mybigday inspect-dataset
Never train on a dataset you haven't audited. Looking at the data is the single most reliable way to avoid format mismatches, KeyError-on-step-1 failures, and silently-bad runs.
When to fire
- Before any
train_sft.py/train_dpo.py/train_grpo.pyinvocation. - When the user says "use dataset X" — audit X immediately.
- When a training job fails with
KeyError,ValueError: column not found, or bad samples.
What to run
python scripts/inspect_dataset.py <repo_id> # default audit
python scripts/inspect_dataset.py <repo_id> --split train --n 5
python scripts/inspect_dataset.py <repo_id> --config wikitext-2-raw-v1
The script prints schema, row counts per split, length statistics on text /
prompt / messages columns, duplicate counts, label distribution, and n sample
rows.
What to verify against the chosen training method
| Method | Required columns |
|---|---|
| SFT | messages (chat) or text or prompt+completion |
| DPO | prompt, chosen, rejected |
| GRPO | prompt (+ a reward function) |
If the columns don't match the method, stop. Tell the user; don't silently remap or substitute the dataset.
Things to surface to the user
- Class imbalance, missing values, duplicate rows.
- Sequence-length p95 — drives the
max_seq_lengthchoice. - Dataset license / gating.
- Any unexpected nesting (e.g.
messagesalready wrapped in a list of dicts vs. flat text).