BoggersTheFish

/

TensionLM-117M-Curriculum

like 0

Hey everyone,

I'm an independent researcher who just submitted "TensionLM: Sigmoid Tension as Constraint Relaxation for Language Modelling" to arXiv (cs.AI + cs.LG).

• 117M model with full public code + weights
• Sigmoid tension attention (replaces softmax) + TS-native auxiliary losses
• Logic → Language → Math curriculum gives 96× better first-contact math PPL

Would really appreciate an endorsement so the paper can go live.

Endorsement code: RLU86M

Paper + code: https://github.com/BoggersTheFish/bozo
Model: https://huggingface.co/BoggersTheFish/TensionLM-117M-Curriculum

Thanks so much!