Datasets:

NagaYu
/

stemma-bench

a stringclasses 19 values	b stringclasses 19 values	related bool 2 classes	direction stringclasses 3 values	relation stringclasses 17 values	a_family stringclasses 3 values	b_family stringclasses 3 values	a_op stringclasses 15 values	b_op stringclasses 15 values	weak_weight_lineage bool 2 classes
gpt2-root	distilgpt2	true	a->b	distilled	gpt2	gpt2	base	distilled	true
smollm2-135m-root	smollm2-sft	true	a->b	sft	smollm2	smollm2	base	sft	false
smollm2-135m-root	smollm2-lora-merged	true	a->b	lora_merged	smollm2	smollm2	base	lora_merged	false
smollm2-135m-root	smollm2-cpt	true	a->b	ancestor	smollm2	smollm2	base	continued_pretrain	false
smollm2-sft	smollm2-cpt	true	a->b	continued_pretrain	smollm2	smollm2	sft	continued_pretrain	false
smollm2-135m-root	smollm2-int8	true	a->b	quantized_int8	smollm2	smollm2	base	quantized_int8	false
smollm2-135m-root	smollm2-sft-int4	true	a->b	ancestor	smollm2	smollm2	base	quantized_int4	false
smollm2-sft	smollm2-sft-int4	true	a->b	quantized_int4	smollm2	smollm2	sft	quantized_int4	false
gpt2-root	gpt2-int8	true	a->b	quantized_int8	gpt2	gpt2	base	quantized_int8	false
smollm2-135m-root	smollm2-prune-mag30	true	a->b	pruned_magnitude	smollm2	smollm2	base	pruned_magnitude	false
smollm2-135m-root	smollm2-sft-prune-struct10	true	a->b	ancestor	smollm2	smollm2	base	pruned_structured	false
smollm2-sft	smollm2-sft-prune-struct10	true	a->b	pruned_structured	smollm2	smollm2	sft	pruned_structured	false
smollm2-135m-root	smollm2-vocab-ext	true	a->b	vocab_extended	smollm2	smollm2	base	vocab_extended	false
smollm2-135m-root	smollm2-vocab-ext-trained	true	a->b	vocab_extended_trained	smollm2	smollm2	base	vocab_extended_trained	false
smollm2-135m-root	smollm2-merge-slerp	true	a->b	ancestor	smollm2	smollm2	base	merge_slerp	false
smollm2-lora-merged	smollm2-merge-slerp	true	a->b	merge_slerp	smollm2	smollm2	lora_merged	merge_slerp	false
smollm2-sft	smollm2-merge-slerp	true	a->b	merge_slerp	smollm2	smollm2	sft	merge_slerp	false
smollm2-135m-root	smollm2-merge-ties2	true	a->b	ancestor	smollm2	smollm2	base	merge_ties	false
smollm2-cpt	smollm2-merge-ties2	true	a->b	merge_ties	smollm2	smollm2	continued_pretrain	merge_ties	false
smollm2-sft	smollm2-merge-ties2	true	a->b	merge_ties	smollm2	smollm2	sft	merge_ties	false
smollm2-135m-root	smollm2-merge-ties3	true	a->b	ancestor	smollm2	smollm2	base	merge_ties	false
smollm2-cpt	smollm2-merge-ties3	true	a->b	merge_ties	smollm2	smollm2	continued_pretrain	merge_ties	false
smollm2-lora-merged	smollm2-merge-ties3	true	a->b	merge_ties	smollm2	smollm2	lora_merged	merge_ties	false
smollm2-sft	smollm2-merge-ties3	true	a->b	merge_ties	smollm2	smollm2	sft	merge_ties	false
smollm2-135m-root	smollm2-merge-dare2	true	a->b	ancestor	smollm2	smollm2	base	merge_dare	false
smollm2-lora-merged	smollm2-merge-dare2	true	a->b	merge_dare	smollm2	smollm2	lora_merged	merge_dare	false
smollm2-sft	smollm2-merge-dare2	true	a->b	merge_dare	smollm2	smollm2	sft	merge_dare	false
smollm2-135m-root	smollm2-distil-half	true	a->b	distilled	smollm2	smollm2	base	distilled	true
distilgpt2	gpt2-int8	true	sibling	sibling	gpt2	gpt2	distilled	quantized_int8	true
smollm2-sft	smollm2-lora-merged	true	sibling	sibling	smollm2	smollm2	sft	lora_merged	false
smollm2-sft	smollm2-int8	true	sibling	sibling	smollm2	smollm2	sft	quantized_int8	false
smollm2-sft	smollm2-prune-mag30	true	sibling	sibling	smollm2	smollm2	sft	pruned_magnitude	false
smollm2-sft	smollm2-vocab-ext	true	sibling	sibling	smollm2	smollm2	sft	vocab_extended	false
smollm2-sft	smollm2-vocab-ext-trained	true	sibling	sibling	smollm2	smollm2	sft	vocab_extended_trained	false
