ibm-granite

/

granitelib-core-r1.0

like 12

Follow

IBM Granite 4.27k

Models Added

• uncertainty/lora/granite-4.0-micro - LoRA adapter
• uncertainty/alora/granite-4.0-micro - aLoRA adapter

How It Works

1. Base model generates a response to a user query
2. UQ adapter is invoked with <certainty> token
3. Model outputs {"score": "0-9"} indicating confidence level
4. Score is transformed to calibrated likelihood (0.05-0.95) via io.yaml

Training Details

• Base model: ibm-granite/granite-4.0-micro
• LoRA config: rank=32, alpha=32, target_modules=[q_proj, k_proj, v_proj]
• Training samples: ~215K examples with certainty labels
• Evaluation: 57 MMLU subsets (14,042 samples)

Files

• adapter_config.json - PEFT adapter configuration
• adapter_model.safetensors - Trained adapter weights
• io.yaml - Output schema and transformation rules