Datology claims 6x compute multiplier from data curation alone; 12B beats 30B baseline
pratyushmaini · x · 2026-11-17
Headline results
Datology AI published a tech report on its Curation Studio claiming a 6x pretraining compute multiplier from data curation alone: at every model size tested, models trained on curated data beat a strong baseline built from the best publicly curated datasets.
Key numbers
- +9.1 average benchmark score: pretraining both models on 1T tokens from identical datasets and changing only curation, a fully post-trained 30B-A3B model scored 46.8% vs the baseline's 37.7%.
- 12B beats 30B: a 12B-A1.4B model trained on 440B curated tokens beat the 30B-A3B baseline on overall math and code ability with just a fifth of the pretraining compute and fewer than half the active parameters.
Methodology
The team stresses that data research demands more scientific control than training one model — small changes can confound results.
More from Infra
- Evolving Infra beats NVIDIA Dynamo on agent workloads: 4.1x lower latency, 14x lower P95 TTFT — YouJiacheng · 2026-11-17
- Open-source models hit 30% enterprise token share in 5 months, up from near zero — pwendell · 2026-11-17
- Databricks opens its cost-vs-quality Pareto benchmark powering the Unity Gateway Smart Router — pwendell · 2026-11-17
- Together AI's Five-Step Playbook for Migrating From Closed to Open Models — togethercompute · 2026-11-17
- Why tensor parallelism breaks RNG reproducibility in model init — ezyang · 2026-11-17
- Bittensor's Engy tops Venice's AI inference volume despite 28x smaller valuation — markjeffrey · 2026-11-17