Half the compute for the same model.
A discovery I made, turned into a service. Before a single GPU-hour is spent, Echo reads a few million words sampled from your target and your data pool — on CPU, in seconds, without touching your model — and issues a sealed, timestamped verdict: curate toward this target, or don't.
When the call is GO, the curated run reaches the baseline's best quality with up to 63% less training compute; when it's DON'T, it keeps you from curating into a run that would never catch up. Head to head, it beat the most advanced published curation methods — including one that spent three times the compute.
- Medical (PubMed)63%
- Legal60%
- Recipes58%
- Medical · 8× larger model56%
- Python documentation52%
- News35%
- Encyclopedia, tiny sampleDON'T · called right
Steps the curated run needs to reach the random baseline's best held-out loss, four seeds per arm, every call sealed before training. Validated so far on text models up to 41.7M parameters and a 256M vision-language model; larger models are the next sealed tests.
