Qwen's Junyang Lin: there's no deterministic way to judge a base model's potential

JustinLin610 · x · 2026-11-18

Junyang Lin (JustinLin610) of Alibaba's Qwen team says the old trick of light SFT followed by benchmarks still over-indexes on post-training benchmark scores instead of assessing a base model's underlying potential—its knowledge and reasoning sharpness. He argues there is no deterministic way to solve this evaluation problem.

Related event: Qwen's Lin: Lightweight SFT Benchmarks Can't Reveal Base Model Potential(2 posts)→

Original post →

More from Models

Models channel →