Module export 

Deterministic JSON Lines export and simple train / test row splits (Phase 2).

This module does not implement tokenizers or model-specific chat templates. Callers align exported text with their trainer’s expected fields.

Functions

dataset_to_jsonl

Serialize each row as one JSON object per line (UTF-8), columns in column_order.

filter_rows_max_utf8_chars

Keep only rows whose UTF-8 value in column has at most max_chars Unicode scalars; other rows dropped.

train_test_row_indices

Deterministic split: first train_count rows are train, remaining rows are test, where train_count = row_count - round(row_count * test_fraction.clamp(0..=1)).