Find out what really changed when someone fine tuned, quantized, merged or edited a model.
Documentation | Quickstart | How it works | Layer by layer | CLI reference | Roadmap
What is this?
You start with a model, the base. Someone builds a new model from it, the derivative. That can be a fine tune, a 4 bit quantized build, a merge of two models, or a model with some behavior edited out.
The model card usually says something like "fine tuned for medical questions" or "4 bit, almost no quality loss". Modeldiffr checks that claim. It gives both models the exact same questions, compares the answers, tells you which differences are real and which are just noise, and writes a report that anyone can check later.
pip install "modeldiffr[hf]"
modeldiffr audit Qwen/Qwen2.5-0.5B-Instruct your-org/your-qwen-finetune
Status: v0.9 preview. All the features are built. Version 1.0 comes after the public study and the backdoor test results are published. See the roadmap.
Why you need it
Hugging Face has hundreds of thousands of derivative models. Picking one today means trusting a sentence on a model card. Running benchmarks yourself is not much better, for three reasons:
- Unfair setup. The two models often get different prompts, chat templates or settings. Then you are measuring the setup, not the model.
- No sense of noise. A score that moves from 79% to 75% on 24 questions could easily be luck. A plain benchmark number does not tell you.
- No way to check. Someone posts a table of results. You cannot tell which exact weights they ran, or whether the numbers were edited.
Modeldiffr fixes all three. Same inputs for both models, a confidence range on every number, and a report that is locked to the exact model files.
How it works, in plain words
Think of it as a careful experiment where the only thing that changes is the model.
- Take fingerprints. Modeldiffr hashes every weight tensor, the tokenizer and the chat template of both models. Now the report points to the exact files, not just a name on the Hub that someone could overwrite.
- Ask the same questions. Both models get byte for byte identical prompts. If their chat templates differ, both get the prompt written with one shared template.
- Score fairly. For multiple choice questions, Modeldiffr does not read the model's written answer. It asks the model how likely each option is and picks the most likely one. So a model that writes "Sure! The answer is B" is not punished for its style.
- Separate real change from luck. For every question it looks at the difference between the two models. Then it reshuffles those differences thousands of times (a bootstrap) to see how much the average could move by chance. If the range of likely values does not include zero, the change is real.
- Lock the report. Every question and every answer is hashed into one summary hash (a Merkle tree). If anyone edits a single number later,
modeldiffr verifywill catch it. - Look inside, if you want. Optional tools compare the models layer by layer, find which internal features are new or lost, test which of them actually cause the behavior change, and scan for hidden triggers.
Longer version: How it works.
Layer by layer comparison
Benchmarks tell you that something changed. The layer by layer tools tell you where inside the model it changed.
There are two views:
| View | Command | What it compares | Needs to run the model? |
|---|---|---|---|
| Weights | runs inside every audit, and fingerprint |
The stored numbers of every tensor, one by one | No |
| Hidden states | modeldiffr layers |
What each layer actually outputs when both models read the same text | Yes |
The weights view hashes each tensor in both models and lists which ones are different, added or removed. A LoRA fine tune that only touched attention shows up as "only attention tensors changed". A "near lossless" quant shows up as "every tensor changed". You learn this before any benchmark runs.
The hidden states view feeds the same text to both models and records what each layer outputs for every token. Then, for each layer, it measures how similar the two outputs are:
| Number | Plain meaning | Identical models give |
|---|---|---|
cka |
Do the two layers organize the information the same way overall? | 1 |
cosine |
Do the token vectors point in the same direction? | 1 |
relative_change |
How big is the difference, compared to the size of the base output? | 0 |
mean_shift |
Did the average output move? | 0 |
norm_ratio |
Did the outputs get bigger or smaller? | 1 |
modeldiffr layers Qwen/Qwen2.5-0.5B-Instruct your-org/your-qwen-finetune
layer CKA cosine rel change mean shift norm ratio
0 1.0000 1.0000 0.0000 0.0000 1.0000
...
14 0.9871 0.9802 0.1543 0.0611 0.9930
...
largest change at layer 14
Why this works. A transformer is a stack of layers. Each layer reads the output of the one before it and passes its own output on. If two models read the exact same tokens, any difference at layer N must come from weights at layer N or before it. So the first layer where the numbers jump is where the derivative starts thinking differently. Small fine tunes usually leave early layers almost untouched and show a clear jump somewhere in the middle.
Why it is needed. Two models can score the same on every benchmark and still be different inside. A quant might lose precision in one specific layer. An "abliterated" model might keep its benchmark scores but have its refusal behavior cut out at a few layers. The layer view finds those places, and tells the more expensive tools (crosscoders and causal patching) which layer to study.
