Sequence-to-Omics Run Explorer

Compare the recorded experiments. Inspect the curves behind the results.

20 runs · offline · no account
Loading embedded run data…

Drag across a plot to zoom all charts. Scroll to zoom around the pointer. Double-click to reset. Hover for exact raw and smoothed values. Click a run name to show only that run.

Metric summary

Raw logged values, independent of smoothing and zoom. Click a column heading to sort.

How to read this viewer

All data is embedded in this file. The viewer uses only the 20 allowlisted runs exported with W&B scan_history; no metric rows are sampled for export. Curves preserve each metric's actual logging frequency and are not interpolated. A metric may be absent from a run.

EMA uses s₀ = y₀ and sₜ = αsₜ₋₁ + (1−α)yₜ. Moving average uses the most recent N logged values, including the current point; its window counts metric observations, not optimizer steps. This is an explicit smoothing definition, not a promise of W&B's exact internal smoothing behavior. Lines are visually thinned for large curves by retaining bucket minima and maxima; hover and CSV export use every recorded point.

Missing optimizer steps are excluded when that axis is selected. Duplicate X values use the last logged observation for that metric. Elapsed time is shown in hours. Wall-clock range inputs use Unix seconds. Log axes omit nonpositive values. Summary statistics use the original metric observations, including repeated X values.

Dataset and parameter scores are test summaries. Depth curves use validation metrics. The native-test paired track benchmark and official-model cosine analysis have separate source files under results/ and analysis/raw/transformer_usage/; this viewer displays the training-run scalar exports. Settings and hidden runs are saved in your browser when local storage is available. Nothing is transmitted.

Analysis methods and source data