05. Data Model & Storage¶

Artifact locations¶

Location	Contents
`scratch/` (gitignored)	Local ONNX bundles during build
Hugging Face Hub	Published `-ONNX` / `-ONNX-{fp16,int8,q4f16}` repos
`fixtures/*-golden.jsonl`	1100 × 3 directions = 3300 golden lines
`fixtures/parity-report-*.json`	FP32 parity
`fixtures/benchmark-*.json`	Quantization benches + mismatches
`fixtures/live-browser-benchmarks.json`	Browser-lab live runs

Bundle file set¶

graph TD
  classDef default fill:#1e293b,stroke:#38bdf8,stroke-width:2px,color:#f8fafc
  classDef highlight fill:#065f46,stroke:#34d399,stroke-width:2px,color:#f0fdf4

  Enc["encoder_model.onnx + .data"]
  Dec["decoder_model.onnx"]
  Past["decoder_with_past_model.onnx"]
  Shared["decoder_shared.onnx.data"]:::highlight
  Tok["tokenizer_src/tgt/meta.json"]

  Enc --> Infer["ORT sessions"]
  Dec --> Infer
  Past --> Infer
  Shared --> Dec
  Shared --> Past
  Tok --> Infer

  linkStyle default stroke:#64748b,stroke-width:2px

Consumer path¶

Export → HF → browser-demo / local-voice-chat on-demand download (Cache Storage in demo).

Interview Q&A¶

Q: Why commit fixtures?
A: Reproducible parity/benchmark claims without re-running multi-hour exports.