Mergeability/merge-sweep-results

Heterogeneous Mergeability — merge sweep results

Per-merge outcomes from the sweep behind Heterogeneous Mergeability: A Quotient-Space Theory of When Neural Networks Compose Across Tokenizers and Architectures. Every row is one merge that was actually executed and scored: a pair of independently trained LMs, a merge operator, and an alignment arm (merge naively, or align the two models into a shared frame first).

This dataset is a snapshot of a sweep that is still running. Snapshot: 2026-08-19 03:51 UTC · current sweep zoo_latest: 2,656 rows, 274 pairs. The sweep was roughly 40% complete when this snapshot was taken and appends new pairs continuously, so every number below will move. Nothing here is a final result. Repo built from mergeability @ cc9bfa6 by src/mergeschool/tools/publish_results.py.

Why this exists, and what the measurement is

The project's claim is that a failed merge decomposes into a coordinate component — the two models put the same knowledge in different bases, and aligning them removes it — and an irreducible knowledge component that no alignment can remove. The experiment that tests this is not any one merge; it is the contrast between two merges of the same pair: one naive, one after alignment, scored identically. So the dataset is built around that contrast rather than around the sweep's file layout.

alignment/naive_vs_aligned.csv is the table that carries the measurement — one row per (pair, operator), both arms on it, and the gap between them.

Read this before the numbers

1. Alignment is close to inert in the current sweep. Of 1,096 (pair, operator) cells that have both a naive and an aligned score, only 138 (12.6%) differ at all — the remaining 87.4% are bit-identical between the arms. Per pair, the best aligned merge beats the best naive merge for 16 of 274 pairs and loses for 2. Median relative damage: naive 0.0579, aligned 0.0559. This is a null result for this configuration, not for the hypothesis. The only alignment family exercised here is permutation over hidden+heads, and most pairs in this run are Beetle models that differ in training curriculum or language, not in initialisation — there is no neuron permutation between them for the solver to undo. In the archived beetle_seed regime, where pairs do differ by seed, alignment moves 458 of 620 cells. Read the null as "permutation alignment bites where a permutation symmetry exists", and wait for the cross-tokenizer cells still queued in this sweep before reading anything stronger.

2. The QMD coordinate/knowledge split is degenerate here. coord_fraction (the share of parameter distance that alignment removes) has median 0.00e+00 and max 3.75e-04 across all 2,656 rows. It is reporting ~zero coordinate component, consistent with caveat 1. Do not cite coord_fraction from this run as a measured decomposition; treat it as "the permutation solver found nothing to remove".

3. transport and transport+aligned are the same rows. Across all 232 cells where both arms exist, they are bit-identical in 100.0% of cases — 0 of the 140 numeric columns differ. (For comparison, naive vs aligned differ on 41 of 164.) The transport+aligned arm currently carries no information beyond transport. Treat it as a duplicate, not as a second condition.

4. There is no jointly-trained ceiling, so MS and nll_ceiling are empty everywhere. has_ceiling is 0 for every row of every sweep in this repo. All quality is measured floor-relative: nll_floor = min(nll_a, nll_b) is the better of the two parents, delta_floor = nll_floor - nll_merge (nats saved; >0 means the merge beat its best parent), and rel_damage = (nll_merge - nll_floor)/nll_floor (scale-free; lower is better). Ignore the MS column.

5. nll_* is not comparable across regimes. beetle_curriculum is scored in nats/token and beetle_language in nats/char (nll_unit says which). Only rel_damage, a ratio, is comparable across the two. Never pool raw nll_merge across regimes.

6. pair_id is truncated. The writer caps it at 120 characters, so 214 of 274 pair ids lose the tail of the second model's name. It is still a unique key within a sweep (no collisions were found), but do not parse model identities out of it. Use pair_label (from the sweep's note field, e.g. classroom vs late, deu vs ell) and eval_langs instead.