smollm2-sft	smollm2-distil-half	true	sibling	sibling	smollm2	smollm2	sft	distilled	true
smollm2-lora-merged	smollm2-cpt	true	sibling	cousin	smollm2	smollm2	lora_merged	continued_pretrain	false
smollm2-lora-merged	smollm2-int8	true	sibling	sibling	smollm2	smollm2	lora_merged	quantized_int8	false
smollm2-lora-merged	smollm2-sft-int4	true	sibling	cousin	smollm2	smollm2	lora_merged	quantized_int4	false
smollm2-lora-merged	smollm2-prune-mag30	true	sibling	sibling	smollm2	smollm2	lora_merged	pruned_magnitude	false
smollm2-lora-merged	smollm2-sft-prune-struct10	true	sibling	cousin	smollm2	smollm2	lora_merged	pruned_structured	false
smollm2-lora-merged	smollm2-vocab-ext	true	sibling	sibling	smollm2	smollm2	lora_merged	vocab_extended	false
smollm2-lora-merged	smollm2-vocab-ext-trained	true	sibling	sibling	smollm2	smollm2	lora_merged	vocab_extended_trained	false
smollm2-lora-merged	smollm2-merge-ties2	true	sibling	cousin	smollm2	smollm2	lora_merged	merge_ties	false
smollm2-lora-merged	smollm2-distil-half	true	sibling	sibling	smollm2	smollm2	lora_merged	distilled	true
smollm2-cpt	smollm2-int8	true	sibling	cousin	smollm2	smollm2	continued_pretrain	quantized_int8	false
smollm2-cpt	smollm2-sft-int4	true	sibling	sibling	smollm2	smollm2	continued_pretrain	quantized_int4	false
smollm2-cpt	smollm2-prune-mag30	true	sibling	cousin	smollm2	smollm2	continued_pretrain	pruned_magnitude	false
smollm2-cpt	smollm2-sft-prune-struct10	true	sibling	sibling	smollm2	smollm2	continued_pretrain	pruned_structured	false
smollm2-cpt	smollm2-vocab-ext	true	sibling	cousin	smollm2	smollm2	continued_pretrain	vocab_extended	false
smollm2-cpt	smollm2-vocab-ext-trained	true	sibling	cousin	smollm2	smollm2	continued_pretrain	vocab_extended_trained	false
smollm2-cpt	smollm2-merge-slerp	true	sibling	sibling	smollm2	smollm2	continued_pretrain	merge_slerp	false
smollm2-cpt	smollm2-merge-dare2	true	sibling	sibling	smollm2	smollm2	continued_pretrain	merge_dare	false
smollm2-cpt	smollm2-distil-half	true	sibling	cousin	smollm2	smollm2	continued_pretrain	distilled	true
smollm2-int8	smollm2-sft-int4	true	sibling	cousin	smollm2	smollm2	quantized_int8	quantized_int4	false
smollm2-int8	smollm2-prune-mag30	true	sibling	sibling	smollm2	smollm2	quantized_int8	pruned_magnitude	false
smollm2-int8	smollm2-sft-prune-struct10	true	sibling	cousin	smollm2	smollm2	quantized_int8	pruned_structured	false
smollm2-int8	smollm2-vocab-ext	true	sibling	sibling	smollm2	smollm2	quantized_int8	vocab_extended	false
smollm2-int8	smollm2-vocab-ext-trained	true	sibling	sibling	smollm2	smollm2	quantized_int8	vocab_extended_trained	false
smollm2-int8	smollm2-merge-slerp	true	sibling	cousin	smollm2	smollm2	quantized_int8	merge_slerp	false
smollm2-int8	smollm2-merge-ties2	true	sibling	cousin	smollm2	smollm2	quantized_int8	merge_ties	false
smollm2-int8	smollm2-merge-ties3	true	sibling	cousin	smollm2	smollm2	quantized_int8	merge_ties	false
smollm2-int8	smollm2-merge-dare2	true	sibling	cousin	smollm2	smollm2	quantized_int8	merge_dare	false
smollm2-int8	smollm2-distil-half	true	sibling	sibling	smollm2	smollm2	quantized_int8	distilled	true
smollm2-sft-int4	smollm2-prune-mag30	true	sibling	cousin	smollm2	smollm2	quantized_int4	pruned_magnitude	false
smollm2-sft-int4	smollm2-sft-prune-struct10	true	sibling	sibling	smollm2	smollm2	quantized_int4	pruned_structured	false
smollm2-sft-int4	smollm2-vocab-ext	true	sibling	cousin	smollm2	smollm2	quantized_int4	vocab_extended	false
smollm2-sft-int4	smollm2-vocab-ext-trained	true	sibling	cousin	smollm2	smollm2	quantized_int4	vocab_extended_trained	false
smollm2-sft-int4	smollm2-merge-slerp	true	sibling	sibling	smollm2	smollm2	quantized_int4	merge_slerp	false
