{
  "schema": "aner.native-pretrained-qualification/1",
  "date": "2026-10-30",
  "status": "development source; not included in released 0.0.5 packages",
  "profile": {
    "training": "CPU FP32",
    "runtime": "LibTorch C++ 2.9.0",
    "host_architecture": "arm64",
    "python_runtime_required": false,
    "native_engine_revision": "5047748f114564fcf271e6a648f60e2d2eae6264"
  },
  "pretrained": {
    "repository": "HuggingFaceTB/SmolLM2-135M-Instruct",
    "revision": "12fd25f77366fa6b3b4b768ec3050bf629380bac",
    "parameters": 134515008,
    "source_safetensors_sha256": "5af571cbf074e6d21a03528d2330792e532ca608f24ac70a143f6b369968ab8c",
    "converted_f16_gguf_sha256": "c953b337d61834a459884f51250e8242c7a4d2d7f74e465e84eb30e0e51d6d59"
  },
  "observations": {
    "training_loss_before": 1.94235477448,
    "training_loss_after": 1.9411190033,
    "import_max_logit_error": 0.000202178955078,
    "adapter_max_logit_error": 0.000185966491699,
    "merged_max_logit_error": 0.000185966491699,
    "generation": "The clear daytime sky",
    "checks": 442421
  },
  "validation": {
    "native_training_numeric_and_contract_checks": 110039,
    "native_artifact_checks": 609,
    "independent_tiny_cases": [
      "F32 GQA",
      "F16 GQA",
      "BF16 GQA",
      "tied embedding GQA"
    ],
    "hf_full_vocab_logit_cells": 147456,
    "hf_max_absolute_logit_error": 4.52995e-05,
    "held_out_quality_improvement_claimed": false
  },
  "scope": [
    "Native floating dense-Llama import, full/LoRA tuning, masked CSV/causal examples",
    "Immutable source, token-weighted accumulation, held-out validation",
    "Self-contained exact optimizer/data-cursor checkpoint and resume on qualified same-host profile",
    "Merged F32 GGUF and exact-base GGUF adapter deployment through llm.load",
    "F16/Q4_0 deployment quantization; existing Ollama/cloud clients retained"
  ],
  "limits": [
    "Pretrained qualification uses one 135M model and short LoRA updates, not 14B training",
    "Full tuning and quantization numerical tests also use nontrivial synthetic Llama/GQA fixtures",
    "No claim of general language-quality improvement or training speed",
    "Training is synchronous in interpreter; standalone compiler retains serial native inference",
    "No GPU training, QLoRA, volunteer network or distributed aggregation"
  ],
  "reproduce": {
    "fixtures": "python3 tools/create_modern_llm_fixture.py BUILD/modern-fixtures",
    "standard": "ctest --test-dir BUILD -R \"aner.modern\" --output-on-failure",
    "optional_oracles": "Set ANER_REFERENCE_ML_PYTHON to a local development interpreter with torch/numpy; no downloads",
    "real_pretrained": "Prepare pinned official source and offline F16 conversion; run aner_modern_pretrained_probe SOURCE.gguf NEW_OUTPUT, then tests/modern_pretrained_reference.py with local source/GGUF/report paths",
    "oracle_precision": "Strict direct-engine FP32 oracle disables flash attention because the pinned engine AUTO path casts attention K/V to F16. Public default inference is tested separately."
  }
}