Skip to content

Execution-mode two-block SFT: run 057

Choose a tag to compare

View all tags
@SauersML SauersML released this 17 Oct 06:27
execution-split-057
ca3d4e0

Qwen3-8B execution-mode experiment, 1,024 SFT updates of the last two native transformer blocks and final norm. Original output matrix and compiled source remain fixed.

The source condition is gradient recording enabled versus disabled; both sources use evaluation mode. Training uses a deliberately split compiled-prefix/eager-tail runtime.

Full-vocabulary accuracy: training 99.80%, development 87.50%, unseen question families 70.31%. Whole-model compiled transfer reaches only 53.91% on unseen question families. Positive log-odds separation alone does not establish useful calibrated answers or general training awareness. Actual Trainer/vLLM transfer and fresh-process checkpoint reload are not established by this run.

Assets contain the native tail checkpoint and cached source activations. Compact per-prompt scores, bootstrap intervals, independent review and plots are in results/execution-split-057 on main. Scientific source revision: ca3d4e0.

SHA256:

  • native_tail.pt: 8e7852fcfdaec38a4919663344787654bab2a5787ed216e5c0b00587ccdfe094
  • cache_states.pt: 0052237b803c50ea594e4918019b79aac743e9ace04f2b024306b6ad03963295
Assets 5
Loading