7. Three columns are aliases. pair_closeness is literally the cka column. param_coverage is 1.0, align_method is permutation, n_eval is 400 and n_calib is 6144 on every row.

8. Merged model weights for this run are NOT published. Merged checkpoints are scored and then discarded: publishing is gated on an env var that this run does not set, so hf_publish.publish_merged() is a no-op. The 246 model repos currently in the Mergeability org are from an earlier (2026-08-11) sweep and do not correspond to the rows here. Merges in this dataset are reproducible from the parents and the recipe, not downloadable.

Layout

path	rows	one row is	config
overall/by_pair.csv	274	a pair: closeness in, best merge per arm out, gap between	by_pair (default)
alignment/naive_vs_aligned.csv	1096	a (pair, operator) with both arms and the gap — the measurement	naive_vs_aligned
overall/by_operator.csv	40	an operator x arm (pooled and per regime)	by_operator
overall/by_language.csv	58	a cross-lingual pair in beetle_language	by_language
overall/summary.csv	23	a headline number, with its own n	summary
by-pair/<regime>.csv	2656	one executed merge (pair x arm x operator), readable columns	per_merge_long
raw/zoo_latest_shard0.csv	2656	the unmodified 178-column sweep row	raw_current_sweep
archive/previous_sweeps.csv	8116	a row from an earlier, completed sweep	archive_previous_sweeps
provenance.csv	27	a source shard: mtime, sha256, where it went	provenance

Why the pair is the primary axis and not the model. Model identity is not recoverable from this sweep: the only place a model name appears is inside pair_id, which is truncated on 214 of 274 pairs, and there are no parent_a/parent_b columns. More importantly the naive-vs-aligned contrast — the thing the paper is about — is defined on a pair, not on a model; a per-model view would have to average over partners and would destroy exactly the comparison the dataset exists to expose. by-pair/ is split by regime rather than into 274 one-row directories, because 274 single-row CSVs are harder to read than two grouped tables and would defeat the dataset viewer.

What is in the current sweep

regime	rows	pairs	nll unit	arms	median rel_damage
beetle_curriculum	1728	216	token	naive · aligned	0.0403
beetle_language	928	58	char	naive · aligned · transport · transport+aligned	0.2689

• beetle_curriculum — same language pair, different training curriculum (balanced / sequential / simultaneous / classroom / late, and the tiso* data-order schedules).
• beetle_language — different language monolingual and bilingual Beetle models merged against each other.

Operators x arms (pooled over regimes — see caveat 5 before pooling anything else)

operator	arm	n	median rel_damage	beats best parent
average	aligned	274	0.0727	15.7%
average	naive	274	0.0727	15.7%
average	transport	58	0.2327	1.7%
average	transport+aligned	58	0.2327	1.7%
qmd_guided	aligned	274	0.0481	15.7%
qmd_guided	naive	274	0.0505	15.7%
qmd_guided	transport	58	0.1883	1.7%
qmd_guided	transport+aligned	58	0.1883	1.7%
task_arithmetic	aligned	274	0.0317	22.3%
task_arithmetic	naive	274	0.0310	22.3%
task_arithmetic	transport	58	0.1210	12.1%
task_arithmetic	transport+aligned	58	0.1210	12.1%
ties	aligned	274	0.1292	4.4%
ties	naive	274	0.1292	4.4%
ties	transport	58	0.3393	0.0%
ties	transport+aligned	58	0.3393	0.0%

rel_damage is scale-free and lower is better; 0 means the merge matched its better parent. "beats best parent" is the fraction of merges with delta_floor > 0.

