uniform(50, 150), len(t))
	phase1, amp1 = rng.uniform(0, 6), rng.uniform(10, 30)
	phase2, amp2 = rng.uniform(0, 6), rng.uniform(2, 8)
	signal = (
	trend # trend
	+ amp1 * np.sin(2 * np.pi * t / PERIOD + phase1) # 64-step season
	+ amp2 * np.sin(2 * np.pi * t / 12 + phase2) # 12-step season
	+ rng.normal(0, 1.0, len(t)) # noise
	).astype(np.float32)

	context, truth = signal[:CONTEXT], signal[CONTEXT:]

	forecast = pipeline.predict(
	context,
	prediction_length=HORIZON,
	freq="D", # a model input, not metadata: pandas offset alias
	seed=0, # makes the call reproducible
	)

	print(forecast.median) # (256,) point forecast
	print(forecast.quantile(0.9)) # (256,)
	print(forecast.interval()) # {"lower": (256,), "upper": (256,)}
	print(forecast.to_dataframe().head())
	```

	Several series at once (lengths and frequencies may differ):

	```python
	forecast = pipeline.predict(
	[series_a, series_b, series_c],
	prediction_length=48,
	freq=["D", "D", "H"],
	item_ids=["a", "b", "c"],
	)
	forecast.values # (3, n_quantiles, 48)
	```

	With covariates — stack channels into a 2-D array, target first:

	```python
	multivariate = np.stack([target, covariate_1, covariate_2])
	forecast = pipeline.predict([multivariate], prediction_length=48)
	```

	Covariate forecasts degrade faster with horizon than univariate ones. For
	long horizons the univariate path is often the better answer — measure both.

	In a GluonTS benchmark:

	```python
	from yug.gluonts_adapter import YugPredictor

	predictor = YugPredictor.from_pretrained(
	"birlaailabs/yug", freq="H", prediction_length=48
	)
	forecasts = list(predictor.predict(dataset))
	```

	## Training data

	- Subset of [Chronos Datasets](https://huggingface.co/datasets/autogluon/chronos_datasets)
	- Subset of [GIFT-Eval Pretrain](https://huggingface.co/datasets/Salesforce/GiftEvalPretrain)
	- Proprietary synthetic data

	## Citation

	If you find Yug useful for your research, please consider citing it:

	```bibtex
	@software{yug_2026,
	title = {Yug: a foundation model for zero-shot probabilistic time-series forecasting},
	author = {Aaditya Jain* and Debdeep Sanyal* and Aaryan Nagpal and Dhruv Kumar and Murari Mandal and Saurabh Deshpande},
	year = {2026},
	organization = {Birla AI Labs},
	url = {https://github.com/birla-ai-labs/yug}
	}
	```
	*Equal contribution.

	## Acknowledgements

	We thank Karthik Sridhar, Deepak Lenka, Akash Kokare, Manya Pandey, and Varya Srivastava, as well as the extended Birla AI Labs team for their support.

	## License

	Model weights : [CC-BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)

	Code : licensed separately — see [GitHub repository](https://github.com/birla-ai-labs/yug) for its license.