Models

Code / Tool / Agentic Models (v2)

A unified family trained on the code + tool-output + prose benchmark. One model covers coding-agent answers, tool output, and prose, and emits typed spans (category + subcategory).

Model	Base	Output	Notes
lettucedect-v2-qwen-2b	Qwen3.5-2B	typed spans (+ reasoning)	generative; detection and typing in one pass
lettucedect-v2-mmbert-base	mmBERT-base	binary spans	fast encoder detector
lettucedect-v2-taxonomy-head	mmBERT-base	span typing	types encoder spans (cascade) — see Quick Start

English Models

Model	Base	Max Tokens	Example F1	Span F1
lettucedect-base-modernbert-en-v1	ModernBERT-base	4K	76.8%	SOTA
lettucedect-large-modernbert-en-v1	ModernBERT-large	4K	79.2%	SOTA

Multilingual Models

One checkpoint per language, in two sizes:

Model family	Base	Languages	Max Tokens
lettucedect-210m-eurobert-<lang>-v1	EuroBERT-210M	de, fr, es, it, pl, cn	8K
lettucedect-610m-eurobert-<lang>-v1	EuroBERT-610M	de, fr, es, it, pl, cn	8K

English is covered by the ModernBERT models above. See the multilingual collection on Hugging Face for every checkpoint. The EuroBERT models load remote code that requires transformers<5.

TinyLettuce (Distilled)

Smaller models for resource-constrained environments. See TinyLettuce docs.

Using a Model

from lettucedetect.models.inference import HallucinationDetector

detector = HallucinationDetector(
    method="transformer",
    model_path="KRLabsOrg/lettucedect-large-modernbert-en-v1"
)

Models are downloaded automatically from HuggingFace Hub on first use.