Full explanation: Layer by layer comparison.
What it can answer
| Question | Commands | Read more |
|---|---|---|
| Which exact models and settings were compared, and can anyone check it? | verify sign anchor |
Fingerprints and proof |
| Did scores change on thirteen standard benchmarks, and is the change bigger than noise? | audit |
Statistics |
| Which layers changed, and by how much? | layers directions |
Layer by layer |
| Which internal features are new, lost or reshaped? | features |
Features |
| Which of those layers and features actually cause the change? | causal |
Causal tests |
| Is there a strange input where the derivative acts very differently? | scan zoo |
Anomaly scan |
Install
pip install "modeldiffr[hf]"
| Extra | Adds | Use it for |
|---|---|---|
hf |
PyTorch, Transformers | Running audits and the layer tools |
tasks |
Datasets | The lite, standard and full benchmark suites |
vllm, llamacpp |
vLLM, llama.cpp | Fast GPU generation, GGUF builds |
proof |
jsonschema | Full schema checks in verify |
sign, anchor |
cryptography, Sigstore, OpenTimestamps, web3 | Signing and public timestamps |
batch |
PyYAML | YAML batch files |
all |
Everything above except vLLM and llama.cpp |
The core package has no dependencies at all, so a CI job that only reads reports stays tiny. See Installation.
Quickstart
modeldiffr audit Qwen/Qwen2.5-0.5B-Instruct Qwen/Qwen2.5-0.5B --out reports/qwen
audit id mdr1:71e153dc730bb9ec20b7159dc8845afab76de27ddf94b99cada26c713aa3170d
weights: 290 of 290 shared tensors changed, 0 added, 0 removed
capability_accuracy base 0.7917 derivative 0.7500 delta -0.0417 95% CI [-0.1667, +0.0833] within noise
over_refusal_rate base 0.1000 derivative 0.0000 delta -0.1000 95% CI [-0.2500, +0.0000] within noise
output_divergence_kl base 0.0000 derivative 0.4120 delta +0.4120 95% CI [+0.3051, +0.5233] CHANGED
report: reports/qwen/report.html
json: reports/qwen/report.json
These numbers only show the format. Run the command to get real values.
How to read one line: delta is derivative minus base. 95% CI is the range the true difference most likely sits in. If that range includes zero, the result is within noise. If it does not, the result is CHANGED.
report.html is a single page you can open anywhere or attach to a pull request. report.json holds every question, setting and hash. Full walkthrough: Quickstart.
Common tasks
Run real benchmarks
pip install "modeldiffr[hf,tasks]"
modeldiffr audit BASE DERIVATIVE --suite standard --plan-only --estimate
modeldiffr audit BASE DERIVATIVE --suite standard
Covers MMLU, ARC Challenge, HellaSwag, TruthfulQA, GSM8K, MGSM, IFEval, XSTest, OR Bench, HumanEval, a long context test and output divergence. Every dataset is pinned to a fixed commit. --plan-only --estimate tells you how long it will take before you commit a GPU. Try it free on Colab.
Check a quantized build
modeldiffr audit Qwen/Qwen2.5-1.5B-Instruct \
Qwen/Qwen2.5-1.5B-Instruct-GGUF:qwen2.5-1.5b-instruct-q4_k_m.gguf \
--base-backend hf --suite lite
See Backends.
Check someone else's report
modeldiffr verify report.json
modeldiffr verify report.json --models
--models also downloads and hashes the real model files to confirm they match. See Verify a report.
Sign and publish
modeldiffr keygen --out keys
modeldiffr sign report.json --key keys/modeldiffr.key
modeldiffr hub push audit --repo your-org/your-model
modeldiffr registry build audits --out registry_site --key me=keys/modeldiffr.pub
See Sign and verify and Public registry.
Look inside the model
modeldiffr layers BASE DERIVATIVE
modeldiffr features BASE DERIVATIVE --layer 14 --out audit --report audit/report.json
modeldiffr causal BASE DERIVATIVE --layer 14 --features audit --tasks mmlu --out audit
modeldiffr scan BASE DERIVATIVE --out audit
Run them in this order. layers finds the layer, features finds what changed at that layer, causal tests whether it matters, scan looks for hidden triggers.
Block a bad checkpoint in CI
permissions:
id-token: write
contents: read
steps:
- uses: actions/checkout@v4
- uses: amareshhebbar/modeldiffr@v0.9.4
with:
base: org/base_model
derivative: ./checkpoints/candidate
suite: lite
The action runs the audit, fails the job if something changed for real, signs the report and uploads it. See CI gate.
