Dataset Viewer
Auto-converted to Parquet Duplicate
model_a
stringlengths
13
43
model_b
stringlengths
13
40
p_derived
float64
0.13
1
⌀
sigma_r
float64
0.46
1
⌀
vector_cos
float64
0.98
1
⌀
pcs_cos
float64
0.41
1
⌀
spectra_r
float64
0.99
1
⌀
detected
bool
2 classes
documented_kind
stringclasses
6 values
child
stringclasses
17 values
merge_alpha
float64
0.29
1
⌀
codellama/CodeLlama-7b-hf
meta-llama/Llama-2-7b-hf
0.9832
0.9494
0.9983
0.525
0.9908
true
continued-pretrain
codellama/CodeLlama-7b-hf
null
codellama/CodeLlama-7b-hf
NousResearch/Nous-Hermes-llama-2-7b
0.9832
0.9494
0.9983
0.5249
0.9908
true
null
null
null
cognitivecomputations/dolphin-2.9-llama3-8b
deepseek-ai/DeepSeek-R1-Distill-Llama-8B
0.9976
0.9362
0.9998
0.9741
0.9981
true
null
null
null
cognitivecomputations/dolphin-2.9-llama3-8b
meta-llama/Llama-3.1-8B
0.9977
0.94
0.9999
0.9823
0.9984
true
null
null
null
cognitivecomputations/dolphin-2.9-llama3-8b
meta-llama/Llama-3.1-8B-Instruct
0.9977
0.9397
0.9999
0.9819
0.9984
true
null
null
null
cognitivecomputations/dolphin-2.9-llama3-8b
meta-llama/Meta-Llama-3-8B
0.998
0.9477
1
0.9963
0.9997
true
fine-tune
cognitivecomputations/dolphin-2.9-llama3-8b
null
cognitivecomputations/dolphin-2.9-llama3-8b
meta-llama/Meta-Llama-3-8B-Instruct
0.998
0.9477
1
0.9957
0.9997
true
null
null
null
DeepHat/DeepHat-V1-7B
deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
0.9704
0.941
0.9859
0.4096
0.9915
true
null
null
null
DeepHat/DeepHat-V1-7B
Qwen/Qwen2-7B
0.9707
0.9466
0.9827
0.4062
0.9907
true
null
null
null
DeepHat/DeepHat-V1-7B
Qwen/Qwen2.5-7B
0.9795
0.9536
0.9844
0.4786
0.9934
true
null
null
null
DeepHat/DeepHat-V1-7B
Qwen/Qwen2.5-7B-Instruct
0.9795
0.9535
0.9844
0.4786
0.9934
true
null
null
null
DeepHat/DeepHat-V1-7B
Qwen/Qwen2.5-Coder-7B
0.9985
1
1
0.9996
1
true
null
null
null
DeepHat/DeepHat-V1-7B
Qwen/Qwen2.5-Coder-7B-Instruct
0.9985
1
1
0.9979
0.9997
true
null
null
null
deepseek-ai/DeepSeek-R1-Distill-Llama-8B
meta-llama/Llama-3.1-8B
0.9984
0.9987
1
0.9909
0.9993
true
null
null
null
deepseek-ai/DeepSeek-R1-Distill-Llama-8B
meta-llama/Llama-3.1-8B-Instruct
0.9984
0.9989
1
0.9905
0.9993
true
null
null
null
deepseek-ai/DeepSeek-R1-Distill-Llama-8B
meta-llama/Meta-Llama-3-8B
0.9982
0.9887
0.9998
0.9771
0.998
true
null
null
null
deepseek-ai/DeepSeek-R1-Distill-Llama-8B
meta-llama/Meta-Llama-3-8B-Instruct
0.9982
0.9888
0.9998
0.9767
0.998
true
null
null
null
deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
Qwen/Qwen2-7B
0.9773
0.9273
0.9934
0.4845
0.9907
true
null
null
null
deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
Qwen/Qwen2.5-7B
0.9849
0.9391
0.9958
0.5618
0.9922
true
null
null
null
deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
Qwen/Qwen2.5-7B-Instruct
0.9849
0.9391
0.9958
0.5618
0.9922
true
null
null
null
deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
Qwen/Qwen2.5-Coder-7B
0.9704
0.941
0.9859
0.4095
0.9915
true
null
null
null
deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
Qwen/Qwen2.5-Coder-7B-Instruct
0.9703
0.9406
0.9859
0.4097
0.9917
true
null
null
null
google/gemma-2-9b
google/gemma-2-9b-it
0.9983
0.999
1
0.9748
0.9987
true
official-instruct
google/gemma-2-9b-it
null
google/gemma-2b
google/gemma-2b-it
0.9426
0.4562
0.9888
0.8475
0.9922
true
official-instruct
google/gemma-2b-it
null
HuggingFaceH4/zephyr-7b-beta
mistralai/Mistral-7B-Instruct-v0.3
0.9984
0.9966
1
0.9835
0.9983
true
null
null
null
HuggingFaceH4/zephyr-7b-beta
mistralai/Mistral-7B-v0.1
0.9985
1
1
0.9993
0.9999
true
fine-tune
HuggingFaceH4/zephyr-7b-beta
null
HuggingFaceH4/zephyr-7b-beta
mistralai/Mistral-7B-v0.3
0.9984
0.9967
1
0.9837
0.9984
true
null
null
null
HuggingFaceH4/zephyr-7b-beta
mlabonne/AlphaMonarch-7B
0.9984
0.9961
1
0.9967
0.9997
true
null
null
null
HuggingFaceH4/zephyr-7b-beta
mlabonne/Monarch-7B
0.9984
0.9961
1
0.9967
0.9997
true
null
null
null
HuggingFaceH4/zephyr-7b-beta
mlabonne/NeuBeagle-7B
0.9985
0.9972
1
0.9972
0.9997
true
null
null
null
HuggingFaceH4/zephyr-7b-beta
mlabonne/NeuralHermes-2.5-Mistral-7B
0.9985
1
1
0.999
0.9999
true
null
null
null
HuggingFaceH4/zephyr-7b-beta
mlabonne/NeuralOmniBeagle-7B
0.9984
0.9954
1
0.9967
0.9997
true
null
null
null
HuggingFaceH4/zephyr-7b-beta
mlabonne/NeuralPipe-7B-slerp
0.9985
0.9999
1
0.9986
0.9998
true
null
null
null
HuggingFaceH4/zephyr-7b-beta
mlabonne/OmniTruthyBeagle-7B-v0
0.9984
0.9954
1
0.9968
0.9998
true
null
null
null
HuggingFaceH4/zephyr-7b-beta
OpenPipe/mistral-ft-optimized-1218
0.9985
0.9998
1
0.9981
0.9998
true
null
null
null
HuggingFaceH4/zephyr-7b-beta
teknium/OpenHermes-2.5-Mistral-7B
0.9985
1
1
0.999
0.9999
true
null
null
null
meta-llama/Llama-2-7b-hf
NousResearch/Nous-Hermes-llama-2-7b
0.9985
1
1
0.9998
1
true
fine-tune
NousResearch/Nous-Hermes-llama-2-7b
null
meta-llama/Llama-3.1-8B
meta-llama/Llama-3.1-8B-Instruct
0.9985
1
1
0.9991
0.9999
true
null
null
null
meta-llama/Llama-3.1-8B
meta-llama/Meta-Llama-3-8B
0.9983
0.9934
0.9999
0.9859
0.9984
true
null
null
null
meta-llama/Llama-3.1-8B
meta-llama/Meta-Llama-3-8B-Instruct
0.9983
0.9935
0.9999
0.9852
0.9983
true
null
null
null
meta-llama/Llama-3.1-8B-Instruct
meta-llama/Meta-Llama-3-8B
0.9983
0.9932
0.9999
0.9851
0.9984
true
null
null
null
meta-llama/Llama-3.1-8B-Instruct
meta-llama/Meta-Llama-3-8B-Instruct
0.9983
0.9932
0.9999
0.9852
0.9984
true
null
null
null
meta-llama/Meta-Llama-3-8B
meta-llama/Meta-Llama-3-8B-Instruct
0.9985
1
1
0.9992
0.9999
true
official-instruct
meta-llama/Meta-Llama-3-8B-Instruct
null
mistralai/Mistral-7B-Instruct-v0.3
mistralai/Mistral-7B-v0.1
0.9984
0.9964
1
0.9841
0.9983
true
null
null
null
mistralai/Mistral-7B-Instruct-v0.3
mistralai/Mistral-7B-v0.3
0.9985
1
1
0.9998
1
true
official-instruct
mistralai/Mistral-7B-Instruct-v0.3
null
mistralai/Mistral-7B-Instruct-v0.3
mlabonne/AlphaMonarch-7B
0.9983
0.9887
1
0.9826
0.9983
true
null
null
null
mistralai/Mistral-7B-Instruct-v0.3
mlabonne/Monarch-7B
