LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B
Reported by Hugging Face1 source ↓
Hugging Face released experimental LLMON preview versions of its Granite 4.1 3B and 8B models, plus a toolkit. It says the models resisted more indirect prompt-injection attacks on tested benchmarks while matching Granite 4.1 Instruct on standard benchmarks.
Key points
- Hugging Face released experimental Granite 4.1 3B and 8B previews trained to read and generate LLMON
- The release includes a toolkit, examples, and tools to convert between JSON and LLMON
- Hugging Face reports fewer successful static prompt injections on its tested benchmarks
- Hugging Face says general benchmark performance was on par with Granite 4.1 Instruct
Details
LLMON uses reserved special tokens to mark the boundaries and meanings of text components, such as instructions, data and requirements. Hugging Face says this makes their relationships explicit to the model, rather than relying on it to infer structure from a flattened prompt. The format complements chat templates rather than replacing them.
Hugging Face reports that its previews avoided 47% more attacks on the OPI benchmark than the best frontier model it tested, and 95% more than the best open-source model. The company says similar trends appeared on the ASB and StruQ benchmarks. It describes these results as applying to static indirect prompt injections and says adaptive attacks unfolding over multiple turns remain an area of work.
The release also includes a GitHub toolkit with workflow examples and templates. Hugging Face says its conversion tools translate JSON to LLMON and back without data loss.
Written from coverage as of
Images
- Hugging Face
- Hugging Face
- Hugging Face
- Hugging Face