File size: 3,422 Bytes
76b6bfc
2ac142a
 
 
 
 
 
 
 
76b6bfc
2ac142a
a29c2a2
2ac142a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a29c2a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
---
license: apache-2.0
task_categories:
- other
tags:
- model-provenance
- lineage
- ai-bom
- model-merging
---

# 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:

1. **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.
2. **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.
3. **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.