Current model catalog

This catalog covers all five artifacts in the current MoM multilingual embedding collection and all 14 artifacts in the classifier collection. A row represents one logical architecture and lists all of its published release shapes.

Embedding and reranking artifacts

Logical model	Published artifact	Architecture	Training method
mmBERT-32K foundation	mmbert-32k-yarn	ModernBERT masked-language encoder with a 32K YaRN context	Continued multilingual masked-language modeling
mmBERT-32K embedder	mmbert-embed-32k-2d-matryoshka	Bi-encoder with layer- and dimension-selectable embeddings	Multiple-negatives ranking with 2D Matryoshka supervision
mmBERT-32K reranker	mmbert-rerank-32k-2d-matryoshka	Cross-encoder with 20 layer/dimension scoring heads	Binary relevance loss averaged across all heads
Small multimodal embedder	multi-modal-embed-small	MiniLM, SigLIP, and Whisper-tiny towers with two-layer fusion; 384 dimensions	Staged image-text and audio-text contrastive alignment with Matryoshka loss
Large multimodal embedder	multi-modal-embed-large	mmBERT-32K, SigLIP2-SO400M, and Whisper-medium tri-encoder; 768 dimensions	Cached multiple-negatives ranking with hard negatives

See mmBERT-32K models and multimodal embeddings for the data flow, objectives, configuration, and commands.

Classifier artifacts

The first six classifiers below use the multilingual mmBERT-32K/ModernBERT encoder. The two published safety artifacts use jhu-clsp/mmBERT-base; the current safety workflow can train 32K successor artifacts. Sequence classifiers predict one label for the request; the PII model predicts a BIO label for each token.

Logical model	Labels or output	Published artifacts	Training method
Intent classifier	14 subject areas	merged, LoRA	LoRA sequence classification on MMLU-Pro plus fallback-intent examples
Jailbreak detector	benign, jailbreak	merged, LoRA	LoRA sequence classification on benign/toxic chat, attack data, and pattern augmentation
Feedback detector	Four feedback states	merged, LoRA	Class-weighted LoRA sequence classification
Modality router	AR, DIFFUSION, BOTH	merged, LoRA	LoRA with focal loss, class balancing, and optional synthetic mixed-modality prompts
Fact-check classifier	FACT_CHECK_NEEDED, NO_FACT_CHECK_NEEDED	merged, LoRA	Balanced LoRA sequence classification on information-seeking and non-information-seeking prompts
PII detector	17 entity types represented by 35 BIO labels	merged, LoRA	LoRA token classification with character-offset-to-token alignment
Safety Level 1	safe, unsafe	LoRA adapter	Deterministic prompt-only LoRA sequence classification
Safety Level 2	Nine hazard outputs	LoRA	Deterministic prompt-only LoRA sequence classification with a fixed taxonomy crosswalk

See classifier models for the first six tasks and safety classifiers for the two-level safety pipeline.

Choose a release shape

Use a merged artifact when the runtime expects a standalone Transformers model. Use a LoRA artifact when the runtime can load PEFT adapters and you want a smaller task-specific artifact. Both shapes must retain the same tokenizer, label order, base-model compatibility, and preprocessing contract used during training.

Model cards describe the released weights. The checked-in training configuration describes a new run. When they differ, treat the release as an existing artifact and the in-tree configuration as the source of truth for retraining; do not assume a new checkpoint will be bit-identical without the original data and run receipt.

The Level 1 safety artifact is a PEFT adapter even though its historical name ends in -merged. Inspect artifact contents and metadata instead of inferring the loading method from a suffix.