tokenizer/PREPROCESSING.md

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.