nanojev-coreml / source /STATUS.md
alexwengg's picture
Publish NanoJev Core ML conversion source without trained weights
dd6f7d9 verified
|
Raw
History Blame Contribute Delete
981 Bytes

NanoJev status

  • Pinned current checkpoint and audited native source; full checkpoint SHA256 matches upstream and contains 596,250,498 trained parameters across 322 tensors. Historical tracker revision unresolved.
  • Pinned full checkpoint downloaded. Native Qwen3 encoder plus decision-head exporter matched the trained model on Choice, Boolean, and Score requests: same argmax for all three, maximum logit error 1.67e-6.
  • FP16 Core ML L128/K4 encoder and full decision-head packages exported in 22.7 seconds. Encoder package is about 1.1 GiB; head package is about 420 KiB.
  • Core ML runtime parity passed for Choice, Boolean, and Score: 3/3 chosen answers matched trained PyTorch. Maximum probability errors were 0.003851, 0.000847, and 0.005882, respectively.
  • No full Decision Index or 2048 evaluation has been run for this Core ML package.
  • Trained-weight redistribution license unresolved. Public HF model artifact is not yet authorized by the source metadata.