Small T5 (google/t5-efficient-small) trained from scratch for chat sidebar title Prompt format is exactly User: {message}\nTitle: (note the trailing space). Beam search (numbeams=4) is what the eval used. An int4 export lives in the int4/ subfolder (see below). The root repo is the fp16 model. The residual stream of this checkpoint reaches ~82k at the last encoder block, which overflows fp16 max (65504) and produces nan. Train/eval it in bf16, never fp16. fp32 also works. The stored weights are fp16/bfloat16-compatible safetensors. - 2 epochs, 2xT4, bf16, batch 16 x grad-accum 4 - best evalloss 2.7184 (Trainer-reported; inflated ~1.79x by grad-accum reporting, true eval loss ~1.52) Measured…