model-merging-benchmark / sources.json
suchirsalhan's picture
Initial release: summer merge experiments (18,171 merges, 64,180 evaluations)
1e80dec verified
Raw History Blame Contribute Delete
9.05 kB
{
"beetle": [
"results/beetle-tables/figures/merge_frame.csv",
"results/beetle-merge-eval/beetle_merge_eval_long.csv",
"results/beetle-merge-eval/arm_identity_audit.csv",
"results/beetle-merge-eval/coverage_by_merge.csv",
"results/beetle_zoo_catalogue.csv",
"results/provenance/provenance_labels.csv",
"results/provenance_merge_bench/provenance_merge_bench.csv",
"results/provenance_merge_bench/transport_seed_shard.csv",
"results/provenance_merge_bench/provenance_merge_bench_v2.csv"
],
"polypythia_toksuite": [
"results/compose-audit/results/set1_14m.jsonl",
"results/compose-audit/results/set1_31m.jsonl",
"results/compose-audit/results/set1_70m.jsonl",
"results/compose-audit/results/set1_160m.jsonl",
"results/compose-audit/results/set1_410m.jsonl",
"results/compose-audit/results/set1x_410m.jsonl",
"results/compose-audit/results/set1_pairs.csv",
"results/compose-audit/results/blimp_14m.jsonl",
"results/compose-audit/results/blimp_31m.jsonl",
"results/compose-audit/results/blimp_70m.jsonl",
"results/compose-audit/results/blimp_160m.jsonl",
"results/compose-audit/results/blimp_410m.jsonl",
"results/compose-audit/results/blimpB_410m.jsonl",
"results/compose-audit/results/slerp_14m.jsonl",
"results/compose-audit/results/slerp_31m.jsonl",
"results/compose-audit/results/slerp_70m.jsonl",
"results/compose-audit/results/slerp_160m.jsonl",
"results/compose-audit/results/repair_14m.jsonl",
"results/compose-audit/results/repair_31m.jsonl",
"results/compose-audit/results/repair_70m.jsonl",
"results/compose-audit/results/repair_160m.jsonl",
"results/compose-audit/results/corpus_14m.jsonl",
"results/compose-audit/results/corpus_160m.jsonl",
"results/compose-audit/results/abl_160m-data.jsonl",
"results/compose-audit/results/abl_160m-weight.jsonl",
"results/merge-accuracy/results/ledger.jsonl",
"results/merge-accuracy/code/pairs_r4.json",
"data/sweeps/toksuite/zoo_shard0.csv",
"data/sweeps/toksuite/zoo_shard1.csv",
"data/sweeps/toksuite/zoo_shard2.csv",
"data/sweeps/toksuite/zoo_shard3.csv",
"data/sweeps/toksuite/zoo_shard4.csv",
"data/sweeps/toksuite/zoo_shard5.csv",
"results/compose-audit/README.md",
"results/compose-audit/RESULTS_COMPOSE_AUDIT.md",
"results/compose-audit/results/rung_summary.csv",
"results/compose-audit/code/set1_polypythia.py",
"results/compose-audit/code/set1_blimp.py",
"results/compose-audit/code/set1_slerp.py",
"results/compose-audit/code/set1_repair.py",
"results/compose-audit/code/set1_corpus_robustness.py",
"results/compose-audit/code/common.py",
"results/compose-audit/code/run_set1.sh",
"results/compose-audit/code/run_blimp.sh",
"results/compose-audit/code/run_blimp410.sh",
"results/compose-audit/code/run_repair.sh",
"results/compose-audit/code/run_abl.sh",
"results/merge-accuracy/README.md",
"results/merge-accuracy/RESULTS_MERGE_ACCURACY.md",
"results/merge-accuracy/code/run_pairs.py",
"results/toksuite/TOKSUITE_RESULTS.md",
"results/toksuite/PREFLIGHT.md",
"results/toksuite/toksuite_pair_level.csv",
"src/mergeschool/controlled/emit_lm.py",
"src/mergeschool/core/models.py",
"src/mergeschool/core/merge.py",
"src/mergeschool/core/eval.py",
"src/mergeschool/core/data.py",
"results/exp39_results.json",
"results/exp40_results.json"
],
"xarch_llama_mergebench": [
{
"path": "results/crossarch-1b-diagnostics/results/table_model_roster.csv",
"family": "cross_architecture",
"used": true,
"note": "the five ~1B models (Pythia-1.4B, Zh-Pythia-1.4B, Tucano-1b1, Bielik-1.5B-v3, Minerva-1B): HF id, language, architecture, vocab"
},
{
"path": "results/crossarch-1b-diagnostics/results/transport_merge.csv",
"family": "cross_architecture",
"used": true,
"note": "OT fusion (Transport-and-Merge reimplementation), 20 directed pairs x alpha {0,0.01,0.03,0.1}; learned plan and shuffled-plan control; alpha=0 is the unfused target (reference). Per-direction transport_<t>__<d>.csv files hold the same rows and are not read separately."
