Pharia

No specific accounting or listing available. Training data described in broad terms as 'web-crawled data and structured datasets with a total size of 7.7T, with a cutoff date 04/2023' alongside 'some additional web scraping'.

No data provided except a very generic reference to 'source-available, commercially usable datasets, as well as self-created and procured proprietary datasets'.

Pharia 1 base model made available through HuggingFace

Aligned model made available through HuggingFace

Aleph Alpha claims Pharia has been trained using the Scaling code base, which it made available as a repository mirrored from an unknown source; no specific repository found documenting the training or instruction-tuning of Pharia.

No specific code for training and tuning found, so no documentation of code found. Scaling code base is reasonably documented.

Training process, architecture and resource usage are documented at above-average levels of detail.

No preprint or in-depth scientific documentation found outside model card.

No peer-reviewed paper found

Modelcard on HF provides a fair bit of documentation of training data, architecture, evaluation, and risks and limitations.

No datasheet or other detailed accounting of pretraining, finetuning or instruction tuning data found. Scant details given in model card.

Package available through a bespoke Intelligence Layer SDK

Paid API access available through Intelligence Layer SDK.

Licensed under Open Aleph License 1.0, which limits use to personal, non-profit, research and educational uses.