Show HN: Bongard-mini – An open-weight Jev-like model built on T5
Category: ai-ml
Tags: judgment-model, t5, open-weights, decision-making, calibrated-probabilities, inference
Score: 7.0/10 (Innovation: 7, Technical: 8, Documentation: 8, Utility: 5)
Bongard-mini is an open-weight judgment model built on a T5 encoder-decoder (7.51B parameters) that accepts a shared situation and multiple questions, returning calibrated probability distributions without text generation. It introduces a full-model training approach across judgment, semantic relationship, and action outcome stages, and claims competitive or superior calibration on benchmarks like DecisionBench and typed-decisions despite lower top-label agreement. The design emphasizes shared bidirectional evidence encoding with parallel question branches, offering millisecond-level latency for multiple decisions on one context.
Target audience: AI researchers, ML engineers, and developers building decision-support or agent systems
Repository: https://huggingface.co/AgentBull/bongard-mini · Python · Apache-2.0 · 3 stars
View on Hacker News