FRAME NL → Intent Compiler (English)

This model converts natural language into structured intent JSON:

{ "intent": "string", "params": { "key": "value" } }

Important:

• The model is NOT trusted for correctness
• Runtime MUST enforce substring validation
• Runtime MUST compute missing params

Example:

Input: send bob hello

Output: { "intent": "message.send", "params": { "to": "bob", "text": "hello" } }

Note: this is not perfect for a reason, frames runtime will fix it entirely over time

---

Run locally (llama.cpp)

Requirements: llama.cpp built (llama-cli binary available)

Example:

git clone https://huggingface.co/frameprotocol/frame-intent-english

/path/to/llama-cli \
  -m model.gguf \
  -p "send bob 5 dollars" \
  -n 100 \
  --temp 0.0

Expected output (approx):

{"intent":"payment.send","params":{"to":"bob","text":"5"}}

Run with validation (this repo):

python infer.py "send bob 5 dollars"

Expected output:

{"intent":"payment.send","params":{"to":"bob"}}

Notes:

• Output is strict JSON only
• Params not present in input are removed by validation

---

How it works

flowchart LR
    A[Natural language input] --> B[GGUF model\nllama-cli]
    B --> C[Raw JSON output]
    C --> D[Validation\ninfer.py]
    D --> E[Cleaned intent JSON]