How a DQU is calculated
A Data-Qualifying Unit is round(max(estimated logical GB, 0.01) × the sum of the task multipliers, to two decimal places). Logical size comes from schema preflight. Axolotl does not bill from compressed file size alone.
Tier allowances, prices, and overage rates are in the Pricing section at https://www.adequata.ai/. This page only explains the measurement behind the estimate you confirm before a run.
Logical GB for CSV and Parquet
Billing uses logical size. For CSV that is the file size. For Parquet it is the uncompressed size of the row groups, not the compressed file.
| Format | Logical size |
|---|---|
| CSV | File byte size on storage (logical ≈ physical) |
| Parquet | Sum of row-group uncompressed sizes from metadata |
| Parquet (metadata unreadable) | Fallback: 4 × file size so usage is not under-reported |
Task multiplier (by stage range and options)
Each kind of work in the stage range you selected adds its own multiplier. The total is the sum of those multipliers, not the largest one.
- For example, 2.00 logical GB with basic checks and fuzzy deduplication only uses multipliers of 0.5, 0.5, and 10. The sum is 11.0, so the estimate is 22.00 DQU.
- If you select the full pipeline, fuzzy deduplication, every anomaly method, and hyperparameter tuning, the multipliers add up to 34 times the logical size.
- A run that is blocked because the dataset is too large does not consume DQUs. You are only charged for work Axolotl is allowed to start.
| Work item | Multiplier |
|---|---|
| Basic: invalid values | 0.5 |
| Basic: null / sparse rows | 0.5 |
| Dedup skipped | 0 |
| Dedup exact only | 1.0 |
| Dedup fuzzy (incl. fuzzy + exact prefilter) | 10.0 |
| Anomaly stage, no method selected | 0.25 |
| Each column method (IQR, Z-score) | 0.25 |
| Dataset methods block (if any dataset method selected) | 0.5 + 2.0 per method |
| Synthetic imputation | 7.0 |
| Synthetic + hyperparameter tuning | 14.0 |