Verified75% confidenceFactExact time
In 2022, Hoffmann et al.'s Chinchilla paper redid experiments on four hundred models and concluded that, under the premise of compute-optimality, each parameter should correspond to approximately 20 tokens.
3
Sources
75%
Confidence
Long-term
Relevance
8/19/2026
First Seen
Sources
Rethinking Scaling Laws: Parameters Are Not the Only Answer
twitterjietang8/19/2026
Related Entities
Related Claims
UnverifiedChinchilla论文通过对超过400个不同规模模型的系统性实验得出结论77% similarUnverified2022年Hoffman等人的Chinchilla论文重新校准了算力最优的参数-数据比例，指出此前许多大模型训练不充分75% similarUnverifiedThe paper conducts experiments on the FineWeb-Edu dataset at four model scales, from 60M to 1B parameters.72% similarUnverifiedThe original Anthropic text recommends using a reranking model to narrow from the top 150 down to 20, and they tried passing 5, 10, and 20 chunks and found 20 worked best, but advise readers to test in their own scenarios.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/796925API
curl https://kongchang.com/api/v1/knowledge/claims/796925MCP
get_claim(id=796925)