InternLM releases Sage-SciPerception projectors linking scientific encoders to language models
Hugging Face release maps crystals, spectra, sequences and biosignals into LM space; reported scores include 97.35% BEACON ncRNA accuracy and 49.75% QM9S Acc@1.
Why it matters
Shanghai AI Lab’s InternLM organization published Sage-SciPerception on Hugging Face, a set of trainable projectors that map scientific encoder states—molecular graphs, spectra, crystal structures, protein/nucleic-acid sequences, and physiological signals—into a language model’s embedding space while freezing the scientific and language backbones during the mapping stage. Task-specific checkpoints report metrics such as 97.35% accuracy on BEACON ncRNA family classification and 49.75% Acc@1 on QM9S joint-spectrum structure identification. Serving examples use vLLM prompt_embeds; license metadata is not yet specified on the card.
Key facts
- HF repo internlm/Sage-SciPerception created 2026-11-06T07:34:30Z
- Modalities: crystals, spectra, sequences, biosignals, molecular graphs
- BEACON ncRNA family classification accuracy 97.35% (author-reported)
- QM9S joint-spectrum structure ID Acc@1 49.75% (author-reported)
- Inference path uses vLLM prompt_embeds; license TBD on card