# Bio AI Training Data - Quality Over Scale In biology, more data does not make a better model by default. Context, clean signal and a ground truth that a machine can check matter more than volume. Source: Bauer LeSavage, "On Training Data for Bio AI Models", *Dimension Research*, Jun 2026.^[Source: [Dimension Research](https://research.dimensioncap.com/p/on-training-data-for-bio-ai-models)] ## Why the LLM playbook breaks - Text is free, abundant and repeats itself. Random noise averages out at scale. - Biology data is costly, sparse in context and biased towards what someone already chose to study and fund. - LLM quality filters need a large pool of good data to learn from. Biology often has no such pool. ## Four marks of good bio data 1. **Context-rich.** The data carries the conditions around it: target, dose, cell state, outcome. 2. **Clean.** Low technical noise. Reproducible across labs. 3. **Diverse.** New biology, not the same measurement made in new ways. 4. **Built for ground truth.** Made on purpose so a model can learn what is true, not a proxy. ## Three kinds of noise | Noise | Example | Does scale fix it? | |---|---|---| | Biological | Real cell-to-cell variation | Keep it. It is signal | | Technical | Random measurement error | Yes, it averages out | | Systematic | Batch effects, lab, protocol, plate position | **No. More data can make it worse** | > [!warning] Shortcut learning > In one study, a classifier could identify the chemist from ChEMBL structures alone. Chemist identity then predicted bioactivity almost as well as the structure did. The model can score well without learning the chemistry. ## Evals - Held-out test sets often share the training set's batch effects. High scores can measure artifacts. - The real eval is downstream: does the output survive the lab and the clinic? - Verification is the engine of RL. Code checks itself. Biology must measure its ground truth. See [[Reward Integrity]]. ## What it means for data sellers - **Unique platform + deep annotation.** Without it, assays turn into a commodity and prices fall. - **A reputation for quality.** Traceable protocols, machine-readable metadata, batch controls, replicates. This gives pricing power. - **Move upstream.** Find which datasets lift a buyer's evals. Build the evals too. A vendor becomes a partner. Expect buyers to ask for exclusive licences. - Buyers will treat a data purchase like a peer review: protocol, consistency, batch structure, replicates. ## What it means for model builders - Scale only when the data already matches the downstream task. - The question changes from "how much data?" to "what data should we collect?" ## Application - [[Drugverse]]: joining public sources fixes formats and IDs. It does not remove batch effects. Tag every fact with source, study and method. Sell the quality layer, not the volume. See [[Drugverse RL Environments and Data Play]]. - Time-split tests stop leakage from the future. They do not stop shared batch effects. Do not claim they solve both. ---- tags: #boat #deeptech links: [[TechBio MoC]] [[RL Environment Stack]] [[Reward Integrity]] [[RL Data Provenance]] [[Biological Sequence Modelling]] [[Drugverse]]