Fineprint: a 1.2B model that flags unfair terms on your own laptop

Fineprint is a Chrome extension and a fine-tuned 1.2B-parameter model that reads a terms-of-service page on your own machine. It quotes each clause that costs you money or rights and grades the page from A to E. It's for people who click "I agree", and nothing leaves the device.

I built the model, data, evaluation and pipeline. Mukesh Agrawal built the Chrome extension. We made Fineprint at Moonfire Build Night in London on 22 October 2026, a 3-hour hackathon, and it won the White Circle track and placed 2nd overall.

What it flags

Free trials that turn into paid plans, no refunds, price rises, hard cancellation, forced arbitration, data shared for ads. The labels have 18 categories, and 7 are red flags. The grade depends only on how many distinct red flags the page has. It isn't legal advice.

On real pages from the README, Spotify Premium terms get a D, and the New York Times and Netflix get a C.

Why it's fast

The model has one narrow job: answer one question per clause, in a fixed format.

1. The extension splits the page into clauses and lowercases them, because the training text is lowercase.
2. The model is fully fine-tuned on a 12-word system prompt: "Flag this Terms of Service or privacy policy clause as JSON." A longer instruction used to cost about 210 tokens per request. The short one costs about 80 and made it about 3 times faster.
3. The reply is {"flags":[...],"plain":"..."}. The request sets a stop string, so a harmless clause ends right after {"flags": [], about 5 output tokens.
4. The extension sends 32 clauses at a time to a local llama-server.
5. Clauses with money words (renew, trial, cancel, refund, fee, arbitration) go first, so the first red card shows up in 1 to 2 seconds.

Clicking a card highlights the clause with the CSS Custom Highlight API, so the page itself doesn't change.

Where the labels come from

Expert-labelled public datasets: UNFAIR-ToS (via LexGLUE), extra online terms, and OPP-115 privacy policies. I added 503 clauses from 20 real subscription services and 1,200 synthetic ones. Claude Sonnet 5.5 wrote the plain-English sentence for each clause and didn't choose any flags.

Scores

I scored the quantized GGUF model through llama-server, so the numbers match what runs on a laptop. The test set is 1,607 clauses from UNFAIR-ToS.

Model	Micro-F1	Unfair-detection F1	Harmless clauses flagged
Base LFM2.5-1.2B	0.056	0.216	970 of 1,435
Fineprint v1	0.778	0.823	21

The base model, with no fine-tuning, flags most harmless clauses, so most of what it flags is wrong. Fineprint v1 is a full fine-tune: 3 epochs at a learning rate of 2e-5, on White Circle's Halo tool and one H200. Its weakest category is arbitration, at 0.667 F1.

The shipped model is v5. The README scores it on a validation set and the repo has no metrics file for it, so I'm not reporting its scores here.

On an M4 Pro with the Q4_K_M model (about 700 MB), the 170-clause Spotify Premium terms take about 14 seconds. My first version took 56 seconds on a similar page. That's one run on one machine.

Limits

• The fine-tuned model file isn't in the repo, and the README has no download link, so you can't run the fine-tuned model from the README alone. I still need to host it.
• It sorts clauses into a fixed set of 18 labels. It doesn't write free-form summaries.
• The extension doesn't read iframes, shadow DOM or the PDF viewer, so a page that builds its terms with JavaScript can come back empty.
• The clause splitter breaks after "e.g.".
• It needs llama-server running on your machine.
• The labels come from 50 mostly English terms-of-service documents.

Against cloud summarizers

Cloud AI summarizers write free-form summaries, which Fineprint doesn't, and they send every page to a third party. Fineprint gives up the flexibility to keep the text on your machine.

Try it

Source: https://github.com/hridaya423/fineprint. The packaged extension is in the v1.0 release as fineprint-extension.zip.

brew install llama.cpp
llama-server -m fineprint-v5-Q4_K_M.gguf --port 8081 -np 32 -c 32768 -b 4096 -ub 2048 -fa on

Then open chrome://extensions, turn on Developer mode, and choose Load unpacked on the chrome/ folder. Open a terms page, click the icon, choose Fine-tuned, and click "Check this page".

For an offline demo, run python3 -m http.server 8765 --bind 127.0.0.1 -d chrome/fixtures and open spotify_premium.html.