developerscoffee/airo-media-actions-smollm2-135m

Airo Media Actions SmolLM2 135M

Task-specific local intent model for Airo TV edge intelligence. It translates natural-language media requests into deterministic media-action JSON consumed by the slm_edge_intelligence Rust/Flutter runtime.

Base model: HuggingFaceTB/SmolLM2-135M-Instruct

Files

• peft-adapter/ - PEFT LoRA adapter and tokenizer files.
• gguf/airo-media-actions-lora-f16.gguf - llama.cpp LoRA adapter for use with a compatible SmolLM2 135M Instruct GGUF base model.
• gguf/airo-media-actions-smollm2-135m-merged-f16.gguf - standalone merged F16 GGUF for easiest edge deployment.
• reports/airo_media_actions_rule_vs_slm.json - strict eval summary.
• reports/airo_media_actions_slm_predictions.jsonl - generated eval predictions.

Expected Output

Prompt:

### Instruction
Translate the Airo TV user request into a media action JSON object. Output JSON only. Do not answer conversationally.

### Input
Show Hindi news

### Response

Output:

{"clarification_required":false,"confidence":0.91,"constraints":{"genre":"news","language":"hi","live":true},"intent":"search","missing_fields":[],"tool":"media.search"}

llama.cpp

Use llama-completion for raw completion mode:

llama-completion \
  -m gguf/airo-media-actions-smollm2-135m-merged-f16.gguf \
  -p $'### Instruction\nTranslate the Airo TV user request into a media action JSON object. Output JSON only. Do not answer conversationally.\n\n### Input\nShow Hindi news\n\n### Response\n' \
  -n 128 \
  --temp 0 \
  --no-display-prompt

Base plus LoRA also works:

llama-completion \
  -m smollm2-135m-instruct-f16.gguf \
  --lora gguf/airo-media-actions-lora-f16.gguf \
  -p "$PROMPT" \
  -n 128 \
  --temp 0 \
  --no-display-prompt

Do not use llama-cli chat mode for this adapter; it can wrap prompts as conversation turns and produce malformed schema output.

Edge Runtime Configuration

For slm_edge_intelligence / edge-intent:

export EDGE_INTELLIGENCE_INTENT_BACKEND=llama.cpp+rule
export EDGE_INTELLIGENCE_LLAMA_CPP_BIN=/path/to/llama-completion
export EDGE_INTELLIGENCE_INTENT_MODEL=/path/to/airo-media-actions-smollm2-135m-merged-f16.gguf

Or, when using base plus LoRA:

export EDGE_INTELLIGENCE_INTENT_BACKEND=llama.cpp+rule
export EDGE_INTELLIGENCE_LLAMA_CPP_BIN=/path/to/llama-completion
export EDGE_INTELLIGENCE_INTENT_MODEL=/path/to/smollm2-135m-instruct-f16.gguf
export EDGE_INTELLIGENCE_INTENT_LORA=/path/to/airo-media-actions-lora-f16.gguf

The hybrid backend falls back to the Rust rule backend if local model output is invalid or low confidence.

Evaluation

Strict eval on 500 generated media-action examples:

Rule intent accuracy: 1.000
SLM intent accuracy: 1.000
SLM constraint exact accuracy: 1.000

No SLM failures were observed in reports/airo_media_actions_rule_vs_slm.json.