{
  "name": "TinyReceiptVQA-StructuredQA-22M-BPE1536-DirectNoValue-LoRA-Router-e100",
  "format": "tiny_receipt_vqa_split_onnx_v1",
  "library": "onnxruntime",
  "opset": 18,
  "architecture": {
    "vocab_size": 1536,
    "d_model": 320,
    "heads": 8,
    "enc_layers": 6,
    "dec_layers": 4,
    "ff_mult": 4,
    "dropout": 0.1,
    "max_q_len": 192,
    "max_out_len": 192,
    "img_tokens": 210,
    "vision_stride": 32,
    "use_adapters": true,
    "use_router": true,
    "adapter_bottleneck": 64,
    "adapter_families": 8,
    "lora_r": 8,
    "lora_alpha": 16.0,
    "lora_dropout": 0.05,
    "lora_targets": "transformer"
  },
  "source": {
    "selection": "epoch-100 last.pt; heldout was excluded from training and intermediate checkpointing, then used post hoc to compare e79 best.pt with e100 last.pt; e100 last.pt was selected for higher heldout answer exact",
    "target_schema": "direct_answer_no_value_v1",
    "training_precision": "BF16 autocast; saved source tensors are FP32",
    "tokenizer": {
      "type": "byte_fallback_bpe",
      "version": 1,
      "vocab_size": 1536,
      "normalization": "NFC",
      "tokenizer_hash": "5801826ac9092bdc984210af805d2ff20f0398b80adaf5963ef13b04771a8584"
    },
    "heldout_usage_note": "Heldout was excluded from training and intermediate checkpointing. After training it was used to compare e79 best.pt and e100 last.pt and to select e100 last.pt, so it is a development test rather than a final blind benchmark.",
    "pytorch_or_safetensors_checkpoint_distributed": false
  },
  "artifacts": {
    "encoder": {
      "path": "encoder_model.onnx",
      "bytes": 58500824,
      "sha256": "0c6469ef75eeb266c3dbc77c414d658f98f704264b72bdc7dfeed9cf37217db7"
    },
    "decoder": {
      "path": "decoder_model.onnx",
      "bytes": 29924405,
      "sha256": "3fff897777fa5e9805c01a6779812f4adf33e1294f4e2003dd823c04a3300b05"
    },
    "encoder_int8_w8a8": {
      "path": "encoder_model_int8.onnx",
      "bytes": 17808595,
      "sha256": "924a60cfa8172456131aedb2e672af2ad1de7e4fa975bf14f4837a49e09cec94",
      "quantize_linear_nodes": 120,
      "dequantize_linear_nodes": 167
    },
    "decoder_int8_w8a8": {
      "path": "decoder_model_int8.onnx",
      "bytes": 10917677,
      "sha256": "24dbf7411719fea928f1f249db9b981eedf4b3af6751bec0970f242ca3f99d0a",
      "quantize_linear_nodes": 109,
      "dequantize_linear_nodes": 146
    }
  },
  "evaluation": {
    "onnx_parity": {
      "status": "passed",
      "scope": "graph-level numerical parity on dynamic shapes",
      "exporter_validation": {
        "split_vs_model_max_abs_difference": 4.0531158447265625e-06,
        "onnx_checker": "passed",
        "onnxruntime": "1.23.2",
        "provider": "CPUExecutionProvider",
        "validation_shapes": {
          "image": [
            1,
            1,
            320,
            672
          ],
          "question_ids": [
            1,
            11
          ],
          "decoder_input_ids": [
            1,
            3
          ]
        },
        "explicit_family_ids_validated": [
          0,
          1,
          2,
          3,
          4,
          5,
          6,
          7
        ],
        "explicit_family_encoder_memory_max_abs_difference": 1.996755599975586e-05,
        "explicit_family_decoder_logits_max_abs_difference": 1.049041748046875e-05,
        "encoder_memory_max_abs_difference": 1.9311904907226562e-05,
        "encoder_mask_max_abs_difference": 0.0,
        "router_logits_max_abs_difference": 1.1444091796875e-05,
        "selected_family_ids_max_abs_difference": 0.0,
        "decoder_logits_max_abs_difference": 6.198883056640625e-06
      },
      "lora_merge": {
        "enabled": true,
        "layers_merged": 20,
        "module_names": [
          "encoder.layers.0.linear1",
          "encoder.layers.0.linear2",
          "encoder.layers.1.linear1",
          "encoder.layers.1.linear2",
          "encoder.layers.2.linear1",
          "encoder.layers.2.linear2",
          "encoder.layers.3.linear1",
          "encoder.layers.3.linear2",
          "encoder.layers.4.linear1",
          "encoder.layers.4.linear2",
          "encoder.layers.5.linear1",
          "encoder.layers.5.linear2",
          "decoder.layers.0.linear1",
          "decoder.layers.0.linear2",
          "decoder.layers.1.linear1",
          "decoder.layers.1.linear2",
          "decoder.layers.2.linear1",
          "decoder.layers.2.linear2",
          "decoder.layers.3.linear1",
          "decoder.layers.3.linear2"
        ],
        "max_abs_difference": 5.245208740234375e-06
      }
    },
    "int8_w8a8": {
      "format": "tiny_receipt_vqa_onnx_w8a8_u8s8_qdq_v1",
      "onnx_checker": "passed",
      "runtime_validation": {
        "onnxruntime": "passed",
        "provider": "CPUExecutionProvider",
        "family_ids": [
          -1,
          0,
          1,
          2,
          3,
          4,
          5,
          6,
          7
        ],
        "dynamic_prefix_lengths": [
          1,
          10,
          20
        ],
        "onnxruntime_version": "1.23.2"
      },
      "heldout_evaluation": {
        "status": "passed",
        "dataset": "eval/heldout supported phone/address records",
        "records_total": 2000,
        "routing": "learned_router",
        "evaluation_mode": "final_split_onnx_files",
