← Back
Hugging Face Blog

LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B

•9 min read•
#llm
Level:Advanced
For:AI Engineers
✦TL;DR

The announcement introduces LLMON, an LLM‑native markup language that uses reserved special‑token delimiters to make context structure explicit, and ships two experimental preview models—granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview—trained to consume and generate LLMON‑marked text. Benchmarks show indirect prompt‑injection attack success rates drop to nearly zero on the LLMON models, while regular instruction‑tuned Granite 4.1 models remain highly vulnerable, all without sacrificing general capability. Engineers can now replace fragile JSON framing with LLMON, leverage automatic “LLMONizing” in Mellea, and fine‑tune custom tags via LoRA.

⚡ Key Takeaways

  • granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview are the released preview models.
  • LLMON uses reserved special‑token delimiters (XML‑like tags) to separate structure from content.
  • Attack success rates on benchmark prompt‑injection tests drop to nearly zero compared to high hijack rates on standard models.
  • Use the provided GitHub toolkit’s converters to transform existing JSON into LLMON markup.
  • Custom tag schemas can be learned by applying a LoRA on top of the LLMON models.

🔧 Tools & Libraries

GitHub repoMelleaLoRAconverters
💡 Why It Matters

By giving models a clear, token‑level representation of instructions, data, and requirements, LLMON dramatically reduces prompt‑injection risk and eliminates the need for brittle JSON handling, enabling more reliable and secure LLM integration in production systems.

✅ Practical Steps

  1. Clone the released GitHub repo and run the provided converter to turn your JSON payloads into LLMON markup before sending them to a Granite 4.1 LLMON model.
  2. Deploy granite-4.1-8b-LLMON-preview (or the 3B variant) via your preferred inference platform and feed it LLMON‑formatted prompts to benefit from the built‑in injection resistance.
  3. If you need custom structures, train a LoRA on top of the LLMON model using your own tag definitions as demonstrated in the repo examples.

Want the full story? Read the original article.

Read on Hugging Face Blog ↗

More like this

Introducing Gemini 4 Neon

Google DeepMind•#llm

DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency

Hugging Face Blog•#llm

NVIDIA RTX Spark Compact Desktops Out Now

NVIDIA Blog•#compute

Helping AI text detectors keep up with changing conditions

MIT News AI•#llm

EXPLORE AI NEWS

Daily hand-picked stories on LLMs, RAG, agents and production AI — curated for engineers who ship.

BROWSE NEWS

GET THE WEEKLY DIGEST

Join engineers getting the Monday signal-over-noise AI breakdown. No spam, unsubscribe anytime.

LEARN AI ENGINEERING

Curated courses, research papers, repos and tutorials built for engineers leveling up in AI.

START LEARNING