LineageBench

A small, honestly-scoped benchmark for weight-based lineage verification: given a suspect open-weight model and a set of candidate parents, decide which foundation base it actually descends from — from the weights alone.

The hard part of a lineage benchmark is the ground truth. It cannot come from a model's own base_model tag, because auditing that tag is the entire point of the exercise. LineageBench instead labels every pair from the publishing organization's own documentation — technical reports, official release notes, model cards written by the org that trained the parent — and ships a citation for each label.

• 15 suspect models across 5 foundation families
• 8 candidate parents (curated foundation bases)
• Reference implementation: modelDNA v0.1.0
• Headline: AUROC 1.0 · TPR@FPR1% 1.0 · 0 false positives at the reporting threshold, 13/13 top-1 parent attribution

Generated 2026-07-11T06:11:25+00:00.

Why this exists

Open-weight model lineage is mostly undocumented or unverifiable: a large fraction of Hub models carry missing, self-reported, or contradicted parentage. "Which base is this really derived from?" recurs in license-compliance disputes, merge archaeology, and safety provenance — and the honest answer has to be reconstructed from weights, not taken on trust. LineageBench is a named, cited, reproducible yardstick for that task, so different methods can be compared on the same ground truth instead of on anecdotes.

What's here

file	rows	what
ground_truth.jsonl	15	one row per suspect: suspect, parent (true base), kind, in_positive_pool, and a citation to org documentation
reference_results.jsonl	15	modelDNA's output per suspect: calibrated score_vs_true_parent, verdict, verdict_best_candidate, top1_correct
parents.json	8	the candidate foundation parents every suspect is judged against
metrics.json	—	the headline ship-gate metrics for the reference implementation

The suspects

The pairs cover the derivation regimes a scanner meets in the wild, on purpose:

• fine-tunes — community SFT/DPO derivatives (Zephyr-7B-β, OpenHermes-2.5, Nous-Hermes, Dolphin-2.9)
• continued-pretrain — CodeLlama-7B, Qwen2.5-Coder-7B-Instruct
• official-instruct — the orgs' own instruct releases (Llama-3, Qwen2.5, Mistral, Gemma)
• merge-chain — AlphaMonarch-7B, a DARE-TIES merge whose every ancestor traces to one Mistral base
• depth-upscale — SOLAR-10.7B (a limitation case, see below)
• quantized-copy — a GPTQ int4 repack (a limitation case)

Two cases (in_positive_pool: false) are out-of-scope by design and scored separately against their documented expected behavior, which is abstention: depth-changing derivations (SOLAR's 32→48 layer up-scale) and packed-int4 quantization (GPTQ) both fall outside the shape-matched, safetensors-only comparison the reference tool performs today. A method that scored them as confident positives would be wrong; abstaining is the correct answer.

The scoring protocol

Each suspect is judged by the actual verdict engine against all 8 candidate parents — the true parent must win, not merely score well in isolation. The positive pool is the 13 in-scope suspects. The negative pool is 107 hard negatives: every positive suspect scored against each cross-family wrong parent, plus all cross-family parent-versus-parent pairs.

Same-family wrong parents (e.g. Zephyr against Mistral-v0.3 rather than v0.1) are deliberately excluded from the negative pool — they genuinely share lineage, and scoring them as negatives would punish a method for detecting something true.

Reference results (modelDNA v0.1.0)

metric	value
AUROC (13 positives vs 107 hard negatives)	1.0
TPR at 1% FPR	1.0
false positives at the p ≥ 0.9 reporting threshold	0 / 107
weakest positive	0.9426
strongest negative	0.6596
top-1 parent attribution	13/13

Clean separation: the entire abstention band (0.50–0.90) is empty on this data. The two limitation cases behaved exactly as documented — SOLAR produced an honest NO_MATCH on the layer-count mismatch; the GPTQ copy abstained on the packed tensors while still ranking the true parent first.

Reproduce it

Every label carries a citation and every fingerprint is cached, so the numbers regenerate offline in seconds — no Hub round-trip:

pip install modeldna
git clone https://github.com/AwaisAdilKhokhar/modelDNA
python modelDNA/benchmarks/real_lineagebench.py --no-fetch
# -> ship gates: PASS

Load it

from datasets import load_dataset

gt = load_dataset("AwaisAdilKhokhar/lineagebench", "ground_truth", split="train")
res = load_dataset("AwaisAdilKhokhar/lineagebench", "reference_results", split="train")

Scope, honestly

This is a deliberately small slice — 15 pairs, not the ≥300 the project's planning document sets as the eventual target. Every label is cited, but AUROC 1.0 should be read as "no errors at this scale", not as a claim that errors are impossible. Growing the benchmark is mechanical (the fingerprint pipeline is the expensive part, and it is fast); contributions of new org-documented, cited pairs are welcome via PR to the manifest.

Absence of a detected edge is not evidence of independence: distillation is invisible to every weight-space method by construction, and depth/quantization mismatches abstain rather than resolve. Read a verdict as statistical consistency with derivation, never as an accusation.

Cite

@misc{lineagebench,
  title  = {LineageBench: org-documented lineage ground truth for open-weight LLMs},
  author = {Awais Bin Adil, Muhammad and Aamir, Saad},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/AwaisAdilKhokhar/lineagebench}},
  note   = {Reference implementation: modelDNA, \url{https://github.com/AwaisAdilKhokhar/modelDNA}}
}