Qwen's Lin: Lightweight SFT Benchmarks Can't Reveal Base Model Potential

Justin Lin of Alibaba's Qwen team says lightweight SFT followed by benchmarks is a flawed way to evaluate base models, as the community over-focuses on post-training scores and no deterministic method exists to judge a base model's potential.

2026-11-18 ~ 2026-11-19 · 2 related posts