Easiest and hardest pairs (best operator per arm)

pair	eval langs	regime	closeness (CKA)	best naive	best aligned	gap
tiso2-burst vs tiso3-block	eng+zho	beetle_curriculum	0.829	-0.1360 (ties)	-0.1360 (ties)	+0.0000
hin vs rus	hin+rus	beetle_language	0.811	-0.0753 (task_arithmetic)	-0.0753 (task_arithmetic)	+0.0000
sequential vs late	eng+zho	beetle_curriculum	0.878	-0.0422 (ties)	-0.0422 (ties)	+0.0000
sequential vs late	eng+spa	beetle_curriculum	0.866	-0.0392 (average)	-0.0392 (average)	+0.0000
sequential vs late	eng+est	beetle_curriculum	0.848	-0.0318 (average)	-0.0318 (average)	+0.0000
tiso2-burst vs tiso3-block	deu+eng	beetle_curriculum	0.644	-0.0259 (task_arithmetic)	-0.0259 (task_arithmetic)	+0.0000
classroom vs late	eng+spa	beetle_curriculum	0.805	-0.0215 (average)	-0.0215 (average)	+0.0000
classroom vs late	eng+hin	beetle_curriculum	0.811	-0.0214 (average)	-0.0214 (average)	+0.0000
…
est vs kor	est+kor	beetle_language	0.792	2.2052 (task_arithmetic)	2.2052 (task_arithmetic)	+0.0000
deu vs kor	deu+kor	beetle_language	0.818	1.9056 (task_arithmetic)	1.9056 (task_arithmetic)	+0.0000
est vs hin	est+hin	beetle_language	0.801	1.6644 (task_arithmetic)	1.6644 (task_arithmetic)	+0.0000
fin vs hin	fin+hin	beetle_language	0.768	1.3783 (task_arithmetic)	1.3783 (task_arithmetic)	+0.0000
deu vs hin	deu+hin	beetle_language	0.781	1.3641 (task_arithmetic)	1.3641 (task_arithmetic)	+0.0000
fin vs kor	fin+kor	beetle_language	0.785	1.3133 (task_arithmetic)	1.3133 (task_arithmetic)	+0.0000
est vs jpn	est+jpn	beetle_language	0.760	1.2774 (task_arithmetic)	1.2774 (task_arithmetic)	+0.0000
fin vs jpn	fin+jpn	beetle_language	0.788	1.1633 (task_arithmetic)	1.1633 (task_arithmetic)	+0.0000

What you may and may not conclude

May. That across 2,656 real merges of 274 pairs of small multilingual LMs, task_arithmetic is the least damaging operator and ties the most, in both arms; that most merges do not beat their better parent (12.7% of 2,656 rows do); that pairs with higher CKA are cheaper to merge; and that permutation alignment over hidden+heads changes almost nothing for models that share an initialisation.

May not. That alignment does not help in general — the cells that would test that (cross-tokenizer, cross-architecture) are still queued, and the archived seed regime already shows alignment biting when a permutation symmetry exists. That coord_fraction ≈ 0 measures a real absence of coordinate mismatch (caveat 2). That transport+aligned is evidence about transport-plus-alignment (caveat 3). That any absolute nll is comparable across regimes (caveat 5). That these numbers are final — the sweep is ~40% done.

One relationship that does not survive pooling

Representational closeness predicts merge cost — but only inside beetle_curriculum. Correlating pair_closeness (CKA) against the pair's best naive rel_damage:

cut	n pairs	Pearson	Spearman
beetle_curriculum	216	-0.579	-0.430
beetle_language	58	-0.027	+0.038
(pooled)	274	-0.092	-0.309

Within the curriculum regime the effect is real and strong. Within the language regime it is absent. Pooled across both it is a mirage produced by the two regimes sitting at different damage levels — the classic Simpson's-paradox shape, and a reason to cut every predictor by regime before believing it.

