| # 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. |
|
|
