Exact proposal revision
With documents paired across arms, as in the 0.5B analysis, is the pooled Polish bits-per-byte cost of a 30% math and code share in 1B tokens of recovery pretraining of an APT4 transplant of Qwen2.5-1.5B at most 0.05?
Recomputation only: the per-document scores of the final checkpoints of the 0% and 30% arms of the 1.5B confirmation on the pooled 800 Polish documents (four texts of 200 documents each, the same documents in the same order in both arms), byte-weighted bits per byte, 2,000 bootstrap resamples of paired documents with seed 20260707, the estimator of the 0.5B analysis. No training or scoring.
Access and suggested protocol
- Access needs
- No restricted materials: the per-document score files are public files of experiments/E06-math-code-rescue.
- Suggested protocol
- Commit a design that states the estimator before computing. Inputs: the conf00_final and conf30_final score files of the four Polish texts in experiments/E06-math-code-rescue/results/perdoc; check that document byte counts match between arms before resampling. Estimator: the paired pooled-cost bootstrap of scripts/06_analyze.py applied to those files. Bin as the 0.5B rule: confirmed if the upper 95% bound is at most 0.05, refuted if the lower bound exceeds 0.05, partial otherwise. Report the paired estimate, interval and bin next to the reported unpaired interval.
Selected exact hypotheses and premises
Hypothesis · H15
ed4b79d2-8435-4a8b-b3c2-f466d5084d7bThe cost bound of 0.05 bits per byte, applied to the 1.5B confirmation.
Reason for this revision
Initial proposal.