Captured by the VN Studio dataset-export pipeline. Each row is byte-identical to what the corresponding stage's LLM receives at inference time, so a model fine-tuned on these rows is a drop-in replacement for the pipeline's LLM.

Row counts (4349 total)

localizer-rewrite: 4349 rows

Layout

/ train.sharegpt.jsonl # ShareGPT chat format train.instruction.jsonl # flat prompt/completion format… See the full description on the dataset page: https://huggingface.co/datasets/SubTL-AI/subtl-dsg1-en.