0.9983
0.9887
1
0.9826
0.9983
true
null
null
null
mistralai/Mistral-7B-Instruct-v0.3
mlabonne/NeuBeagle-7B
0.9983
0.9907
1
0.9833
0.9984
true
null
null
null
mistralai/Mistral-7B-Instruct-v0.3
mlabonne/NeuralHermes-2.5-Mistral-7B
0.9984
0.9964
1
0.9839
0.9983
true
null
null
null
mistralai/Mistral-7B-Instruct-v0.3
mlabonne/NeuralOmniBeagle-7B
0.9983
0.9875
1
0.9825
0.9983
true
null
null
null
mistralai/Mistral-7B-Instruct-v0.3
mlabonne/NeuralPipe-7B-slerp
0.9984
0.9968
1
0.9835
0.9983
true
null
null
null
mistralai/Mistral-7B-Instruct-v0.3
mlabonne/OmniTruthyBeagle-7B-v0
0.9983
0.9876
1
0.9826
0.9983
true
null
null
null
mistralai/Mistral-7B-Instruct-v0.3
OpenPipe/mistral-ft-optimized-1218
0.9983
0.9968
1
0.983
0.9983
true
null
null
null
mistralai/Mistral-7B-Instruct-v0.3
teknium/OpenHermes-2.5-Mistral-7B
0.9984
0.9964
1
0.9839
0.9983
true
null
null
null
mistralai/Mistral-7B-v0.1
mistralai/Mistral-7B-v0.3
0.9984
0.9965
1
0.9843
0.9984
true
null
null
null
mistralai/Mistral-7B-v0.1
mlabonne/AlphaMonarch-7B
0.9984
0.9963
1
0.9973
0.9998
true
merge-chain
mlabonne/AlphaMonarch-7B
null
mistralai/Mistral-7B-v0.1
mlabonne/Monarch-7B
0.9984
0.9963
1
0.9973
0.9998
true
null
null
null
mistralai/Mistral-7B-v0.1
mlabonne/NeuBeagle-7B
0.9985
0.9973
1
0.9979
0.9998
true
null
null
null
mistralai/Mistral-7B-v0.1
mlabonne/NeuralHermes-2.5-Mistral-7B
0.9985
1
1
0.9997
1
true
null
null
null
mistralai/Mistral-7B-v0.1
mlabonne/NeuralOmniBeagle-7B
0.9984
0.9956
1
0.9974
0.9998
true
null
null
null
mistralai/Mistral-7B-v0.1
mlabonne/NeuralPipe-7B-slerp
0.9985
0.9999
1
0.9993
0.9999
true
null
null
null
mistralai/Mistral-7B-v0.1
mlabonne/OmniTruthyBeagle-7B-v0
0.9984
0.9956
1
0.9975
0.9998
true
null
null
null
mistralai/Mistral-7B-v0.1
OpenPipe/mistral-ft-optimized-1218
0.9985
0.9998
1
0.9988
0.9998
true
null
null
null
mistralai/Mistral-7B-v0.1
teknium/OpenHermes-2.5-Mistral-7B
0.9985
1
1
0.9997
1
true
fine-tune
teknium/OpenHermes-2.5-Mistral-7B
null
mistralai/Mistral-7B-v0.1
TheBloke/Mistral-7B-v0.1-GPTQ
0.1301
null
1
null
null
false
quantized-copy
TheBloke/Mistral-7B-v0.1-GPTQ
null
mistralai/Mistral-7B-v0.3
mlabonne/AlphaMonarch-7B
0.9983
0.9888
1
0.9827
0.9984
true
null
null
null
mistralai/Mistral-7B-v0.3
mlabonne/Monarch-7B
0.9983
0.9888
1
0.9827
0.9984
true
null
null
null
mistralai/Mistral-7B-v0.3
mlabonne/NeuBeagle-7B
0.9983
0.9908
1
0.9834
0.9984
true
null
null
null
mistralai/Mistral-7B-v0.3
mlabonne/NeuralHermes-2.5-Mistral-7B
0.9984
0.9964
1
0.984
0.9984
true
null
null
null
mistralai/Mistral-7B-v0.3
mlabonne/NeuralOmniBeagle-7B
0.9983
0.9876
1
0.9826
0.9984
true
null
null
null
mistralai/Mistral-7B-v0.3
mlabonne/NeuralPipe-7B-slerp
0.9984
0.9968
1
0.9836
0.9983
true
null
null
null
mistralai/Mistral-7B-v0.3
mlabonne/OmniTruthyBeagle-7B-v0
0.9983
0.9876
1
0.9827
0.9984
true
null
null
null
mistralai/Mistral-7B-v0.3
OpenPipe/mistral-ft-optimized-1218
0.9983
0.9968
1
0.9831
0.9983
true
null
null
null
