Slomin/slophammer_350m

SlopHammer 350M

A quantised (q4f16) ONNX detector for AI-generated text, small enough to run locally in the browser. Built for the SlopHammer extension.

The archive is self-contained: the graph, its tokenizer, and a seq_cls_contract.json describing how the model is fed and read. The model emits four buckets — fully human, lightly AI, moderately AI, fully AI — over the last 512 tokens of the input. Decision thresholds ship as calibration metadata inside the contract.

file	backbone	params	size
slophammer_350m_v0_1.zip	LFM2.5-350M-Base	350M	206 MB

Licence

Released under the LFM Open License v1.0, inherited from the LFM2.5-350M backbone. Full text: https://huggingface.co/LiquidAI/LFM2.5-350M-Base/blob/main/LICENSE

Two clauses are worth reading before use:

• Commercial use by an organisation at or above US$10,000,000 annual revenue requires a separate licence from Liquid AI.
• Derivative works must retain the original attribution notices and carry prominent notice that the files were modified.

Notice of modification: these weights are modified from LiquidAI/LFM2.5-350M-Base — supervised fine-tuning for sequence classification with a four-class head. This is not the unmodified base model.

Attribution

• Liquid AI — LFM2.5-350M-Base

Intended use and limits

Built for flagging AI-generated text in web content. It is a statistical signal, not proof of authorship, and should never be the sole basis for an accusation against a person. Known limitations, all measured:

• Entity-dense writing evades it. Recall drops sharply on text that is mostly proper nouns — travel and restaurant recommendations, sports recaps, news summaries, product round-ups.
• Short text is unreliable. Below ~40 words there is no measurement at all; accuracy falls off well before that.
• Polished human writing is the hard case. The threshold is set by the most AI-looking human writing measured — professional assistant-register prose — so careful, well-structured human text is where false accusations concentrate.
• English only.