05. Data Model & Storage¶

Artifact flow¶

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

  P["data/sft/prompts.jsonl"] --> R["data/raw/{name}/traces.jsonl"]
  R --> C["data/compressed/{name}/traces.jsonl"]
  C --> V["data/validated/{name}/traces.jsonl"]
  V --> T["data/train.jsonl + valid.jsonl"]:::highlight
  T --> A["adapters/{name}/{timestamp}/"]:::highlight
  A --> E["results/{name}/gsm8k*.json"]
  E --> Rep["report/ committed plots + REPORT.md"]

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

Schemas (logical)¶

Artifact	Typical fields
Prompt	id, source, prompt, choices, ground_truth
Trace	prompt id, thinking, answer, correctness metadata
SFT row	`{"text": "..."}` chat-formatted string
Adapter run	safetensors, `metrics.json`, loss plot
Eval JSON	per-example predictions + aggregate metrics

What is committed vs external¶

In git	Gitignored / HF
`report/` plots, metrics, REPORT.md	`data/`, `adapters/`, `results/`, `data-and-models/`
Scripts + configs	Large weights and full JSONL

HF dataset/model repo: hari31416/qwen-grug-finetune via sync_hf.py.

Sources and eval¶

Train mix: StrategyQA, LogiQA, BoolQ, ANLI, PIQA, ReClor (sizes in config.yaml).

Leakage control: block GSM8K test (and ARC in sampler) with fuzzy Jaccard matching (sample_sft_prompts.py).

Eval in code: GSM8K only.

Interview Q&A¶

Q: Where is the database?
A: There isn't one — the experiment is file-backed JSONL plus adapter directories, with HF as the distribution store.