Hardware and Data Requirements

In order to effectively and efficiently utilize BaseModel you need sufficient amount of behavioral data and adequately powerful infrastructure.

Data

• Minimum requirements:

  • At least 10k individual behavioral profles (eg. customers, subscribers, members etc.)
    They will act as the main entity and level of the model's predictions,
  • One or more event data sources containing behavioral data.
    These are interactions between the main entity and other entities (eg. products, services, stores etc.)
  • All records in event data sources must have entity ID and timestamp of the event.
  • There should be at least 100k interactions observed per month, tied to behavioral profiles.

• Recommended information and data volume:

  • At least 3 months' worth of event data for industries where interactions are daily, such as banking, telco, FMCG retail.
  • At least 1 year's worth of data for industries where the interactions are less frequent, eg. fashion, insurance, automotive.
  • Attributes related to behavioral profile and other entities (see below).
  
  

Hardware

• An environment capable of running Docker containers
• A100 GPU or better with CUDA 11.7 or higher
• Minimum of 240 GB RAM
• At least 32 CPU cores
• 1 TB hard disk space
  
  

Example infrastructure and illustrative performance

Our proprietary algorithms - Cleora and emde - enable a notably efficient and fast training process, allowing to process vast amount of behavioral data. Below is a example of an end-to-end cycle of foundation model and downstream model training based on a real-world retail customer which should help to assess resource requirements and expected costs in your case:

Data Size

• Ca. 8 Bn events
• Ca. 18 Mn unique Client IDs
• Ca. 1 Mn unique Product IDs

Performance

• Self-supervised foundation model training: 12h @ 1x Nvidia A100
• Full (non-LoRA) fine-tuning for a supervised task (brand propensity for 16,000 brands): 10h @ 1x Nvidia A100
• Full model (non-quantized) inference throughput: 2718 clients / sec per 1 GPU
• Training and inference scale linearly with the number of GPUs.
  
  

If necessary, additional optimizations may be applied depending on customer needs, at the cost of slightly impaired model quality:

• Low rank fine-tuning,
• Model quantization for inference.