        "runtime": {
          "library": "onnxruntime",
          "version": "1.23.2",
          "provider": "CUDAExecutionProvider"
        },
        "batch_size": 24,
        "artifacts": {
          "fp32": {
            "encoder": {
              "filename": "encoder_model.onnx",
              "bytes": 58500824,
              "sha256": "0c6469ef75eeb266c3dbc77c414d658f98f704264b72bdc7dfeed9cf37217db7"
            },
            "decoder": {
              "filename": "decoder_model.onnx",
              "bytes": 29924405,
              "sha256": "3fff897777fa5e9805c01a6779812f4adf33e1294f4e2003dd823c04a3300b05"
            }
          },
          "int8": {
            "encoder": {
              "filename": "encoder_model_int8.onnx",
              "bytes": 17808595,
              "sha256": "924a60cfa8172456131aedb2e672af2ad1de7e4fa975bf14f4837a49e09cec94"
            },
            "decoder": {
              "filename": "decoder_model_int8.onnx",
              "bytes": 10917677,
              "sha256": "24dbf7411719fea928f1f249db9b981eedf4b3af6751bec0970f242ca3f99d0a"
            }
          }
        },
        "summary": {
          "fp32": {
            "address": {
              "n": 1032,
              "answer_exact": 0.9496124031007752,
              "recomputed_n": 1032,
              "recomputed_answer_exact": 0.9486434108527132,
              "target_exact": 0.015503875968992248
            },
            "overall": {
              "n": 2000,
              "answer_exact": 0.9705,
              "recomputed_n": 2000,
              "recomputed_answer_exact": 0.97,
              "target_exact": 0.453
            },
            "phone_number": {
              "n": 968,
              "answer_exact": 0.9927685950413223,
              "recomputed_n": 968,
              "recomputed_answer_exact": 0.9927685950413223,
              "target_exact": 0.9194214876033058
            }
          },
          "int8": {
            "address": {
              "n": 1032,
              "answer_exact": 0.9437984496124031,
              "recomputed_n": 1032,
              "recomputed_answer_exact": 0.9457364341085271,
              "target_exact": 0.01065891472868217
            },
            "overall": {
              "n": 2000,
              "answer_exact": 0.968,
              "recomputed_n": 2000,
              "recomputed_answer_exact": 0.969,
              "target_exact": 0.4505
            },
            "phone_number": {
              "n": 968,
              "answer_exact": 0.993801652892562,
              "recomputed_n": 968,
              "recomputed_answer_exact": 0.993801652892562,
              "target_exact": 0.9194214876033058
            }
          }
        },
        "comparison": {
          "answer_agreement": 0.9925,
          "changed_predictions": 388,
          "n": 2000,
          "prediction_exact_agreement": 0.806,
          "selected_family_agreement": 1.0
        },
        "timing": {
          "fp32": {
            "encoder_seconds": 4.9500676593743265,
            "decoder_seconds": 20.797384655103087,
            "total_seconds": 25.747452314477414,
            "decoder_calls": 2953,
            "providers": [
              "CUDAExecutionProvider",
              "CPUExecutionProvider"
            ],
            "encoder": "encoder_model.onnx",
            "decoder": "decoder_model.onnx"
          },
          "int8": {
            "encoder_seconds": 5.913507376797497,
            "decoder_seconds": 26.702236368786544,
            "total_seconds": 32.61574374558404,
            "decoder_calls": 2944,
            "providers": [
              "CUDAExecutionProvider",
              "CPUExecutionProvider"
            ],
            "encoder": "encoder_model_int8.onnx",
            "decoder": "decoder_model_int8.onnx"
          }
        },
        "timing_note": "Static W8A8 speed depends on QDQ fusion and the selected execution provider."
      }
    },
    "source_heldout_reference": {
      "status": "source_pytorch_reference_not_full_onnx_rerun",
      "dataset": "eval/heldout supported phone/address records",
      "records_total": 2000,
      "routing": "learned_router",
      "source_model": "epoch-100 last.pt; heldout was excluded from training and intermediate checkpointing, then used post hoc to compare e79 best.pt with e100 last.pt; e100 last.pt was selected for higher heldout answer exact; the source checkpoint container is not distributed in this package",
      "onnx_fp32_agreement_note": "The source PyTorch reference and the final ONNX FP32 rerun agree exactly on all nine reported accuracy metrics across the overall, address, and phone subsets.",
      "summary": {
        "address": {
          "n": 1032,
          "answer_exact": 0.9496124031007752,
          "recomputed_n": 1032,
          "recomputed_answer_exact": 0.9486434108527132,
          "target_exact": 0.015503875968992248
        },
        "overall": {
          "n": 2000,
          "answer_exact": 0.9705,
          "recomputed_n": 2000,
          "recomputed_answer_exact": 0.97,
          "target_exact": 0.453
        },
        "phone_number": {
          "n": 968,
          "answer_exact": 0.9927685950413223,
          "recomputed_n": 968,
          "recomputed_answer_exact": 0.9927685950413223,
          "target_exact": 0.9194214876033058
        }
      }
    }
  },
  "license": "Apache-2.0"
}