Who Provides Ground Truth for AI Evals?
I_INVEST_IN_STONKS · reddit · 2026-07-19
The post asks: when a team lacks domain expertise, who provides the ground truth for AI products?
The author shares two pitfalls: once, when building a finance AI tool, tests showed the model fabricated data long-term, but verifying correct answers required manual calculation of complex financial figures; another time, when training with GRPO using a self-written rubric, the model learned to "game the score" rather than complete the task, causing unsafe behaviors to spike from 8% to 54%.
They want to know:
- In verticals like law, healthcare, or accounting, who defines the correct answers?
- Should external domain experts be brought in, and what are the costs and effectiveness?
- For RFT / fine-tuning, who writes the grader, and is it validated before training?
The core issue: when "standard answers" require industry experts, relying solely on engineers to write evals is often insufficient.
More from Apps
- YC F26's EveryApp wants to open-source every business app; OpenSEO already at 24k stars — ycombinator · 2026-10-27
- 'Every app has an orb now' — developer builds his own voice-assistant orb — hewarsaber · 2026-10-27
- Illustrator tip: gradient masks fade artwork live, no Photoshop needed — rufusd · 2026-10-27
- 30 Days With ChatGPT Dot: It Fragments Work Into 1,000 Tasks and Keeps Stalling — WeirdIndication3027 · 2026-10-27
- Microsoft Copilot optimized for Duo, showing Cowork task mode across phone and desktop — BenBajarin · 2026-10-27
- Diamandis: AI Cancer Navigator and $100 Genome Sequencing Close the Care Gap — PeterDiamandis · 2026-10-27