{
  "status": "prepared_not_published",
  "base_repository": "https://github.com/vcruz305/llama.cpp",
  "base_revision": "5210c7c5ed61dddaee6ed476623abf4b63093d16",
  "base_source_archive_sha256": "3e0ad09b1f2175a170d87310731ac12d1bf52d8ab0b1f4fbdb19298457ae73e6",
  "patch_sha256": "1a70ed3aacb89fec18ff7fdd263702a767f1e2606b3a57bad425035c1ed29fe6",
  "model": "pfeifferj/DeepSeek-V4.1-Flash-GSQ-RCO-GGUF",
  "model_file": "DeepSeek-V4.1-Flash-GSQ-RCO-3.0bit.gguf",
  "model_sha256": "11f46543370256ef616b6e458b6950e148625b5f8b7545173d124445d3393c53",
  "official_reference_revision": "dba1be0a40aa45a94ad051997016db3960a90277",
  "runtime_artifacts_sha256": {
    "libllama.so.0.4.0": "9abf951448ec289625e27aff805986ea80780afc1c7e9b745162f3fd5cf3d7d8",
    "libggml-cuda.so.0.23.0": "b33b8b786f7c61c1cc10a6699d1f5d778eedfa094dec34f9013c38a82c8f5648",
    "mc-score-deepseek-f32": "e5b8a91ba2547b02a14668335b590ddcac185b6b606c3b10f50b0cf29834fb4c",
    "llama-perplexity": "51197334ee7c016569cbbb893eb20150cc2714ef46a161c34f00ae5175da3632"
  },
  "evaluation_settings": {
    "context": 2048,
    "batch": 2048,
    "microbatch": 2048,
    "sequences": 1,
    "threads": 16,
    "flash_attention": false,
    "cache_k": "f32",
    "cache_v": "f32",
    "CUDA_VISIBLE_DEVICES": "2,3",
    "NVIDIA_TF32_OVERRIDE": "0",
    "GGML_CUDA_MMF_F32_DISABLE": "1",
    "GGML_CUDA_REFERENCE_F32": "1",
    "engram_placement": "lazy CPU mmap",
    "method": "single prompt forward, raw answer-letter single-token log-probabilities, variable option counts from archived task index, no chat template",
    "tokenization": {
      "scorer_add_special": true,
      "gguf_add_bos_token": false,
      "gguf_add_eos_token": false,
      "effective_raw_prompt": "No forced BOS or EOS; preserve original Spark scorer and model metadata behavior",
      "chat_prompt": "Pinned official encoder supplies BOS explicitly"
    }
  },
  "throughput_pilot": {
    "questions": 20,
    "decode_seconds": 32.749756,
    "seconds_per_question": 1.6374878,
    "projected_2000_decode_minutes": 54.582926666666665
  },
  "ppl_sidecar_patch_sha256": "2ed02e1d9c7a77303ef50153792b3a40da19cadc394ae0d79aa76d59c8ba2068",
  "question_state_isolation": "research/question-state-isolation.md",
  "tokenizer_parity": {
    "status": "PASS",
    "vocabulary_size": 129280,
    "compared": "Every native token ID, exact token string, and decoded byte piece including special tokens",
    "token_string_mapping_sha256": "ad4e020abc59468b66b83039c5663753efe861c0af3cda9034bfc283d298059c",
    "token_byte_mapping_sha256": "dced44b89abfc32450c7398912681c477f0da4db310b0caa1bbdcf14ed0e9c7d",
    "tokenizer_sha256": "c90dfa01249db1be4245780a052ede752e1361c612ac6d08e2bdada7d599476b",
    "native_dump_sha256": "c53162bbc06168cdb73b3e69f41c13fb1ed4b8603dd33ac0f85e54e78df9bebc"
  },
  "accepted_mmlu": {
    "correct": 1220,
    "n": 2000,
    "accuracy": 0.61,
    "tsv_sha256": "7ac61bdb80ce8c0c41079c2d2364968442380785bfdb33246d3b0846704254e1",
    "physical_gpu_pairs": [
      [
        2,
        3
      ],
      [
        0,
        1
      ]
    ],
    "cross_pair_parity": "All20 predictions and every printed option log probability match exactly",
    "final_merge": {
      "status": "PASS",
      "rows": 2000,
      "primary": [
        0,
        1500
      ],
      "tail": [
        1700,
        2000
      ],
      "sha256": "7ac61bdb80ce8c0c41079c2d2364968442380785bfdb33246d3b0846704254e1",
      "reused_initial_rows": 824,
      "second_restart": 1379,
      "middle": [
        1500,
        1700
      ],
      "same_binary_and_settings": true
    }
  }
}