Dataset
Nemotron-SFT-Long-Context-Translation-PG19-v1 is a supervised fine-tuning dataset from NVIDIA containing 4,335 English-to-multilingual translation examples. English source texts are books or retained book excerpts from the PG19 training set. Target translations were generated with DeepSeek V4 Flash into German, Spanish, French, Italian, Japanese, and Chinese.
Each example pairs a translation instruction and long English source text with its model-generated translation. Instructions request preservation of paragraph structure, document order, and approximate sentence alignment, with only the translated text in the response. Multiple examples may share the same source book.
| Translation direction | Examples |
|---|---|
English → German (de_DE) | 728 |
English → Spanish (es_MX) | 766 |
English → French (fr_FR) | 513 |
English → Italian (it_IT) | 757 |
English → Japanese (ja_JP) | 785 |
English → Chinese (zh_CN) | 786 |
| Total | 4,335 |
Locale labels describe the generation targets and do not independently establish dialect fidelity.
The dataset is provided as data/train.jsonl, exposed as the train split. The file contains 4,335 records and occupies approximately 2.22 GB uncompressed.
Each JSONL record contains:
messages: one user message containing the translation instruction and English source text, followed by one assistant message containing the translation. Each message has role and content fields.metadata: uuid, dataset, document_id, source_language, target_language, target_language_name, source_file, and target_file.The dataset value is pg19-long-translation. All records retain a nonempty source document identifier in metadata.document_id. Individual generation chunks are concatenated before creating document-level SFT records.
There are no system messages, tool calls, or separate reasoning-content fields.
To load the JSONL file after downloading it:
from datasets import load_dataset
dataset = load_dataset("json", data_files="data/train.jsonl", split="train")
The dataset supports supervised fine-tuning for long-document translation and multilingual long-context capabilities, including document-structure preservation and consistency of terminology.
Some source texts are retained prefixes rather than complete books. Translations are synthetic and may contain errors, omissions, or formatting differences. Human translation review and independent semantic-quality assessment have not been established. Source books and generated material can recur across records; record identifiers do not establish content deduplication. This dataset is training-derived and is not an independent held-out evaluation benchmark.
Historical source texts may contain dated language and biases. Developers should assess suitability for their intended use case.
NVIDIA's contributions are provided under the OpenMDW License Agreement, version 1.1 (OpenMDW-1.1). The source corpus is derived from the PG-19 Language Modelling Benchmark, provided by DeepMind under the Apache License, Version 2.0, using the emozilla/pg19 distribution. PG19 contains historical books from Project Gutenberg published before 1919. The applicable upstream Apache-2.0 terms remain in effect and are not superseded by OpenMDW-1.1.
NVIDIA's modifications include selecting PG19 training-set texts, retaining book excerpts where applicable, generating translations with DeepSeek V4 Flash, formatting source/translation pairs as SFT records, and filtering records. See NOTICE for attribution and the upstream reference.
The underlying English source works are public domain in the United States. Copyright status may differ in other jurisdictions; users should check applicable local law before use or redistribution. This statement concerns the underlying source works, not the licensing of the PG19 dataset or NVIDIA's contributions.
Please report quality, risk, security vulnerabilities, or NVIDIA AI concerns through NVIDIA's reporting channel.