YaRusty/distilroberta-fallacy-classifier

katana-guy fallacy classifier

distilroberta-base fine-tuned to spot 14 logical fallacies (or none) in short chat messages. Trained on LOGIC, CoCoLoFa and synthetic Discord-style messages.

file	use
model.safetensors, config.json, tokenizer	transformers: pipeline("text-classification", model="YaRusty/distilroberta-fallacy-classifier")
model.onnx	ONNX Runtime, inputs input_ids, attention_mask, fixed shape [1, 128]
model.rknn	RK3588 NPU (rknn-toolkit-lite2 2.3.2), fp16, also takes position_ids

Test F1

"real" is the LOGIC + CoCoLoFa test split, "discord" a held-out slice of the synthetic messages.

label	real	discord
ad hominem	0.69	0.63
appeal to authority	0.73	0.98
appeal to emotion	0.67	0.84
appeal to majority	0.78	0.86
appeal to nature	0.90	0.89
appeal to tradition	0.81	0.85
appeal to worse problems	0.78	0.72
circular reasoning	0.57	0.91
false causality	0.63	0.82
false dilemma	0.81	0.90
hasty generalization	0.60	0.91
none	0.76	0.90
red herring	0.42	0.81
slippery slope	0.77	0.88
straw man	0.74	0.95

License

Apache-2.0, like the base model. LOGIC and CoCoLoFa are published without a license, so their terms are not covered by it; check them with the dataset authors before commercial use.