Project summary

You know the common devs who wanna get the model from huggingface, and integrate into their project and later says model finetuning or model isn;t good from base or original model, or even when a guy finetune it, he doesn;t get to know where and how it;s done or get a proof that what;s model upto unless he go create benchmark with some questions

Now here comes my framework called Modeldiffr.

For now you need to give questions where this modeldiffr checks both model layer by layer, and tells you everything.It also tells which parts inside the model were edited. Every report is signed and tied to the exact model files, so anyone can verify it without trusting me

Check the code and you will get trust.
Currently now it comes with its own questions (13 standard benchmarks plus a quick built in set), and you still add your own too. but in future i want it to create test questions by itself for each model.

First result: I tested Qwen2.5 0.5B Instruct against a popular copy with its safety removed

• Unsafe requests: 12% for the original, 90% for the copy
• Toxic prompts answered, here 24% for the original, 100% for the copy
• General knowledge (MMLU) and reasoning (ARC) there;s no real change
• Modeldiffr also found exactly which weights were edited: 46 of 290, two per layer

So the copy kept its skills and lost its safety, and the tool proved it. This ran on my 4 GB vram laptop GPU, that;s how it;s done, now you know what it is...

And it;s just begining, still i got more ideas

Links:

• Code: https://github.com/amareshhebbar/modeldiffr
• Docs: https://modeldiffr.netlify.app
• Install: https://pypi.org/project/modeldiffr (pip install modeldiffr)
• Signed audit: https://huggingface.co/AmareshHebbar/modeldiffr_audits
• DOI: https://doi.org/10.5281/zenodo.23587022

I;m asking for money to test 20 popular edited models, measure how well the tool finds hidden backdoors, publish the results, and release version 1.0

Just optional proposal, if funding is more then i will keep a dashboard where devs just come and use my env to get the results without paying anything

And even i have the idea of doing to build further it as testing framework where sometimes model just breaks while chatting, or doesn;t give proper answer, then to tell dev where the layers are running through visuals, i got skills to develop that architecture

What are this project's goals? How will you achieve them?

1. Test as much as popular edited models on Hugging Face against their originals, and publish every signed report and this is done as per my budget
2. Measure the backdoor scanner honestly. I'll make 40 small test models, 20 with a hidden trigger planted and 20 clean, and publish how many the scanner catches and how often it raises a false alarm
3. Write up the results as a paper on arXiv and a post on LessWrong
4. Release version 1.0, including a GitHub Action that blocks a model update if its safety gets worse

The tool already works and is released. What's left is compute time, running the tests carefully, and writing

Plan:

• phase 1: create a pipeline to test model under daily budget, and this will be keep going
• phase 2: build & test that backdoor, dashboard and visuals for breaking open weight model in a website
• phase 3: publish what i did, writing paper
• phase 4: getting feedback and improve

How will you know it worked?

• 20 signed reports are public, and anyone can verify them with one command
• The backdoor results are published with clear numbers, good or bad
• The paper is on arXiv and the post is on LessWrong
• Version 1.0 is on PyPI
• Other people start using it by audits on model cards, GitHub stars, issues and contributions, but it;s not in my hand

How will this funding be used?

$6,000: my time, 4 months part time(after reducing tax, i might get less, so i asking this amount)

$3,000: cloud GPUs to run the tests(i need freedom to run)

$300: website, hosting and storage

$700: buffer

Total: $10,000

With the $3,300 minimum, I'll just pay for GPUs and hosting and do the work more slowly in my spare time, but for sure i will do it

Who is on your team? What's your track record on similar projects?

Just me, Amaresh Hebbar, from Bengaluru, India

I;m a Founding AI Engineer at Impossible AI, where I lead a team of 10 building AI products. I fine tune and publish open models on Hugging Face, so I see every day how easily an edit changes a model's behavior, and how rarely anyone checks(check my linkedin)

I built and released Modeldiffr myself. It has tests, CI, full documentation and signed reports, and it's released on PyPI with a citable DOI

• GitHub: https://github.com/amareshhebbar
• Hugging Face: https://huggingface.co/AmareshHebbar
• Modeldiffr: https://github.com/amareshhebbar/modeldiffr

What are the most likely causes and outcomes if this project fails?

1. The backdoor scanner might turn out to be weak. I;ll still publish the numbers, so people know not to rely on it. The rest of the tool works either way
2. GPU money might run out early. I;ll test fewer or smaller models, which are also the most downloaded ones
3. Few people might use it, because everyone are not into the tuning ai model. Everything stays free and public under Apache 2.0

How much money have you raised in the last 12 months, and from where?

Nope, just my money