Frequently Asked Questions

Three things:

• Data — at least one event table with entity IDs, timestamps, and ≥ 100k interactions/month across ≥ 10k entities. See Requirements.
• Infrastructure — a Docker-capable environment with GPU meeting the hardware requirements.
• A data scientist — someone who can configure data sources, define prediction targets, and operate the training process.

At minimum a data scientist. Depending on your organization you may also need:

• DevOps / Infra — to deploy the container and manage GPU resources
• Data Engineer — to grant read-only warehouse access and configure data sources
• Analyst / Product Owner — to define business objectives and evaluate results

If infrastructure and data access are in place, BaseModel can be operational within a few hours. Most teams have a foundation model training on day one. The first scenario model (e.g. churn prediction) is typically ready within a week.

No. BaseModel requires timestamped event streams linked to identifiable entities. Static-only data is better served by traditional models like GBDTs.

Any recurring, timestamped interaction tied to an entity ID: transactions, page views, calls, service activations, support tickets, campaign responses, etc. See Data Connect Sources.

Depends on interaction frequency:

• High-frequency (banking, telco, FMCG) — 3+ months
• Low-frequency (fashion, insurance, automotive) — 1+ year

See Requirements for full volume guidelines.

Yes:

• Temporally aggregated features (e.g. weekly spend) → treat as timestamped events
• Static features (e.g. demographics, product metadata) → include as entity or item attribute tables

BaseModel does not impute missing values. Columns with missing data exceeding a defined threshold are excluded during training. Check the foundation model's fit stage logs for warnings.

All timestamps are converted to UTC. If a timezone is specified, BaseModel translates accordingly. If no timezone is provided, timestamps are interpreted as UTC.

Data splits are enforced chronologically by design. Training and test sets must be separated in time. If this rule is violated, BaseModel raises an error. See Foundation Model — Basic Configuration for split setup.

Run verify_target() on your target function before training to catch runtime errors, data type mismatches, or excessive None returns. See Function Validation.

Training can run for a fixed number of epochs or use configurable early stopping. Track metrics via MLflow or another logger to monitor convergence. See Training Controls.

BaseModel learns from naturally imbalanced data effectively through its representation learning. When needed, you can apply sampling or loss weighting and compare results using validation metrics.

In priority order:

1. Optimize data — use interpretation to identify high-impact columns, remove duplicates, experiment with shorter time windows
2. Parallelise — scale across GPUs, tune number of workers
3. Enable caching — reuse preprocessed batches between runs
4. Reduce model complexity — lower graph quality or network size (may reduce accuracy)

See Training Scaling.

Yes — the target function defines a prediction interval relative to the prediction date. For example, predict churn between 30 and 90 days out. The prediction date itself must not be set beyond your last available data.

No. All computation happens inside your infrastructure. BaseModel reads directly from your data warehouse or local files.

BaseModel connects to all major data platforms (Snowflake, Databricks, BigQuery, ClickHouse, Hive, Azure Synapse) and integrates with MLflow for experiment tracking. See Monitoring with MLflow.

Feature embeddings, model checkpoints, logs, configuration files, and prediction outputs (TSV format). See Inference for output details.

It runs efficiently on a single GPU but scales linearly with multi-GPU and multi-node setups. Preprocessing can run on CPU-only environments; training and inference require GPUs.

Typically every 4 weeks to account for new entities, seasonal patterns, and behavioral shifts. The optimal frequency depends on your data refresh rate and domain dynamics.

BaseModel provides event-level attribution: which individual interactions contributed most to a prediction. This goes beyond aggregate feature importance and maps directly to behavioral drivers. See Interpretation.

Yes — as long as the model checkpoint and corresponding data are still available. Preserving checkpoints for audit purposes is the client's responsibility.

All processing happens within your infrastructure. BaseModel does not transfer data externally, does not perform federated learning, and does not train across clients. Data governance and compliance are the client's responsibility.