{
  "release": "MemoryATHENA-20260922",
  "description": "Compact release metadata and aggregate results for MemoryATHENA QA and general-NLP experiments.",
  "base_model": "mistralai/Mistral-7B-v0.3",
  "seed": 42,
  "collection": "OLAResearchX/memoryathena-6ab21a89e66b6ff2ded749ca",
  "model_repositories": [
    "OLAResearchX/memoryathena-qa-20260922",
    "OLAResearchX/memoryathena-general-nlp-20260922"
  ],
  "training_data": {
    "qa": {
      "reader_and_router": "English Wikipedia-2021 causal next-token-only stream",
      "source_memory_dependency": "OLAResearchX/xmemtransfer-llama2-source-20m",
      "token_budget": 20000000,
      "seed": 42
    },
    "general_nlp": {
      "mixture": [
        "WikiText-103",
        "Amazon Polarity review text",
        "CC-News",
        "IMDB text"
      ],
      "mixing": "equal-token mixture",
      "tokens_per_phase": 20000000,
      "sequence_length": 2048,
      "batch_size": 1,
      "seed": 42
    }
  },
  "test_data": {
    "qa": {
      "NQ": "Natural Questions Open paper-aligned evaluation split",
      "WebQA": "WebQuestions/WebQA paper-aligned evaluation split",
      "TriviaQA": "rc.nocontext validation split",
      "TruthfulQA": "multiple_choice split",
      "HotpotQA": "distractor split"
    },
    "general_nlp": {
      "SST2": {
        "split": "paper-aligned evaluation",
        "examples": 872
      },
      "MR": {
        "split": "paper-aligned evaluation",
        "examples": 2000
      },
      "CR": {
        "split": "paper-aligned evaluation",
        "examples": 2000
      },
      "RT": {
        "split": "paper-aligned evaluation",
        "examples": 1066
      },
      "AGN": {
        "split": "paper-aligned evaluation",
        "examples": 7600
      },
      "Yahoo": {
        "split": "Yahoo Answers Topics paper-aligned evaluation",
        "examples": 60000
      }
    }
  },
  "disclosures": [
    "single training seed",
    "no raw examples or labels redistributed",
    "historical Vanilla NLP scorer differs from the E/router scorer",
    "Yahoo tau=1.0 is a post-hoc test-set sweep"
  ],
  "excluded": [
    "ARTIFACTS.md",
    "raw datasets",
    "raw test examples",
    "labels",
    "raw predictions",
    "credentials",
    "internal filesystem paths"
  ]
}