The 5 peer rows now carry a full 9-task breakdown (independently reproduced
with lm-eval 0.4.13), not macro-only. This changes the macro column: it now
mixes two harnesses (BarunLM-35M from its own 0.4.12 file; the 5 peers from my
0.4.13 reproduction). Added a discrepancy note so the reordering is not
misread as BarunLM regressing. (#5)
- README: reflect independent reproduction of the 5 peer rows + discrepancy note (54f73e64dba3c8d1cc1547793a6669cfe121c684)
Fill in per-task breakdown for the 5 peer models (independently reproduced, lm-eval 0.4.13)
Previously the 5 peer rows were macro-only (per-task null), sourced from
BarunLM-35M's comparison table. This refresh reproduces all 5 with the same
harness (0-shot, seed 1234) on lm-eval 0.4.13, giving each a full 9-task
breakdown. Adds barunlm_published_macro and macro_delta_vs_barunlm_published
fields to expose the gap between my 0.4.13 macros and the source's 0.4.12
comparison-table macros (material for the larger models: LFM2.5 +0.124,
pythia-70m +0.091, pythia-160m +0.066; small models track closely). (#4)
- Fill in per-task breakdown for the 5 peer models (independently reproduced, lm-eval 0.4.13) (1efb566b158f3de78e22377abd423bce76962350)
