PolicyLM

weights

open(Apache-2.0 safetensors on HF musubilabs/policylm-1.7b, ungated)

data

partial(three third-party public safety datasets named in the card, plus synthetic data that is not released)

code

partial(inference helper and quickstart in the model repository

license

Apache-2.0(OSI)

Musubi publishes PolicyLM's weights and inference helper under Apache-2.0 with no access gate, and the card names the three public safety datasets it was trained on. The synthetic data used alongside them and the training code are not released, so this is an open-weights release rather than a fully open model.

• https://huggingface.co/api/models/musubilabs/policylm-1.7b recorded 2026-10-19

  The Hugging Face API record reports "gated":false, carries the tag "license:apache-2.0" and the three "dataset:" tags, and lists "model.safetensors", "LICENSE" and the inference files "inference/policylm_infer.py" and "inference/quickstart.py". No training script is among the files.

• https://huggingface.co/musubilabs/policylm-1.7b/raw/main/LICENSE recorded 2026-10-19

  The LICENSE file in the model repository is the Apache License, Version 2.0, January 2004.

• https://huggingface.co/musubilabs/policylm-1.7b/raw/main/README.md recorded 2026-10-19

  Raw model card front matter declares "license: apache-2.0" and lists the datasets "nvidia/Nemotron-Safety-Guard-Dataset-v3", "Alibaba-AAIG/XGuard-Train-Open-200K" and "ToxicityPrompts/PolyGuardMix". The Training section says the model was fine-tuned "in five stages, on public safety datasets" and "synthetic data"; the synthetic data is not linked. The Files table describes `inference/` as "The helper `policylm_infer.py`, `quickstart.py`, `requirements.txt` and the base encoder's code", and the license section reads "**PolicyLM-1.7B:** Apache-2.0."

Adoption is measured by Hugging Face downloads of the one checkpoint Musubi publishes. A hosted version also runs on Baseten, and its calls are not counted, so the figure is a lower bound on use.

• https://huggingface.co/api/models/musubilabs/policylm-1.7b recorded 2026-10-19

  268 downloads in the trailing 30 days for musubilabs/policylm-1.7b

• https://raw.githubusercontent.com/roostorg/model-community/main/musubi-policylm/README.md recorded 2026-10-19

  The ROOST Model Community page for PolicyLM links a "Hosted version on Baseten" beside the Hugging Face download.

Like gpt-oss-safeguard, PolicyLM checks content against a policy the user writes, and it scores each category of that policy separately in one fast pass. It follows custom policies less accurately on Musubi's own benchmark, gives no written reason and does not check assistant responses, so it sits level with gpt-oss-safeguard rather than ahead.

• https://huggingface.co/musubilabs/policylm-1.7b/raw/main/README.md recorded 2026-10-19

  The card says "You give it a message and a policy and it returns a score from 0 to 1 for each category of the policy", offers "A built-in taxonomy (`aegis` mode)** of 23 categories from NVIDIA's Aegis 2.0", scores "Every category in one pass,** up to 16 per policy", and was "Evaluated on messages in 19 languages". Its custom-policy table gives accuracy 0.842 for PolicyLM-1.7B, 0.909 for gpt-oss-safeguard-20B, 0.829 for CoPE-B-A4B and 0.643 for Shieldstral-1.0-3B, with a 22 ms median on an H100. It lists "moderating assistant responses" as out of scope and points to other models when "The decision needs a written reason".

• https://raw.githubusercontent.com/roostorg/model-community/main/musubi-policylm/README.md recorded 2026-10-19

  The ROOST Model Community page describes PolicyLM-1.7B as "an open-weights (Apache-2.0) moderation classifier from Musubi Labs" that "returns a 0–1 score for every category in one pass, with no text generated".