Prepare features, work with numerical arrays and analyze model outputs in q/k. l brings filtering, statistics and matrix operations into a runtime that considers the physical layout of the data it computes on.
01 / The workload
Filtering observations, constructing features, scoring numerical arrays and inspecting outputs all require repeated passes over data. Intermediate allocations and expanded columns can dominate that work. A programmable array layer makes the transformations visible and gives you a place to measure their cost.
Select observations, transform numerical columns and calculate the statistics used by your feature pipeline. Preserve identifiers so derived values can be traced back to their source.
Express dot products, matrix-vector calculations and elementwise transformations using the documented numerical primitives. Build and inspect the calculations behind a small scoring function.
Analyze scores, labels and evaluation results supplied by your model pipeline. Group observations to see where score distributions and measured outcomes change.
Work with ordered numerical observations using differences, running calculations, correlations and grouped summaries. Prepare the features and comparisons required by your own forecasting workflow.
02 / From data to analysis
Keep the source identifiers, feature expressions and numerical outputs connected through the preparation and analysis stages.
Start with flat observations and the identifiers, cohort fields and supplied labels needed for the task.
Filter observations, derive feature columns and apply the statistics or matrix operations needed by your calculation.
Exchange a feature table with your model tools or inspect a supplied set of scores, retaining the rows behind each comparison.
/ Compose numerical features
l>x:1 2 3 4.0
l>(x-avg x)%dev x
-1.341641 -0.4472136 0.4472136 1.341641
l>w:(0.1 0.2 0.3;-0.4 0.1 -0.2)
l>w mmu 1 2 3.0
1.4 -0.8
l>{0.0|x} w mmu 1 2 3.0
1.4 0
l>{e%sum e:exp x-max x} x
0.0320586 0.08714432 0.2368828 0.6439143
l>03 / Why l
You write q/k. The runtime considers the physical representation of the data, choosing encoding-aware operations where supported and materializing values when needed.
04 / Your first evaluation
Choose a feature transformation or small numerical scoring calculation with reference inputs, outputs and clearly defined numeric types.
Available as a research preview for local evaluation.
The data layout and execution discipline behind wide kernels.
Read the paper Integration guideCall l from Python, pass typed values and bring results back to your application.
Read the guide Integration guideRead, write and stream Arrow IPC files with the native interop library.
Read the guideYour data. Your next question.