mistralai/Mistral-7B-v0.3
teknium/OpenHermes-2.5-Mistral-7B
0.9984
0.9964
1
0.984
0.9984
true
null
null
null
mlabonne/AlphaMonarch-7B
mlabonne/Monarch-7B
0.9985
1
1
1
1
true
null
mlabonne/AlphaMonarch-7B
1
mlabonne/AlphaMonarch-7B
mlabonne/NeuBeagle-7B
0.9985
0.9996
1
0.999
1
true
null
null
null
mlabonne/AlphaMonarch-7B
mlabonne/NeuralHermes-2.5-Mistral-7B
0.9984
0.9963
1
0.9973
0.9998
true
null
null
null
mlabonne/AlphaMonarch-7B
mlabonne/NeuralOmniBeagle-7B
0.9985
0.9999
1
0.9993
1
true
null
null
null
mlabonne/AlphaMonarch-7B
mlabonne/NeuralPipe-7B-slerp
0.9985
0.9962
1
0.9982
0.9999
true
null
null
null
mlabonne/AlphaMonarch-7B
mlabonne/OmniTruthyBeagle-7B-v0
0.9985
0.9999
1
0.9994
1
true
null
null
null
mlabonne/AlphaMonarch-7B
OpenPipe/mistral-ft-optimized-1218
0.9985
0.9966
1
0.9982
0.9999
true
null
null
null
mlabonne/AlphaMonarch-7B
teknium/OpenHermes-2.5-Mistral-7B
0.9984
0.9963
1
0.9973
0.9998
true
null
null
null
mlabonne/Monarch-7B
mlabonne/NeuBeagle-7B
0.9985
0.9996
1
0.999
1
true
null
mlabonne/Monarch-7B
0.347
mlabonne/Monarch-7B
mlabonne/NeuralHermes-2.5-Mistral-7B
0.9984
0.9963
1
0.9973
0.9998
true
null
null
null
mlabonne/Monarch-7B
mlabonne/NeuralOmniBeagle-7B
0.9985
0.9999
1
0.9993
1
true
null
mlabonne/Monarch-7B
0.291
mlabonne/Monarch-7B
mlabonne/NeuralPipe-7B-slerp
0.9985
0.9962
1
0.9982
0.9999
true
null
null
null
mlabonne/Monarch-7B
mlabonne/OmniTruthyBeagle-7B-v0
0.9985
0.9999
1
0.9994
1
true
null
mlabonne/Monarch-7B
0.371
mlabonne/Monarch-7B
OpenPipe/mistral-ft-optimized-1218
0.9985
0.9966
1
0.9982
0.9999
true
null
null
null
mlabonne/Monarch-7B
teknium/OpenHermes-2.5-Mistral-7B
0.9984
0.9963
1
0.9973
0.9998
true
null
null
null
mlabonne/NeuBeagle-7B
mlabonne/NeuralHermes-2.5-Mistral-7B
0.9985
0.9972
1
0.9978
0.9998
true
null
null
null
mlabonne/NeuBeagle-7B
mlabonne/NeuralOmniBeagle-7B
0.9985
0.9991
1
0.9988
1
true
null
null
null
mlabonne/NeuBeagle-7B
mlabonne/NeuralPipe-7B-slerp
0.9985
0.9973
1
0.9987
0.9999
true
null
null
null
mlabonne/NeuBeagle-7B
mlabonne/OmniTruthyBeagle-7B-v0
0.9985
0.9991
1
0.9988
1
true
null
null
null
mlabonne/NeuBeagle-7B
OpenPipe/mistral-ft-optimized-1218
0.9985
0.9977
1
0.9987
0.9999
true
null
null
null
mlabonne/NeuBeagle-7B
teknium/OpenHermes-2.5-Mistral-7B
0.9985
0.9972
1
0.9978
0.9998
true
null
null
null
mlabonne/NeuralHermes-2.5-Mistral-7B
mlabonne/NeuralOmniBeagle-7B
0.9984
0.9955
1
0.9974
0.9998
true
null
null
null
mlabonne/NeuralHermes-2.5-Mistral-7B
mlabonne/NeuralPipe-7B-slerp
0.9985
0.9999
1
0.9994
0.9999
true
null
mlabonne/NeuralPipe-7B-slerp
0.356
mlabonne/NeuralHermes-2.5-Mistral-7B
mlabonne/OmniTruthyBeagle-7B-v0
0.9984
0.9955
1
0.9974
0.9998
true
null
null
null
mlabonne/NeuralHermes-2.5-Mistral-7B
OpenPipe/mistral-ft-optimized-1218
0.9985
0.9998
1
0.9988
0.9998
true
null
null
null
mlabonne/NeuralHermes-2.5-Mistral-7B
teknium/OpenHermes-2.5-Mistral-7B
0.9985
1
1
1
1
true
null
null
null
End of preview. Expand in Data Studio

