llm-jp-4.1-thinking-sft-data

Overview

This dataset is a supervised fine-tuning (SFT) dataset used to train llm-jp-4.1-*-thinking models.

This dataset is constructed from prompts and conversations collected from multiple data sources. For most subsets, reasoning processes and final responses used for LLM-jp-4.1 SFT were generated or augmented using gpt-oss-120b. For the tool-calling and agentic data derived from NVIDIA Nemotron datasets, the original conversations and trajectories were generated using multiple language models. See the "Nemotron Agentic Mix" section below for details.

The splits reasoning_low, reasoning_medium, and reasoning_high correspond to different reasoning effort settings used when generating responses with gpt-oss-120b.

To support the continued development of LLM-jp, we would greatly appreciate it if you could share how you utilize LLM-jp outcomes via the survey form.

Data Sources and Licenses

llm-jp-4.1-*-thinking are constructed based on the following data sources. Since each dataset has its own license, please ensure compliance with the respective licenses when using this dataset. Some subsets cannot be redistributed due to licensing restrictions and are therefore not included.

Dataset	License	Notes
ac_002	answer-carefully-dataset-tou	Not included due to redistribution restrictions
daring_anteater	CC BY 4.0	-
flan	ODC-BY	-
ichikara	-	Requires paid license, not included
jaster_v1.4.1	CC BY-SA 4.0, CC BY-SA 3.0, CC BY 4.0, BSD-3, Apache-2.0, MIT	-
llmjp_extraction_wiki_ja_v0.x	Apache-2.0	-
llmjp_magpie_sft_v1.0	Apache-2.0	-
llmc_math_dataset	-	To be released
logical_math_coding_wizard8x22b	Apache-2.0	-
multiturn_calm3	Apache-2.0, CC BY-SA 3.0, CC0, CC BY 4.0	-
nemotron_post_v2_stem	CC BY 4.0	-
nemotron_post_v3_chat	CC BY 4.0	-
nemotron_post_v3_if	CC BY 4.0	-
nemotron_post_v3_math	CC BY-SA 4.0	-
random_to_fixed_multiturn_calm3	Apache-2.0	-
synthetic_if_ja	-	To be released
self_system_question	CC BY 4.0	Model-related questions
synthetic_jp_en_coding	Apache-2.0	-
system_prompt_question	CC BY 4.0	Model identity and metadata
nemotron3_sft_chat_v3	CC BY 4.0, ODC-BY	-
nemotron3_sft_science_vendor	CC BY-SA 4.0	-
table_gpt_train_large	MIT	-
nemotron3_sft_multilingual_v2_math_ko_stackoverflow	CC BY-SA 4.0	-
nemotron3_sft_multilingual_v2_math_ja_stackoverflow	CC BY-SA 4.0	-
nemotron_agentic_mix_v0.1.1	CC BY 4.0	Derived from NVIDIA Nemotron-Agentic-v1 and NVIDIA Nemotron-SFT-Agentic-v2
nemotron_agentic_mix_v0.1.1_ja	CC BY 4.0	Japanese translation of nemotron_agentic_mix_v0.1.1 generated using gpt-oss-120b

Nemotron Agentic Mix

nemotron_agentic_mix_v0.1.1 is constructed from the following NVIDIA datasets:

• Nemotron-Agentic-v1
• Nemotron-SFT-Agentic-v2

The source data include agentic conversations generated using multiple language models.

For Nemotron-Agentic-v1, the interactive_agent subset was generated using Qwen3-235B-A22B-Thinking-2507, Qwen3-32B, gpt-oss-120b, and Qwen3-235B-A22B-Instruct-2507. The tool_calling subset was generated using Qwen3-235B-A22B-Thinking-2507 and Qwen3-235B-A22B-Instruct-2507.

For Nemotron-SFT-Agentic-v2, Qwen3-235B-A22B-Thinking-2507, Qwen3-32B, Qwen3-235B-A22B-Instruct-2507, deepseek-r1-0528, DeepSeek-V3.2, and gpt-oss-120b.

nemotron_agentic_mix_v0.1.1_ja is the Japanese version of this data, translated using gpt-oss-120b.

Data Format

Each sample has the following format:

{
  "ID": "...",
  "messages": [
    {"role": "system", "name": null, "content": ["..."]},
    {"role": "user", "name": null, "content": ["..."]},
    {"role": "assistant", "name": null, "content": ["..."], "channel": "analysis"},
    {"role": "assistant", "name": null, "content": ["..."], "channel": "final"},
    ...
  ]
}

• channel="analysis" represents the reasoning process (chain-of-thought)
• channel="final" represents the final response

The messages field contains conversational data and is mostly compatible with openai-harmony.

Starting with LLM-jp-4.1, we use a modified version of the Harmony format as the chat template. As a result, compatibility with openai-harmony is no longer guaranteed, particularly for tool-calling datasets such as nemotron_agentic_mix_v0.1.1 and nemotron_agentic_mix_v0.1.1_ja.

Notes

• This dataset contains reasoning processes and responses generated by a large language model (gpt-oss-120b).
• Generated content may include errors, inaccuracies, biases, or harmful outputs.

Send Questions to

llm-jp(at)nii.ac.jp

Model Card Authors

The names are listed in alphabetical order.

Hirokazu Kiyomaru, Takashi Kodama, and Yunang Wu.