smollm2-sft-int4	smollm2-merge-ties2	true	sibling	sibling	smollm2	smollm2	quantized_int4	merge_ties	false
smollm2-sft-int4	smollm2-merge-ties3	true	sibling	sibling	smollm2	smollm2	quantized_int4	merge_ties	false
smollm2-sft-int4	smollm2-merge-dare2	true	sibling	sibling	smollm2	smollm2	quantized_int4	merge_dare	false
smollm2-sft-int4	smollm2-distil-half	true	sibling	cousin	smollm2	smollm2	quantized_int4	distilled	true
smollm2-prune-mag30	smollm2-sft-prune-struct10	true	sibling	cousin	smollm2	smollm2	pruned_magnitude	pruned_structured	false
smollm2-prune-mag30	smollm2-vocab-ext	true	sibling	sibling	smollm2	smollm2	pruned_magnitude	vocab_extended	false
smollm2-prune-mag30	smollm2-vocab-ext-trained	true	sibling	sibling	smollm2	smollm2	pruned_magnitude	vocab_extended_trained	false
smollm2-prune-mag30	smollm2-merge-slerp	true	sibling	cousin	smollm2	smollm2	pruned_magnitude	merge_slerp	false
smollm2-prune-mag30	smollm2-merge-ties2	true	sibling	cousin	smollm2	smollm2	pruned_magnitude	merge_ties	false
smollm2-prune-mag30	smollm2-merge-ties3	true	sibling	cousin	smollm2	smollm2	pruned_magnitude	merge_ties	false
smollm2-prune-mag30	smollm2-merge-dare2	true	sibling	cousin	smollm2	smollm2	pruned_magnitude	merge_dare	false
smollm2-prune-mag30	smollm2-distil-half	true	sibling	sibling	smollm2	smollm2	pruned_magnitude	distilled	true
smollm2-sft-prune-struct10	smollm2-vocab-ext	true	sibling	cousin	smollm2	smollm2	pruned_structured	vocab_extended	false
smollm2-sft-prune-struct10	smollm2-vocab-ext-trained	true	sibling	cousin	smollm2	smollm2	pruned_structured	vocab_extended_trained	false
smollm2-sft-prune-struct10	smollm2-merge-slerp	true	sibling	sibling	smollm2	smollm2	pruned_structured	merge_slerp	false
smollm2-sft-prune-struct10	smollm2-merge-ties2	true	sibling	sibling	smollm2	smollm2	pruned_structured	merge_ties	false
smollm2-sft-prune-struct10	smollm2-merge-ties3	true	sibling	sibling	smollm2	smollm2	pruned_structured	merge_ties	false
smollm2-sft-prune-struct10	smollm2-merge-dare2	true	sibling	sibling	smollm2	smollm2	pruned_structured	merge_dare	false
smollm2-sft-prune-struct10	smollm2-distil-half	true	sibling	cousin	smollm2	smollm2	pruned_structured	distilled	true
smollm2-vocab-ext	smollm2-vocab-ext-trained	true	sibling	sibling	smollm2	smollm2	vocab_extended	vocab_extended_trained	false
smollm2-vocab-ext	smollm2-merge-slerp	true	sibling	cousin	smollm2	smollm2	vocab_extended	merge_slerp	false
smollm2-vocab-ext	smollm2-merge-ties2	true	sibling	cousin	smollm2	smollm2	vocab_extended	merge_ties	false
smollm2-vocab-ext	smollm2-merge-ties3	true	sibling	cousin	smollm2	smollm2	vocab_extended	merge_ties	false
smollm2-vocab-ext	smollm2-merge-dare2	true	sibling	cousin	smollm2	smollm2	vocab_extended	merge_dare	false
smollm2-vocab-ext	smollm2-distil-half	true	sibling	sibling	smollm2	smollm2	vocab_extended	distilled	true
smollm2-vocab-ext-trained	smollm2-merge-slerp	true	sibling	cousin	smollm2	smollm2	vocab_extended_trained	merge_slerp	false
smollm2-vocab-ext-trained	smollm2-merge-ties2	true	sibling	cousin	smollm2	smollm2	vocab_extended_trained	merge_ties	false
smollm2-vocab-ext-trained	smollm2-merge-ties3	true	sibling	cousin	smollm2	smollm2	vocab_extended_trained	merge_ties	false
smollm2-vocab-ext-trained	smollm2-merge-dare2	true	sibling	cousin	smollm2	smollm2	vocab_extended_trained	merge_dare	false
smollm2-vocab-ext-trained	smollm2-distil-half	true	sibling	sibling	smollm2	smollm2	vocab_extended_trained	distilled	true
smollm2-merge-slerp	smollm2-merge-ties2	true	sibling	sibling	smollm2	smollm2	merge_slerp	merge_ties	false
smollm2-merge-slerp	smollm2-merge-ties3	true	sibling	sibling	smollm2	smollm2	merge_slerp	merge_ties	false