},
{
"path": "results/crossarch-1b-diagnostics/results/crossmodel_native_merge.csv",
"family": "cross_architecture",
"used": true,
"note": "Pythia-EN x Zh-Pythia native lerp/slerp of transformer blocks (embeddings excluded)"
},
{
"path": "results/crossarch-1b-diagnostics/results/checkpoint_pairs.csv",
"family": "cross_architecture",
"used": true,
"note": "within-run checkpoint merges (78 Pythia-1.4B + 28 Zh-Pythia-1.4B pairs): NLL of lerp 0.25/0.5/0.75, TIES, task arithmetic, DARE"
},
{
"path": "results/crossarch-1b-accuracy/results/table1_checkpoint_accuracy.csv",
"family": "cross_architecture",
"used": true,
"note": "accuracy of parents and naive alpha=0.5 merge for 105 of the 106 checkpoint pairs (zhpythia 1000|20000 missing)"
},
{
"path": "results/merge-accuracy/results/ledger.jsonl",
"family": "cross_architecture",
"used": true,
"note": "merge-accuracy R4 cross-group pair pythia-1.4b x Zh-Pythia-1.4B: naive/aligned interpolation and TIES, downstream accuracy. Placed in cross_architecture (it is the Pythia x Zh-Pythia pair), not humanscale."
},
{
"path": "results/exp07_results.json",
"family": "cross_architecture",
"used": false,
"note": "mode='demo' synthetic substrates; no checkpoint, no merge -> no rows"
},
{
"path": "results/exp18_results.json",
"family": "cross_architecture",
"used": false,
"note": "mode='demo' synthetic OT-recovery toy; no checkpoint, no merge -> no rows"
},
{
"path": "results/exp31_results.json",
"family": "cross_architecture",
"used": false,
"note": "mode='demo' synthetic width-mismatch toy; no checkpoint, no merge -> no rows"
},
{
"path": "results/extrinsic/extrinsic_evaluations.csv",
"family": "cross_architecture",
"used": false,
"note": "derived union view of the files above; primary files read instead"
},
{
"path": "results/merge-accuracy/results/chatvec.jsonl",
"family": "llama_forks",
"used": true,
"note": "chat-vector transfer (Llama-3.1-8B-Instruct - Llama-3.1-8B) onto Swallow v0.1/v0.2, Typhoon2, SEA-LION; naive and permutation-aligned; lambda 0.5/1.0; permuted-fork controls; Belebele, ARC-Easy, IFEval strict prompt/instruction level"
},
{
"path": "results/merge-accuracy/results/chatvec_summary.csv",
"family": "llama_forks",
"used": false,
"note": "derived from chatvec.jsonl"
},
{
"path": "results/merge-accuracy/results/ecosystem_screen.json",
"family": "llama_forks",
"used": false,
"note": "frame-drift screen of 10 more Llama-3.1-8B derivatives; diagnostics only, no merge"
},
{
"path": "results/mergebench-property-sweep/suite.json",
"family": "mergebench",
"used": true,
"note": "40 MergeBench expert checkpoints, 8 base families"
},
{
"path": "results/mergebench-property-sweep/pairs_w0.csv",
"family": "mergebench",
"used": true,
"note": "base model id and d_model per family (column 'parent'); pair properties only, no pairwise merge was built"
},
{
"path": "results/mergebench-property-sweep/pairs_w1.csv",
"family": "mergebench",
"used": true,
"note": "as pairs_w0.csv"
},
{
"path": "results/mergebench-property-sweep/mergebench_published_scores.csv",
"family": "mergebench",
"used": true,
"note": "published 5-expert merge scores (arXiv:2505.10833 Tables 8, 10-17); identical to results/mergebench/mergebench_published_scores.csv"
},
{
"path": "results/mergebench-property-sweep/pair_metrics.csv",
"family": "mergebench",
"used": false,
"note": "80 pre-merge pair properties; no merge outcome exists for any pair -> no rows"
},
{
"path": "results/mergebench/",
"family": "mergebench",
"used": false,
"note": "earlier, smaller run (40 pairs, 4 families) of the property sweep; pairs_w0/w1, pair_metrics, table_pairs, table_family and ledger_* DIFFER from the property-sweep copies (superseded). Every ' 2.csv'/' 2.json' duplicate in it is byte-identical to its original and is ignored."
},
{
"path": "results/merge-accuracy/results/",
"family": "humanscale",
"used": false,
"note": "merge-accuracy holds no HumanScale merge; its only non-llama rows are the cross-group pair (see cross_architecture)"
},
{
"path": "results/backlog_tier01/e15_*.csv",
"family": "humanscale",
"used": false,
"note": "E15 Sinkhorn-eps sweep on HumanScale pairs is layer STITCHING (run_stitch.hybrid_sd), not weight merging -> no rows"
}
]
}