Dataset: SQuAD v1.1 — 87,599 training Q/A pairs
Hardware: NVIDIA GB10 (Blackwell, sm_121a)
Engine config: CUDA arch 121a, scalar_flash, legacy_bf16, optimizer lion.
IDA family seats (metrics only — no weights)
family
arch
hidden
layers
heads
vocab
phase
final loss
wallclock (s)
micro-steps
tokens seen
tok/s
steps/s
ms/step
skipped opt
grad norm
compute
edge
ida_lattice
384
8
6
32000
success
10.161242
642.088… See the full description on the dataset page:
https://huggingface.co/datasets/KissTheHabit/IDA-TRAIN-V2-mlperf-squad-gb10-results.
