Unverified50% confidenceFactExact time
The pretraining data scale of mainstream large language models often exceeds 1T (trillion) tokens, and lightweight fine-tuning (SFT) also commonly requires billions of tokens.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe capabilities and behavior of large language models are deeply dependent on training data.67% similarUnverifiedResearch shows that under certain conditions (such as a particular piece of text appearing at very high frequency in training data), models may reproduce verbatim content from training data, and in 2023 researchers from Google DeepMind and the University of California, Berkeley successfully extracted original fragments of training data from ChatGPT.67% similarUnverifiedLarge models can be used to automatically generate synthetic training data based on domain documents to construct question-answer pairs, and typically thousands to tens of thousands of training pairs are needed to achieve significant results.66% similarVerifiedThe energy consumption during the training phase of large models is approximately hundreds of times that of the inference phase.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/902377API
curl https://kongchang.com/api/v1/knowledge/claims/902377MCP
get_claim(id=902377)