Slayer1B-Poetry / tutorials /VALIDATION.md
kacperwikiel's picture
Add five native Slayer notebooks: generation, CPT, SFT, RL and measured checkpoint averaging
05c768b verified
|
Raw History Blame Contribute Delete
2.24 kB

What was checked

On 18 November 2026, locally on CPU with PyTorch 2.10.0:

  • Loaded the actual pinned Slayer1B-1.09 update-45,000 checkpoint and verified its checksum.
  • Completed one native 1B-model CPT optimizer update (16-token context). Adapter weights changed, and frozen base parameters had no gradients. Observed training CE: 3.864333.
  • Verified the SFT mask includes the first response token and final EOS, then completed one real-model response-only SFT update. Observed response CE: 4.256570.
  • Completed one on-policy RL optimizer update using two 2-token rollouts. A string-length diagnostic reward deliberately produced unequal rewards (7 and 8), exercising nonzero advantages. Parent-reference KL was 0 at the reset policy. This tests execution, not learning quality or the notebook's verse-form reward.
  • Saved an inference adapter checkpoint.
  • Executed the generation-comparison notebook using local copies of the published artifacts; its displayed sample/metric outputs are embedded. Portable Hub bootstrap cells were restored afterward.
  • Validated all notebook schemas and parsed their ordinary Python cells.

The full default 10-step CPT/SFT lessons and 3-update verse-form RL lesson have not been run as quality experiments. No GPU memory fit, training improvement, human preference gain, Colab end-to-end execution, or distributed scaling result is claimed. Training notebooks contain no invented training outputs.

Checkpoint averaging

  • Inspected the live archive and pinned revision 32a7304dc2208acb3fb4a6e957498116552a4b18.
  • Downloaded 45k, 50k and 55k weights; matched all three SHA256 values to the archive manifest.
  • Confirmed exact source configs, state names/shapes/dtypes, finite values and tied embeddings; produced their equal FP32 average locally.
  • Reloaded the full output, independently spot-checked arithmetic for embedding, attention and normalization tensors, and verified shared embedding/head storage.
  • Checked the helper on small known weighted inputs, including integer-buffer preservation and rejection of negative coefficients and incompatible configs.
  • See AVERAGING_RESULTS.md and merge-comparison.json for the measured pilot; these do not establish a benchmark improvement.