EthicaLabs releases 114M NSFW text classifier on HuggingFace

A new 114-million-parameter text classifier designed to detect NSFW content landed on HuggingFace this week, offering practitioners a lightweight tool for content moderation pipelines.

Echo-SmolTools-114M-NSFW-CLF, from EthicaLabs, is a fine-tune of the company's Echo-DSRN-114M v0.1.2 base checkpoint. At 114 million parameters, the model sits in the "smol" weight class—small enough to run on CPU or modest GPU setups while still handling real-time inference. The weights ship in safetensors format with custom code, and the model integrates with the transformers library.

EthicaLabs has not published benchmark accuracy figures, training-data details, or inference-speed numbers on the model card. The card also does not specify which NSFW categories the classifier targets—whether it distinguishes between violence, sexual content, hate speech, or other adult themes, or simply outputs a binary safe/unsafe label. No download or like counts are visible yet, suggesting the release is fresh.

Classifiers of this size typically serve as first-pass filters in moderation stacks or as guardrails for larger generative models. For developers building chat interfaces, image-captioning pipelines, or user-generated-content platforms, a 114M classifier can run inline without the latency or cost of a cloud API call. The custom-code requirement means users will need to trust the repository's Python modules rather than relying solely on the standard transformers API—a friction point for some production environments.

Adoption will likely depend on whether EthicaLabs publishes eval numbers and example code. The model is tagged "not-for-all-audiences" on HuggingFace, a metadata flag that typically indicates the model handles or detects sensitive content.