[
  {
    "model_tag": "qwen2.5-coder:3b",
    "hotpotqa_accuracy_pct": 60.0,
    "hotpotqa_hallucination_pct": 40.0,
    "alfworld_success_pct": 60.0,
    "alfworld_invalid_action_pct": 0.0,
    "decode_tps": 21.4,
    "median_ttft_ms": 41.12,
    "peak_vram_mb": 2227.0,
    "gtx_1650_vram_pct": 54.4,
    "total_measured_trials": 20
  },
  {
    "model_tag": "qwen2.5-coder:1.5b",
    "hotpotqa_accuracy_pct": 93.3,
    "hotpotqa_hallucination_pct": 6.7,
    "alfworld_success_pct": 46.7,
    "alfworld_invalid_action_pct": 6.7,
    "decode_tps": 43.4,
    "median_ttft_ms": 32.12,
    "peak_vram_mb": 1377.9,
    "gtx_1650_vram_pct": 33.6,
    "total_measured_trials": 30
  },
  {
    "model_tag": "qwen2.5:3b",
    "hotpotqa_accuracy_pct": 100.0,
    "hotpotqa_hallucination_pct": 0.0,
    "alfworld_success_pct": 20.0,
    "alfworld_invalid_action_pct": 0.0,
    "decode_tps": 19.44,
    "median_ttft_ms": 46.75,
    "peak_vram_mb": 2228.2,
    "gtx_1650_vram_pct": 54.4,
    "total_measured_trials": 20
  },
  {
    "model_tag": "qwen2.5:0.5b",
    "hotpotqa_accuracy_pct": 100.0,
    "hotpotqa_hallucination_pct": 0.0,
    "alfworld_success_pct": 0.0,
    "alfworld_invalid_action_pct": 0.0,
    "decode_tps": 142.58,
    "median_ttft_ms": 18.83,
    "peak_vram_mb": 764.0,
    "gtx_1650_vram_pct": 18.7,
    "total_measured_trials": 10
  },
  {
    "model_tag": "qwen2.5:1.5b",
    "hotpotqa_accuracy_pct": 100.0,
    "hotpotqa_hallucination_pct": 0.0,
    "alfworld_success_pct": 0.0,
    "alfworld_invalid_action_pct": 26.7,
    "decode_tps": 48.39,
    "median_ttft_ms": 63.49,
    "peak_vram_mb": 1313.7,
    "gtx_1650_vram_pct": 32.1,
    "total_measured_trials": 30
  },
  {
    "model_tag": "qwen2.5:0.5b-instruct-q8_0",
    "hotpotqa_accuracy_pct": 73.3,
    "hotpotqa_hallucination_pct": 13.3,
    "alfworld_success_pct": 0.0,
    "alfworld_invalid_action_pct": 0.0,
    "decode_tps": 108.74,
    "median_ttft_ms": 181.09,
    "peak_vram_mb": 2014.0,
    "gtx_1650_vram_pct": 49.2,
    "total_measured_trials": 30
  },
  {
    "model_tag": "deepseek-r1:1.5b",
    "hotpotqa_accuracy_pct": 66.7,
    "hotpotqa_hallucination_pct": 13.3,
    "alfworld_success_pct": 0.0,
    "alfworld_invalid_action_pct": 20.0,
    "decode_tps": 37.44,
    "median_ttft_ms": 11159.98,
    "peak_vram_mb": 1872.3,
    "gtx_1650_vram_pct": 45.7,
    "total_measured_trials": 30
  },
  {
    "model_tag": "llama3.2:1b",
    "hotpotqa_accuracy_pct": 40.0,
    "hotpotqa_hallucination_pct": 60.0,
    "alfworld_success_pct": 0.0,
    "alfworld_invalid_action_pct": 40.0,
    "decode_tps": 43.91,
    "median_ttft_ms": 25.03,
    "peak_vram_mb": 1622.0,
    "gtx_1650_vram_pct": 39.6,
    "total_measured_trials": 20
  }
]