Discussion on low-efficiency data transfer suggests a public transfer learning competition

willdepue · x · 2026-08-20

Will Depue discusses the efficiency of data transfer learning, suggesting that even between highly similar data (e.g., English and Welsh global news), efficiency might be low (1/10 or 1/100), and abysmal for vastly different domains (like Python and DNA). He proposes a public competition, similar to the nanogpt speedrun, to test transfer capabilities: using infinite Chinese data with fixed English data to minimize English PPL or English knowledge evals.

Related event: Researchers Debate Weaker-Than-Expected Cross-Lingual Transfer in LLMs(5 posts)→

Original post →

More from Research

Research channel →