marcjon/llmfp-fingerprints

llmfp Fingerprint Library

A library of pre-computed weight-extrema fingerprints for ~360 LLMs, generated by llmfp — a forensic analysis tool that detects AI model lineage by reading weight files (SafeTensors, GGUF, PyTorch, ONNX) without executing the model or requiring a GPU.

This dataset is the canonical store for the fingerprint .json files. Local copies in the llmfp repo are working copies, not committed to git (50–100 MB each). The registry metadata (parent/modification/notes for every entry) lives at index.yaml here and at fingerprints_index.yaml in the llmfp repo.

What's in a fingerprint

A fingerprint is a deterministic, GPU-free summary of a model's weight tensors. Each *.json file contains, per tensor:

• Extrema samples — positions and values of the largest-magnitude weights (the lineage-bearing signal; positions are preserved through fine-tuning but random for independently trained models)
• Grid samples — uniform position grid for cross-validation of similarity metrics
• Topographical samples — multi-scale "peak count" structure for shape analysis
• Per-tensor randomness statistics — kurtosis, Shannon entropy, Gini coefficient, sparsity at three thresholds, gzip compression ratio
• Tensor metadata — shape, dtype, parameter count, semantic category (attention / mlp / norm / embedding / lm_head)

Plus per-fingerprint metadata: source model identifier, sampling mode + sample sizes, model_config snapshot (architecture, vocab_size, etc.), total_parameters, tensor_count, and detected modifications (quantization config, sharding, etc.).

Sampling modes

mode	extrema	grid	topographical	use case
fast	512	–	–	exploratory checks; lower accuracy
balanced (default)	2048	2048	384 (16/64/192 scales × 128)	canonical mode for all entries here
full	8192	8192	384	research / ground-truth reference

All current 359 entries in this dataset are at balanced (2048 extrema samples), so cross-comparison is well-defined. See "Compatibility caveats" below.

Compatibility caveats

Cross-comparison requires matching sample sizes. If fingerprint A was generated with 2048 samples and fingerprint B with 3000 samples, the comparator's value-correlation metrics (cosine, mean-diff, etc.) become meaningless because the stored sample positions diverge — even bit-identical model weights can produce cosine ≈ 0 because the cross-position alignment is undefined.

As of llmfp v3.7.5, the comparator detects this and refuses with IncompatibleSampleSizeError (exit code 4). The --force-mismatched-samples CLI flag bypasses it for exploratory use.

The library was migrated to a uniform 2048-sample baseline in v3.7.5–v3.7.6. Two outliers remain (Qwen3-14B and Mistral-Nemo-Instruct-2407 at the pre-Dec-2025 3000-sample default — too large to resample on the v3.7.5 development box).

Naming convention

Files follow <org>_<model>.json, where the HF model identifier has / replaced with _:

• meta-llama_Llama-2-7b-hf.json ← meta-llama/Llama-2-7b-hf
• microsoft_phi-2.json ← microsoft/phi-2
• TheBloke_Llama-2-13B-GPTQ.json ← TheBloke/Llama-2-13B-GPTQ

To recover the HF model id from a filename, replace the first _ with /.

A few entries use variant disambiguators (<org>_<model>-<variant>.json) when multiple fingerprints share the same HF repo id (e.g., quantization variants).

Repository layout

marcjon/llmfp-fingerprints/
├── README.md                ← this file
├── index.yaml               ← per-entry metadata: parent, modification, notes, model_card url, ...
└── fingerprints/
    ├── meta-llama_Llama-2-7b-hf.json
    ├── meta-llama_Llama-2-7b-chat-hf.json
    └── ... (~360 entries total)

How to consume

With the llmfp library

# Pull a snapshot of this dataset to local disk
python scripts/hf_library_sync.py pull --repo marcjon/llmfp-fingerprints

# Now point llmfp at any fingerprint
llmfp compare fingerprints/meta-llama_Llama-2-7b-hf.json \
              fingerprints/meta-llama_Llama-2-7b-chat-hf.json

Without llmfp

Each .json file is a self-contained, human-readable record:

import json
fp = json.load(open("meta-llama_Llama-2-7b-hf.json"))
print(fp["tensor_count"], fp["total_parameters"])
print(fp["sampling_methods"], fp["sampling_mode"])
for tname, t in list(fp["tensors"].items())[:3]:
    extrema = t["fingerprints"]["extrema"]
    print(tname, "extrema samples:", len(extrema["samples"]))

The compressed variants the llmfp generator can emit (.json.zst ≈ 7× smaller, .parquet ≈ 10× smaller) are bit-identical on round-trip; llmfp loaders auto-detect by magic bytes. This dataset currently stores the plain .json form for transparency.

