library_name: transformers
tags:
- model-merging
- mergeability
- training-free
- quotient-merge-distance
beetle-arch-pico-beetle-arch-moe__task_arithmetic__aligned
Training-free merged checkpoint from the Mergeability sweep
(benchmark/emit_lm.py --real), produced by weight-space merging of two independently
trained parents. No gradient steps were taken.
| field | value |
|---|---|
| pair_id | beetle-arch-pico-beetle-arch-moe |
| parent_a | Beetle-FineWeb-2B/beetle-bilingual-l2-50-sequential-33-67-b3-fineweb-2b-nld-eng |
| parent_b | Beetle-nld-eng-Variants/beetle-bilingual-l2-50-sequential-33-67-b3-fineweb-2b-nld-eng-arch-moe |
| ceiling | (floor-relative) |
| operator | task_arithmetic |
| alignment | aligned |
| align_method | permutation |
| regime | different_corpora |
| eval_langs | eng+nld |
| nll_merge | 6.4539 |
| nll_floor | 5.849 |
| param_coverage | 0.4887 |
| MS | None |
How it was made
Parents were loaded, activations extracted on a shared calibration corpus, and the merge applied
either naive (parents combined in their own coordinates) or aligned (parent B carried into
parent A's residual-stream basis via common.alignment.residual_basis_map before merging —
permutation for same-width pairs, orthogonal/rectangular for cross-width).
MS is the recovery score from common.eval.mergeability_score (merged vs. floor vs. ceiling), the
same normalisation used by Zhou et al., so it is comparable across rows of the sweep.
Caveats
Sub-1B merges are noisy; an aligned signal where the naive one is noise is the finding, not a bug. Rows without a joint ceiling are floor-relative and must not be read as absolute recovery.
Generated automatically — see the mergeability repo.