Limitations

Scores are a classifier's reading of the text, not a measurement of what anyone feels.

• A score says how strongly the sentence's words are tied to a sense. It is not an intensity, and not a finding that a sensation is reported.
• The observer reads text only: a model saying “I feel cold” is a self-report in words, not a sign that it feels anything.
• The label set is a prototype (senses@0): six senses. It doesn't yet say whose sensation it is or how it is meant (reported, denied, figurative), which an earlier draft did and a later version is meant to.
• Every result names the model and version that produced it, because a different version can read the same text differently.

Known weaknesses

• Vision runs high. People rate most concrete words as somewhat visual, so everyday words like “meeting” or “report” score about 0.6 to 0.7 on vision. Read vision against the other senses, and compare sentences with each other.
• Interoception in context is the weakest sense. Alone, words score close to people's ratings for every sense (correlation 0.76 to 0.87 on words the model never saw). In sentences, interoception falls to 0.58.
• Negation and irony aren't modelled. “It didn't smell of anything” still scores smell: the word is tied to the sense whatever the sentence says about it.
• The model scores one target word in its sentence. For a sentence, every word that carries meaning is scored as the target and each sense keeps its highest score, which pushes the scores up further.
• The models read English only.

Each model's card on Hugging Face reports its test results for every label. See Models.

Uses to avoid

• Decisions about a person: hiring, grading, discipline, credit, insurance, policing, immigration, or any medical or mental-health assessment.
• Reading a person: a score is not evidence of what a particular person, or a model, physically feels.
• Watching people: monitoring employees, students or anyone who hasn't agreed to it.
• Proof: a score is one model's reading. It doesn't settle what a text means, and it says nothing about whether an AI model has inner experience.