# Tokenizer Preprocessing

Tokenizer training should sample across files instead of consuming the first
large file only. Use round-robin sampling from checkpoint shards to avoid
language and domain skew.

## Keep

- Exact indentation.
- Blank lines.
- Comments and docstrings.
- Mixed Korean/English comments.
- Repository/file/language metadata when represented in text.

## Drop Or Quarantine

- Empty text.
- Null bytes.
- Obvious credentials, tokens, private keys, and secrets.
- Malformed JSONL.
- Extremely long records that exceed the active model context unless the loader
  has deterministic chunking.