Use it from Python
from modeldiffr import run_audit
from modeldiffr.backend import HFBackend
from modeldiffr.report import write_reports
from modeldiffr.suites import get_suite
base = HFBackend("Qwen/Qwen2.5-0.5B-Instruct")
derivative = HFBackend("your-org/your-qwen-finetune")
result = run_audit(base, derivative, get_suite("quick"))
write_reports(result, "audit")
print(result["verdict"])
See Python usage.
Roadmap
| Version | Shipped |
|---|---|
| 0.2 | Weight fingerprints, audit ids, verify |
| 0.3 | Thirteen benchmarks, four backends, corrected statistics |
| 0.4 | Signing, attestations, public timestamps, bills of materials, GitHub Action, Hub upload |
| 0.5 | Planted change controls, batch audits, a registry that verifies in the browser |
| 0.6 | Layer by layer statistics, direction comparison, crosscoder features |
| 0.7 | Patching and steering tests |
| 0.8 | Anomaly scan and a test zoo that measures its accuracy |
| 0.9 | Study tables, JSON Schemas, compatibility promise, governance, paper draft |
| 1.0 | After the public study and zoo results are published |
What is still open is listed in Roadmap: still pending.
Limitations
- The built in
quicksuite is a smoke test. Uselite,standardorfullfor a real decision. - Bigger models need a GPU. Both models are in memory at the same time.
- A real change tells you something moved, not why. Features and causal shares are strong hints, not final proof.
- The anomaly scan makes no claim about real models until its accuracy on a public zoo is published.
All of them: Limitations.
Help and contact
| Need | Where |
|---|---|
| Something is broken | Report a bug |
| A metric, suite or backend you need | Request a feature |
| Questions | GitHub Discussions |
| A security issue | Private advisory, see SECURITY.md |
| Audits of your models, collaboration, anything else | hebbar.gvamaresh@gmail.com |
Responsible use
Modeldiffr only measures. It never changes model weights, and its reports say what changed and by how much, not how to copy or undo a change. If you find a serious safety problem or hidden behavior in a public model, tell its publisher privately before you post about it. See Responsible use.
Contributing
Everything from bug reports to new benchmark suites is welcome. Start with CONTRIBUTING.md, GOVERNANCE.md and issues labelled good first issue.
Citation
@software{hebbar_modeldiffr_2026,
author = {Hebbar, Amaresh},
title = {Modeldiffr: auditing what changes between base and derivative language models},
year = {2026},
url = {https://github.com/amareshhebbar/modeldiffr},
version = {0.9.4},
doi = {10.5281/zenodo.23587022}
}
Acknowledgements
The feature comparison builds on crosscoder model diffing from Sparse Crosscoders for Cross Layer Features and Model Diffing (Lindsey et al., Anthropic, 2024) and the work that followed. The benchmark side learns from EleutherAI's lm evaluation harness and UK AISI's Inspect. The proof side follows RFC 8785 and RFC 6962 and works with Sigstore and in toto.
License
Apache 2.0. Copyright 2026 Amaresh Hebbar.
Metadata
Release files for modeldiffr 0.9.4
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| modeldiffr-0.9.4.tar.gz | 167.3 kB | With provenance | Details for modeldiffr-0.9.4.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| modeldiffr-0.9.4-py3-none-any.whl | Python 3 | none | any | Details for modeldiffr-0.9.4-py3-none-any.whl |
Total release size: 334.3 kB
Release files / modeldiffr-0.9.4.tar.gz
| Download URL | modeldiffr-0.9.4.tar.gz |
|---|---|
| Size | 167.3 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
a97e7e26442833707013c10266ee63f71ccd2266f177e32fed9a0a8a58a2bd54
|
|
BLAKE2b-256 checksum How to use checksums |
8faedb8ab9525ae6992dfcd85ce949e654eb596c522b01cc273c4924c69b00d1
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 14, 2026.
Transparency logRelease files / modeldiffr-0.9.4-py3-none-any.whl
| Download URL | modeldiffr-0.9.4-py3-none-any.whl |
|---|---|
| Size | 167.0 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
1afc26d6b1e9a5d70b6b346ee9b3fc41a4f257adff014bbc723dcc14c1f27882
|
|
BLAKE2b-256 checksum How to use checksums |
c9f955d374d21679a24c5d5aa98686303b3c5ebffc4d621d3fadb21a0faf5950
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 14, 2026.
Transparency logSecurity for modeldiffr 0.9.4
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for modeldiffr-0.9.4.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 14, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.