Unverified50% confidenceFactExact time
GPT-2的词表规模约5万，GPT-2 small架构为12层、每层12个注意力头
1
Sources
50%
Confidence
Long-term
Relevance
8/20/2026
First Seen
Sources
Related Claims
UnverifiedThe 355M version of GPT-2 uses 16 attention heads, with each head having a dimension of 64.78% similarUnverifiedGPT-2的BPE在字节级别操作，其merge表包含约5万条规则，最终词表为50,257个token76% similarVerifiedGPT-3 has a vocabulary of approximately 50,000 tokens.73% similarUnverifiedGPT-2 was built by scaling up more than 10 times beyond GPT-1, with 1.5 billion parameters.72% similarUnverified2023年研究者在 GPT-2 中识别出负责间接宾语识别的完整电路，涉及不足1%的注意力头71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/777219API
curl https://kongchang.com/api/v1/knowledge/claims/777219MCP
get_claim(id=777219)