{
  "format_version": 3,
  "project": "jolt",
  "distribution": "jolt-model-runtime",
  "version": "0.1.0",
  "model_id": "Jolt-9B",
  "base_model": "Qwen/Qwen3.5-9B",
  "base_revision": "c202236235762e1c871ad0ccb60c8ee5ba337b9a",
  "adapter": "qwen3.5",
  "lora_rank": 8,
  "lora_alpha": 16,
  "lora_dropout": 0.0,
  "training_loss": "direct-composite",
  "image_pixels": 262144,
  "max_length": 2048,
  "backend": {
    "experimental_rocm_sdpa": false,
    "frozen_vision_cache_MiB": 128,
    "fused_norm_and_swiglu": false,
    "linear_patch": false,
    "shared_prefix": true,
    "triton_convolution": false
  },
  "architecture": {
    "readout": "last",
    "candidate_attention": false,
    "adapters": {
      "layers": [],
      "width": 64,
      "activation": "silu"
    },
    "head": "linear",
    "unfused_mlp_layers": [],
    "backbone_activation": {
      "layers": [],
      "initial_mix": 0.0,
      "learning_rate": 0.001
    },
    "gpu_memory_fraction": 0.9
  },
  "prompt_layout": "inline",
  "inference": "merged LoRA, BF16 inference, shared-prefix parallel candidate scoring",
  "source_hashes": {
    "train": "4e8b258c9e5d5a98810868df5d572a824a9d1d6c5aa557ab5123cb7aab5de373",
    "dev": "b736750efff3e30ae8a7772c482fa248899b8a1cf69ace104310e2540671d552"
  },
  "selected_run": "L2, seed 42",
  "selected_step": 3600,
  "selection": "maximum development macro accuracy within the run; no test selection",
  "seed": 42,
  "dev_macro_accuracy": 0.8530649038461539,
  "temperatures": {
    "choice": 1.0,
    "noul": 1.0,
    "score": 1.0
  },
  "weights_sha256": "55061e148dcf3270142a24e70e493ba893994c48de0416ceb92c13b361684d98",
  "training_precision": "BF16 base",
  "training_quantization": null
}