Coverage and quality

As of 2026-05-28:

• 359 fingerprints, 350 indexed entries in index.yaml (9 variant-disambiguated entries share an HF id)
• Sampling mode: 356/359 at balanced (2048 extrema + 2048 grid + 384 topographical); 2 outliers at the pre-Dec-2025 3000-sample default (Qwen3-14B, Mistral-Nemo-Instruct-2407 — too large for in-session resample); 1 at full mode (fdtn-ai/Foundation-Sec-8B-Instruct at 8192 samples)
• Randomness statistics (kurtosis, entropy, Gini, sparsity at 3 thresholds, compression ratio): present in 354/359 (98.6%). 5 older fingerprints predate the compute_randomness_stats default-on switch
• model_config snapshot: present in 359/359 (100%)
• Embedding tensors captured in 343/359 (96%). The 16 absent are mostly encoder-decoder (T5 / FLAN-T5 / mT5 / Cohere Aya) and GGUF/ONNX variants whose tokenizer layer is named differently or moved out of the main weight list; the comparator handles their alignment separately
• Coverage spans LLaMA, Mistral, Qwen, Phi, Gemma, BERT, T5, ViT, CLIP, Whisper, Wav2Vec2, ResNet, EfficientNet, GPT-2 families, plus MoE families (DeepSeek-V2-Lite, OLMoE, LLaMA-MoE, Qwen1.5-MoE, Qwen3-Coder-30B) and quantization variants (GPTQ, AWQ, GGUF k-quant, FP8, ONNX)
• Cisco MPK benchmark roster: 136/137 models present at balanced=2048 (the Qwen3-0.6B base/instruct pair was resampled in v3.7.6 closing the previous 134/137 gap)

Update workflow

The local llmfp working copy and this HF dataset are kept in sync by scripts/hf_library_sync.py:

# Push the entire local library to HF (additive — won't reap retired entries)
python scripts/hf_library_sync.py push --repo marcjon/llmfp-fingerprints

# Push only the registry (fast metadata sync)
python scripts/hf_library_sync.py push-index --repo marcjon/llmfp-fingerprints

# Push only this README
python scripts/hf_library_sync.py push-readme --repo marcjon/llmfp-fingerprints

# Delete specific files from the remote
python scripts/hf_library_sync.py delete fingerprints/foo.json

# Pull a clean copy down
python scripts/hf_library_sync.py pull --repo marcjon/llmfp-fingerprints

Authentication is via the huggingface_hub SDK (hf auth login or HF_TOKEN env var). The dataset is private; access requires a token authorized by the owner.

Sister datasets / projects

• llmfp — the tool that produces and consumes these fingerprints
• marcjon/ckasketch-sketches — sister dataset for ckasketch representational-geometry sketches (different format, complementary use case)
• lineagebench — benchmark harness that consumes both datasets

License

Cisco Proprietary. See the LICENSE file in the llmfp repository.

The fingerprints themselves are derived static summaries of public model weights — they contain no model parameters or generative capacity. Each entry's model_card URL in index.yaml points at the upstream model's HF page and license; consult that for the upstream license of the model being fingerprinted.

Citation

If this dataset informs published work, please cite the llmfp repository and tag the dataset version (HF dataset revisions are immutable Git refs):

@misc{llmfp_fingerprints_2026,
  author = {Jones, Marc},
  title  = {llmfp Fingerprint Library},
  year   = {2026},
  publisher = {Hugging Face Hub},
  url    = {https://huggingface.co/datasets/marcjon/llmfp-fingerprints},
}