suchirsalhan's picture
Initial release: summer merge experiments (18,171 merges, 64,180 evaluations)
1e80dec verified
|
Raw History Blame Contribute Delete
5.72 kB
metadata
license: cc-by-4.0
pretty_name: Model Merging & Capability Transfer Dataset
tags:
  - model-merging
  - permutation-alignment
  - mergeability
  - language-models
  - negative-results
configs:
  - config_name: merged_models
    data_files: data/merged_models.parquet
    default: true
  - config_name: merges
    data_files: data/merges.parquet
  - config_name: evaluations
    data_files: data/evaluations.parquet
  - config_name: models
    data_files: data/models.parquet
  - config_name: summary
    data_files: data/summary.parquet

Model Merging & Capability Transfer Dataset

A dataset of merge experiments, not just checkpoints. Every row says which two models were merged, how, under which alignment and configuration, and how the result scored on downstream tasks next to both parents and the naive merge, measured under the same protocol. Negative results, controls, unverified and incomplete runs are included, and their status is explicit.

It contains 18,171 merge experiments, 64,180 evaluations of 18,507 checkpoints, and 59,087 rows in the flat merged_models table.

Configurations

merged_models is the default configuration: one long-format row per pair × method × alignment × benchmark × task × metric, carrying parent, naive and merged scores and the derived merge loss, gain and success. merges, evaluations and models are the linked tables it is built from. summary holds the derived statistics.

from datasets import load_dataset
flat = load_dataset("Cross-Mergeability/model-merging-benchmark", "merged_models")["train"]
evals = load_dataset("Cross-Mergeability/model-merging-benchmark", "evaluations")["train"]

Experiment families

experiment_family merges methods unverified
beetle_controlled 11364 average, qmd, task_arithmetic, ties 1996
beetle_published 869 average, csp, dare, opcm, other, qmd, svc, task_arithmetic, ties 56
cross_architecture 788 average, dare, interpolation, ot_fusion, slerp, task_arithmetic, ties 121
llama_forks 18 chat_vector 0
mergebench 72 average, dare, other, task_arithmetic, ties 72
polypythia 4028 average, interpolation, other, slerp, task_arithmetic, ties 0
toksuite 1032 average, qmd, task_arithmetic, ties 0

How often a merge beats its better parent

These counts are for comparisons where both parents and the merge were scored in the same experiment:

model_family comparisons beats_best_parent
beetle_controlled 9368 1107
beetle_published 378 2
cross_architecture 1232 65
polypythia 5460 0
toksuite 5160 0

Metrics

All losses are signed so that positive means worse than the better parent:

  • Merge loss. For NLL, merged − min(A, B). For accuracy, max(A, B) − merged.
  • Merge gain. naive_merge_loss − method_merge_loss. Positive means the method improved on plain averaging. It does not mean the result is useful.
  • Merge success. The merge strictly beats the better parent on that metric.
  • Alignment gain. Loss of the unaligned twin (same pair, method and configuration) minus loss of the aligned merge.

NLL and accuracy are never mixed in one row. A parent is used as a baseline only when it was scored in the same experiment_id.

Row status (result_status)

result_status rows
parent_baseline_missing 34537
evaluated 21598
unverified 2053
naive_baseline_missing 809
naive_baseline_not_applicable 90
  • evaluated: the merge and both parents were scored under one protocol.
  • parent_baseline_missing: the merge was scored but a parent was not, in that experiment. This happens, for example, for PolyPythia per-paradigm BLiMP, where only the parents' aggregate BLiMP was recorded, and for Beetle published NLL, where only the better-parent NLL was recorded.
  • naive_baseline_missing: there is no plain-average twin in the same experiment.
  • naive_baseline_not_applicable: the method has no plain-average twin by design (chat-vector transfer, OT fusion).
  • unverified: the source flags the value.

Pair × method combinations that were never run have no row.

Known limitations

  • Copies a parent. copies_parent=True marks rungs whose output is one parent's tensors, for example SVC/OPCM single-parent fills. They are not evidence of composition.
  • Aligned equals naive. In the published Beetle rungs, the aligned merges are byte-identical to the naive ones, because the aligner returned the identity; see notes.
  • Calibration overlap.
    • OT fusion is scored on sentences inside the set used to fit the transport plans, so it is provisional.
    • PolyPythia alignment is fitted on the same FLORES blocks it is scored on.
  • Task arithmetic without a shared base. Where the parents share no base, task arithmetic and TIES are parent interpolations, not conventional task vectors (see merge_config).
  • MergeBench. Rows are published scores, not re-evaluated, and the 5-expert merges have no parent_a/parent_b.
  • Not recorded. Parent checkpoint hashes, code commits, evaluation timestamps and most dataset revisions are absent from the historical result files and are left null.
  • Wikitext-scored transport rungs. The 56 Beetle transport rungs scored on wikitext are unverified.

Provenance and reproduction

Every row has a source_file, and sources.json lists every file each adapter read. Rebuild from the repository with python -m merge_dataset.build_dataset; the scripts are in scripts/. The tables are generated, never maintained by hand.

Citation

To be added by the authors.