modelDNA Atlas — fingerprint DB + inferred lineage graph

The data behind the modelDNA Atlas: weight-space fingerprints of 68 real Hugging Face models and the family graph reconstructed from those fingerprints alone — no model card metadata, no self-reported base_model fields, just the weights.

Built with modelDNA v0.1.0 · generated 2026-07-10T20:21:11+00:00 · refreshed automatically. Scan your own model against this DB in the live Space.

What's here

file rows what
models.jsonl 68 one row per fingerprinted model (architecture, family, fingerprint metadata)
edges.jsonl 123 the inferred graph: 121 detected edges (calibrated P(derived) ≥ 0.9) plus 15 externally documented pairs, including two decomposed merges with fitted mixture weights
pairs.jsonl 501 every depth-compatible pairwise comparison — the full evidence matrix, not just the hits
fingerprints/*.json.gz 68 the raw fingerprints (σ-curves, norm vectors, seeded PCS samples, spectra; ~1.5 MB each)
modeldna-refdb.tar.gz 37 models the bases-only reference DB (57.2 MB) that modeldna db pull installs

Load it

from datasets import load_dataset

edges = load_dataset("AwaisAdilKhokhar/modeldna-atlas", "edges", split="train")
models = load_dataset("AwaisAdilKhokhar/modeldna-atlas", "models", split="train")

Use the reference DB directly with the CLI:

pip install modeldna
modeldna db pull --url https://huggingface.co/datasets/AwaisAdilKhokhar/modeldna-atlas/resolve/main/modeldna-refdb.tar.gz
modeldna scan some-org/suspicious-model

Columns

models — model_id, family, group (coarse visual grouping, never itself a lineage claim), architecture fields (n_layers, hidden_size, n_heads, n_kv_heads, intermediate_size, vocab_size, n_params, torch_dtype), revision, fingerprint_mode/fingerprint_created_at/ fingerprint_file, in_reference_db (true for the curated base-model set), and — where the model is a LineageBench suspect — documented_parent and lineagebench_verdict.

pairs / edges (same schema; edges is the filtered graph) — model_a, model_b, p_derived (calibrated probability that the pair is weight-derived), the four raw signals (sigma_r attention σ-curve correlation, vector_cos norm/bias cosine, pcs_cos seeded parameter-sample cosine, spectra_r SVD spectrum correlation), detected (p_derived ≥ 0.9), documented_kind (fine-tune, official-instruct, merge-chain, … — from the publishing org's own documentation, with citations in benchmarks/lineagebench_pairs.json), child (direction, only where documented or decomposed), and merge_alpha (fitted mixture weight from modeldna decompose, validated against published mergekit configs).

Read the graph honestly

  • Edges are calibrated probabilities of weight derivation, not accusations. A detected edge means the weights are statistically consistent with derivation; the evidence behind it is in the four signal columns.
  • Absence of an edge is not evidence of independence. Pairs with different layer counts are never compared (depth-changing derivations such as pruning or stacking are out of scope), and quantized/packed weights abstain by design. edges.jsonl carries those documented-but-abstained pairs with p_derived: null.
  • Direction is hard from weights alone. child is set only where external documentation or a merge decomposition establishes it; detected edges without it are undirected.
  • Coverage is curated, not exhaustive. 68 models is a validated seed, not the Hub. The benchmark behind the method (AUROC 1.0, 0 false positives at p ≥ 0.9 on 13 positives / 107 hard negatives) is reproducible from the repo.

Cite

@software{modeldna,
  author = {Awais Bin Adil, Muhammad},
  title = {modelDNA: weight-based lineage verification for open models},
  url = {https://github.com/AwaisAdilKhokhar/modelDNA},
  year = {2026}
}
Downloads last month
93