Mila Suicide Prevention Guardrail

weights

open(Apache-2.0 safetensors on HF mila-ai4h/Mila-Suicide-Prevention-Output-Guardrail, ungated)

data

described(the card says a synthetic training set was generated from the published constitution.md

code

partial(load and inference examples on the card only

license

Apache-2.0(OSI)

Mila publishes the classifier's weights under Apache-2.0 with no access gate, together with the written policy its training labels were drawn from. The synthetic training set built from that policy is not released, and the only code is a loading and inference example, so this is an open-weights release rather than a fully open model.

• https://huggingface.co/api/models/mila-ai4h/Mila-Suicide-Prevention-Output-Guardrail recorded 2026-10-19

  The Hugging Face API record reports "gated":false, carries the tag "license:apache-2.0", and lists the files "model.safetensors", "constitution.md", the tokenizer and config. It declares no dataset tag and ships no training script.

• https://huggingface.co/mila-ai4h/Mila-Suicide-Prevention-Output-Guardrail/raw/main/constitution.md recorded 2026-10-19

  The published constitution sorts content into "Harmful in-scope (intentional self-harm)", "Harmless in-scope (self-harm related, not enabling)", "Harmful out-of-scope (not self-harm)" and "Harmless out-of-scope (everyday / benign)", with example topics for each. It is a policy text, not a dataset or generation script.

• https://huggingface.co/mila-ai4h/Mila-Suicide-Prevention-Output-Guardrail/raw/main/README.md recorded 2026-10-19

  Raw model card front matter declares "license: apache-2.0" and "base_model: bert-base-uncased". The change log says the release "Generated a new synthetic training dataset" and "Trained `bert-base-uncased` on the new dataset", and the training section says the constitution "shapes **data and labels only**" and "can be found in the artifacts, in `constitution.md`". The card links no training dataset; its only code is a "Quick Load Example" and an "Inference Example".

Adoption is measured by Hugging Face downloads of the one checkpoint Mila publishes. No other channel is declared for the model, so this count is the only visible measure of its use.

• https://huggingface.co/api/models/mila-ai4h/Mila-Suicide-Prevention-Output-Guardrail recorded 2026-10-19

  2,279 downloads in the trailing 30 days for mila-ai4h/Mila-Suicide-Prevention-Output-Guardrail

The guardrail checks one thing, whether a chatbot's English reply encourages or instructs suicide or self-harm, and is cheap enough to run on every response. Llama Guard screens prompts and replies across fourteen hazard categories in eight languages, suicide and self-harm among them, so a single-harm, reply-only filter sits two steps below it.

• https://huggingface.co/api/models/mila-ai4h/Mila-Suicide-Prevention-Output-Guardrail recorded 2026-10-19

  The Hugging Face API record gives the architecture "BertForSequenceClassification" with 109,483,778 parameters in a single "model.safetensors" file.

• https://huggingface.co/mila-ai4h/Mila-Suicide-Prevention-Output-Guardrail/raw/main/README.md recorded 2026-10-19

  The card describes "a binary sequence classifier that takes an LLM **output** (assistant response text)", says "It is **not** an input guardrail, **not** a general harm classifier", aligns its policy with the "Suicide and Self-Harm" category from MLCommons AILuminate, lists "Languages other than English" as out of scope, gives a "Max sequence length" of 512 tokens, and reports precision 0.908, recall 0.802 and F1 0.852 at "Default threshold τ = 0.5" on 494 internal examples. It says the classifier "does not provide explanations or category breakdowns".

• https://roost.tools/blog/releasing-mila-s-suicide-prevention-guardrail-lightweight-open-source-safeguards-for-real-time-detection/ recorded 2026-10-19

  ROOST's release post calls the guardrail a "low-latency BERT model" that "can run locally, even on an ordinary CPU", "Targeting exclusively model outputs", and describes it as "a high-recall, low-footprint narrow safety add-on filter" rather than "a general-purpose safety moderation filter".