End of preview. Expand in Data Studio

Stemma lineage benchmark (NagaYu/stemma-bench)

Labelled ordered pairs over a generated model-lineage DAG. Every edge was produced by really running the operation on real safetensors weights (short fine-tunes, a hand-written LoRA, INT8/INT4 fake-quantisation round trips, magnitude and structured pruning, vocabulary extension, SLERP/TIES/DARE merges with recorded mixing ratios, and distillation), so the labels are recipes, not annotations.

models: 20
edges: 21
pairs: ancestral 28, sibling 80, unrelated 82
relations: continued_pretrain, distilled, lora_merged, merge_dare, merge_slerp, merge_ties, pruned_magnitude, pruned_structured, quantized_int4, quantized_int8, sft, vocab_extended, vocab_extended_trained

Columns

a, b (model ids), related (bool), direction (a->b | sibling | none), relation, plus the per-side family/op of each model.

Caveats

Distillation edges are marked weak_weight_lineage=true (distilgpt2, smollm2-distil-half). Weight-level lineage across a distillation edge is expected to be weak; it is included so it can be scored honestly rather than quietly excluded.
Direction accuracy must be reported per relation type. Scar-free relations (SFT, LoRA, continued pretraining) are much harder than lossy ones (quantisation, pruning, vocabulary extension).
Unrelated pairs include same-architecture / different-initialisation models. Those are the pairs a false-positive rate should be measured on.
Builds made with --skip-train carry training: synthetic and must not be reported as fine-tuning measurements.

What these labels can and cannot support

Measured on this dataset after it was first published:

Direction is near-deterministic only for lossy operations. Quantisation, pruning and vocabulary extension are recovered at 100%; scar-free sft / lora / continued_pretrain edges abstain (mean |llr| ~0.02). Scoring direction as one aggregate over all relations hides this, which is why the harness reports it per relation.
Outgroup rooting is invalid for the merge rows. Rooting assumes descendants drift monotonically away from the root. Merging is a contraction toward the centroid: smollm2-merge-ties2 = 0.6*sft + 0.4*cpt partly cancels two perturbations of the root and lands closer to the root than either parent (root->sft 0.000820, root->cpt 0.001610, root->merge-ties2 0.000678). Every correctly chosen sibling outgroup then pushes the answer the wrong way. Direction for a merged model has to come from the decomposition, not from distance geometry.
Cross-architecture distillation rows are a known-weak case, not a bug. They are included and scored honestly rather than excluded.

Current reference results on this dataset (python benchmarks/run.py): relatedness AUC 0.994 with 0.000 false-positive rate on the 25 same-architecture / different-seed controls; merge parent-set precision 1.000, recall 0.792, F1 0.867, mixing MAE 0.070 (DARE MAE 0.0004).

Derivations: https://github.com/NagaYu/stemma/blob/main/docs/FINDINGS.md

These labels describe generated derivations, not claims about anyone's published models. Stemma reports statistical evidence with a confidence and never a determination of infringement.

Downloads last month: -

Stemma lineage benchmark (NagaYu/stemma-bench)

Columns

Caveats

What these labels can and cannot support

Space using NagaYu/stemma-bench 1