modeldiffr 0.9.3

Know what changed between a model and anything built from it.

Documentation | Quickstart | CLI reference | Roadmap | Report a bug | Contact

Modeldiffr audits a derivative language model against its base: a fine tune, a quantized build, a merge or an edited checkpoint. Both models answer the same items under the same template. Every change comes with a paired interval, and the report is bound to the exact weights, so anyone can verify it.

pip install "modeldiffr[hf]"
modeldiffr audit Qwen/Qwen2.5-0.5B-Instruct your-org/your-qwen-finetune

Preview, v0.9. Every layer has shipped. Version 1.0 waits for the published study and model zoo results described in the roadmap. Report schema 0.3 is the 1.0 candidate.

Why

Hugging Face hosts hundreds of thousands of derivative models. A model card might say "fine tuned for medical Q&A" or "4 bit, near lossless", but nobody can cheaply check whether general capability dropped, whether safety behavior shifted, or whether a merge is actually better than its parents. Benchmarks run by hand show that a score moved, not whether the move is real, and comparisons are often unfair because prompts, templates or decoding settings differ between runs.

Modeldiffr makes the comparison fair, statistical, repeatable and checkable.

What it answers

Every behavioral metric is a paired delta with a 95% bootstrap interval, corrected for multiple comparisons, with a minimum effect below which a change is called negligible. Identical models produce a clean report.

Install

pip install "modeldiffr[hf]"

The core package has no dependencies, which keeps it light for CI jobs that only read reports. 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 show the format only. Run the command to get real values for your models.

report.html is a self contained page you can open anywhere or attach to a pull request. report.json holds every item, setting and hash. Full walkthrough: Quickstart.

Common tasks

Run standard benchmarks

pip install "modeldiffr[hf,tasks]"
modeldiffr audit BASE DERIVATIVE --suite standard --plan-only --estimate
modeldiffr audit BASE DERIVATIVE --suite standard

MMLU, ARC Challenge, HellaSwag, TruthfulQA, GSM8K, MGSM, IFEval, XSTest, OR Bench, HumanEval, a long context needle test and output divergence, every dataset pinned to a commit. Try it on a free GPU with the Colab notebook. See Suites and tasks.

Audit 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.

Verify any report

modeldiffr verify report.json
modeldiffr verify report.json --models

--models also hashes the real model files at the recorded commits. 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, Publish an audit 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

See Find changed features, Test what drives a change and Scan for hidden behavior.

Gate a checkpoint in CI

permissions:
  id-token: write
  contents: read
steps:
  - uses: actions/checkout@v4
  - uses: amareshhebbar/modeldiffr@v0.9.3
    with:
      base: org/base_model
      derivative: ./checkpoints/candidate
      suite: lite

The action audits, fails the job on a significant change, signs the report with Sigstore and uploads it. See GitHub Action and 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.

How it works

1. Both models are fingerprinted (weights, tokenizer, chat template) and checked for comparability. The audit gets an id computed from everything that defines it, before anything runs.
2. Both models run on exactly the same inputs, with one shared chat template when theirs differ.
3. Multiple choice items are scored by log likelihood, so the result does not depend on answer formatting.
4. Paired bootstrap resampling turns per item results into deltas with intervals and p values, corrected across all metrics.
5. Every item and result is committed to in a Merkle tree, and a statement hash covers every claim, so modeldiffr verify can recompute the report.
6. Optional layers look inside: crosscoders for features, patching for causal shares, and a scan for unusual inputs.

Details: How it works.

Roadmap

Full plan: Roadmap and PLAN.md.

Limitations

• The built in quick suite is a smoke test. Use lite, standard or full for a release decision.
• Large models need a GPU. Both models are held in memory at once.
• A significant change shows that something moved, not why. Features and causal shares are leads, not proofs.
• The anomaly scan makes no claim about real models until its rates on a public zoo are published.

All of them: Limitations.

Help and contact

More: Contact and support.

Responsible use

Modeldiffr is a measurement tool. It never modifies model weights, and its reports describe what changed and how much, not how to reproduce or reverse a change. If you find a serious safety regression or hidden behavior in a public model, please contact its publisher before posting the finding. See Responsible use.

Contributing

Contributions are welcome, from bug reports to new evaluation suites. Start with CONTRIBUTING.md, GOVERNANCE.md and issues labelled good first issue.

Citation

If Modeldiffr helps your research, please cite it using CITATION.cff or:

@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 = {0.9.3},
  doi     = {10.5281/zenodo.23587022}
}

Acknowledgements

The representational layer builds on crosscoder based model diffing, introduced in Sparse Crosscoders for Cross Layer Features and Model Diffing (Lindsey et al., Anthropic, 2024), and on the work that followed it. Behavioral evaluation draws on open harnesses such as EleutherAI's lm evaluation harness and UK AISI's Inspect. The proof layer follows RFC 8785 and RFC 6962 and interoperates with Sigstore and in toto.

License

Apache 2.0. Copyright 2026 Amaresh Hebbar.