What does Claim Splitter do?

It takes an English claim or argument and returns its claim type, core claim, premises, conclusion, hidden assumptions, confounders, evidence needed and inference weaknesses.

Is my text uploaded anywhere?

The tool's own code sends it nowhere: inference runs in your browser with wllama, and the only network request it starts is the one-time 253 MB model download from Hugging Face. The site's analytics tags load on every page.

Why does it need a 253 MB download?

That file is the model: a 4-bit GGUF of a 270-million-parameter fine-tune. The browser caches it after the first load.

It is small and wrong sometimes. The reported section-header F1 of 0.91 on 68 held-out prompts measures whether the output has the right structure, not whether it is correct, and one documented run misstates its own premise.

Does it work for text that is not English?

No. A check in the page and the model's refusal training both decline non-English input.

GemmaClaim-270M, a fine-tune of Google's Gemma 3 270M, published on Hugging Face.

← Home

Claim Splitter: premises and hidden assumptions, in your browser

Claim Splitter takes an English claim and lists its premises, hidden assumptions, confounders and weak points. A 270M-parameter model does it inside this tab, and the tool's code never uploads your text.

Model on Hugging Face ↗

Model not loaded. Downloaded once, then cached in your browser.

Try

How it works

Model: Gemma 3 270M, fine-tuned to decompose claims (full fine-tune, then LoRA).
Data: About 1,400 synthetic examples, including refusals for non-English and off-topic input.
Runtime: 4-bit GGUF (253 MB) run by llama.cpp compiled to WebAssembly, on the GPU through WebGPU where the browser has it. Greedy decoding.
Score: Section-header F1 against reference answers went from 0.42 to 0.91 on 68 held-out prompts.

It is small and it is wrong sometimes. Read the output as a sketch of a claim's structure, not a verdict.

Quick answers

What does Claim Splitter do?: It takes an English claim or argument and returns its claim type, core claim, premises, conclusion, hidden assumptions, confounders, evidence needed and inference weaknesses.
Is my text uploaded anywhere?: The tool's own code sends it nowhere: inference runs in your browser with wllama, and the only network request it starts is the one-time 253 MB model download from Hugging Face. The site's analytics tags load on every page.
Why does it need a 253 MB download?: That file is the model: a 4-bit GGUF of a 270-million-parameter fine-tune. The browser caches it after the first load.
How accurate is it?: It is small and wrong sometimes. The reported section-header F1 of 0.91 on 68 held-out prompts measures whether the output has the right structure, not whether it is correct, and one documented run misstates its own premise.
Does it work for text that is not English?: No. A check in the page and the model's refusal training both decline non-English input.
Which model is it?: GemmaClaim-270M, a fine-tune of Google's Gemma 3 270M, published on Hugging Face.

How it works and what it returns

How a 270M-parameter claim analyzer runs in your browser: the model, the runtime, the prompt and the language check.
Three real runs, one with a visible mistake: the eight-section output, verbatim.