{
  "max_length": 256,
  "model_type": "llama",
  "task": "text-classification",
  "id2label": {
    "0": "NON_EOU",
    "1": "EOU"
  },
  "label2id": {
    "NON_EOU": 0,
    "EOU": 1
  },
  "pipeline_tag": "text-classification",
  "version": "1.0.0",
  "last_updated": "2025-03-13 12:01:15",
  "last_updated_by": "latishab",
  "formats_available": [
    "pytorch",
    "onnx",
    "onnx_quantized",
    "gguf"
  ],
  "inference_config": {
    "task_type": "text-classification",
    "runtime": "2.1",
    "framework": "pytorch",
    "memory": "auto",
    "conversion_config": {
      "quantization": "int8",
      "optimize_memory": true,
      "optimize_inference": true
    },
    "deployment": {
      "instance_type": "cpu-basic",
      "min_replicas": 1,
      "max_replicas": 3,
      "scaling_config": {
        "cpu_utilization": 80,
        "max_concurrent_requests": 100
      }
    },
    "api_config": {
      "timeout": 30,
      "max_tokens": 256,
      "max_batch_size": 32
    }
  }
}