SELECTED RUNS
—Compared side by sidegpt-lab
Training overview.
What changed, what was measured, and the evidence behind it.
Your experiments, in one place.
Import a results snapshot or export your existing W&B runs.
Learning curves, evaluations, and configurations will appear here.
Imported files stay in this browser session. Nothing is uploaded.
0 selected
No runs match this search.
LOWEST RECORDED BPB
—Validation · lower is betterHIGHEST RECORDED CORE
—Centered score · higher is betterRECORDED EVALUATION
—Sum of logged BPB + CORE durationsTraining loss
EMA · 0.9 · NATSValidation BPB
LOWER IS BETTERTraining throughput
TOKENS / SECONDModel FLOPs utilization (MFU)
ESTIMATED · %CORE over training
CENTERED SCOREEvaluation duration
SECONDS PER CORE PASSCORE by task
LATEST RECORDED ACCURACY PER TASKCheck example counts and few-shot settings before comparing scores. Missing results stay blank.
Configuration comparison
Nested settings stay together. Expand JSON to inspect, or download a run’s complete configuration.
Run recap
LAST RECORDED VALUESTraining time excludes evaluation in gpt-lab. Logged elapsed time includes intervals between metric logs; neither is a Modal billing measurement. MFU uses nanochat-compatible rectangular attention and each run’s original peak-compute denominator. Accounting audit and correction provenance. Missing axis values are omitted, not interpolated. Scores are not automatically ranked as comparable experiments.
Loading experiment archive…
0 runs
Reports retain the recorded protocol and units. Download profiler traces and open them in Perfetto. Missing measurements stay missing.