⚠️ DO NOT USE THIS MODEL ⚠️

This model failed its own evaluation. It is kept publicly for research transparency only.

It is an AI-generated-text ("slop") detector fine-tuned from LiquidAI/LFM2.5-Encoder-350M on a 786k-row corpus (v1 data mix). Despite near-perfect benchmark numbers, it learned the wrong thing.

Measured results (proper eval, final weights)

slice AUROC verdict
blogs/gutenberg/writingprompts vs pile 1.00 era+length shortcut, meaningless
same vs storyscope 1.00 register shortcut, meaningless
coai strict (paired, style-isolated) 0.50 coin flip — worse than a 415-pattern regex floor (0.8643)
heldout laguna 0.756 weak
heldout local 0.861 weak
heldout deepseek_eval 0.752 weak

What went wrong

The training mix was 72% The-Pile with a length confound (AI rows avg 2221 chars, human rows ≤984). The model became an era-and-register classifier: it detects "old-fashioned or short prose vs modern long web text," not slop. Mid-training it scored 1.00 on coai's validation slice; by end of training catastrophic forgetting dropped it to chance while the easy registers absorbed all capacity.

Mid-training val AUROC read 0.9998. That number was a lie told by an unstratified metric. Trust nothing without the paired-slice eval above.

Why it's still public

Negative results are results. If you're building an LLM-text detector, this is what register-confounded training data produces: spectacular metrics, coin-flip reality. See the eval harness requirements: paired same-topic slices, unseen-generator holdouts, and a regex baseline you must beat.

  • training data: vstalingrady/itais (train_all.parquet, pile-heavy v1 mix)
  • eval report shape: cross-register matrix + strict-paired slice + held-out files

Do not use for content moderation, academic integrity, or anything else.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train vstalingrady/lfm-ckpt