Provenance

Derived tables are computed only from zoo_latest. The archive sweeps are published alongside, in their own config, with a sweep column — they were produced by earlier revisions of the emitter (160–178 columns) and are never blended into a headline number.

sweep	shard	rows	cols	source mtime (UTC)	status	published as
zoo	zoo_shard0	324	160	2026-08-09T15:47:12Z	complete	archive/previous_sweeps.csv
zoo	zoo_shard1	328	154	2026-08-09T16:02:39Z	complete	archive/previous_sweeps.csv
zoo	zoo_shard2	356	160	2026-08-09T16:00:04Z	complete	archive/previous_sweeps.csv
zoo	zoo_shard3	340	160	2026-08-09T15:55:51Z	complete	archive/previous_sweeps.csv
zoo	zoo_shard4	328	154	2026-08-09T16:04:46Z	complete	archive/previous_sweeps.csv
zoo	zoo_shard5	328	160	2026-08-09T15:48:58Z	complete	archive/previous_sweeps.csv
zoo	zoo_shard6	304	160	2026-08-09T15:56:18Z	complete	archive/previous_sweeps.csv
zoo_seed	zoo_shard0	344	143	2026-08-09T22:59:30Z	complete	archive/previous_sweeps.csv
zoo_seed	zoo_shard1	352	137	2026-08-09T23:03:59Z	complete	archive/previous_sweeps.csv
zoo_seed	zoo_shard2	336	137	2026-08-09T22:57:46Z	complete	archive/previous_sweeps.csv
zoo_seed	zoo_shard3	336	137	2026-08-09T22:56:49Z	complete	archive/previous_sweeps.csv
zoo_seed	zoo_shard4	336	126	2026-08-09T22:57:22Z	complete	archive/previous_sweeps.csv
zoo_seed	zoo_shard5	352	143	2026-08-09T23:07:55Z	complete	archive/previous_sweeps.csv
zoo_seed	zoo_shard6	360	137	2026-08-09T23:08:10Z	complete	archive/previous_sweeps.csv
zoo_seed	zoo_shard7	336	131	2026-08-09T22:59:45Z	complete	archive/previous_sweeps.csv
zoo_seed_v2	zoo_shard0	296	148	2026-08-11T01:49:12Z	complete	archive/previous_sweeps.csv
zoo_seed_v2	zoo_shard1	296	148	2026-08-11T01:42:31Z	complete	archive/previous_sweeps.csv
zoo_seed_v2	zoo_shard2	280	142	2026-08-11T01:34:53Z	complete	archive/previous_sweeps.csv
zoo_seed_v2	zoo_shard3	280	142	2026-08-11T01:37:47Z	complete	archive/previous_sweeps.csv
zoo_seed_v2	zoo_shard4	280	136	2026-08-11T01:37:22Z	complete	archive/previous_sweeps.csv
zoo_seed_v2	zoo_shard5	296	154	2026-08-11T01:51:44Z	complete	archive/previous_sweeps.csv
zoo_seed_v2	zoo_shard6	304	142	2026-08-11T01:55:04Z	complete	archive/previous_sweeps.csv
zoo_seed_v2	zoo_shard7	288	136	2026-08-11T01:37:45Z	complete	archive/previous_sweeps.csv
zoo_2026_08_17	zoo_shard0	256	142	2026-08-17T02:26:28Z	complete	archive/previous_sweeps.csv
zoo_2026_08_17	zoo_shard1	240	142	2026-08-17T02:24:43Z	complete	archive/previous_sweeps.csv
zoo_2026_08_17	zoo_shard2	240	154	2026-08-17T02:16:48Z	complete	archive/previous_sweeps.csv
zoo_latest	zoo_shard0	2656	178	2026-08-19T03:49:28Z	live (appending during snapshot)	raw/zoo_latest_shard0.csv

Reproducing / refreshing

git clone https://github.com/suchirsalhan/mergeability && cd mergeability && pip install -e .

python3 -m mergeschool.tools.publish_results --dry-run   # rebuild every table locally
export HF_TOKEN=...                                      # write access to the Mergeability org
python3 -m mergeschool.tools.publish_results --apply     # rebuild, upload, then poll the viewer

Re-running after the sweep finishes overwrites this repo in place, including every number in this card — they are computed at build time, not typed.