Skip to content
Xinming Tu·Indexed by this site · Originally publishedDate not given· Evaluation →· Its place in the scan record →SelectedAttention49

Blogger's test claims the latest Claude model may have been trained on large-scale protein sequences

Xinming Tu | Claude was trained on protein sequences at scale

AI summary

Researcher Xinming Tu released an evaluation claiming that the latest Claude model has been trained on large-scale protein sequences. The test used 668 bacterial and archaeal proteins from 167 Pfam families of unknown function for next-amino-acid prediction, with shuffled sequences as a control: Claude Fable 5.2 achieved 54% accuracy on real sequences versus only 9% on shuffled ones, and Opus 5.5 showed a similar gap, while GPT, Gemini, and seven open-weight models including DeepSeek V4.1 Flash and Qwen 3.8 were all near zero. Fable 5.2 surpassed all open protein language models tested, and Opus 5.5 was comparable to ProtGPT3-10B. The author states that this capability also holds across 128 families without papers and extends to 56 proteins first made public in August 2026, after the training cutoff, inferring from this that protein sequences account for a considerable portion of its training data; Opus 5.5 only predicts human and mouse coding sequences and did not learn genomes.

Why it was selected

The evaluation design using shuffled sequences as a control separates the model's knowledge of sequence order from amino acid composition, and this type of method can be transferred to inferences about the training data composition of frontier models.

Read originalxinmingtu.cn