gpt-oss-20b-Ja-Fin-Thinking

A Japanese financial domain reasoning model, built through supervised fine-tuning of gpt-oss-20b-Ja-Fin-CPT.

Model Overview

Trained to provide high-quality responses with explicit reasoning traces for Japanese financial domain tasks.

Base Model: gpt-oss-20b-Ja-Fin-CPT
Training Stage: Supervised Fine-Tuning (SFT)
Domain: Japanese Finance
Language: Japanese, English

Benchmark Results

japanese-lm-fin-harness

Model	Avg.	chabsa	cma	cpa	fp2	ss1
gpt-oss-20b (official)	66.93	91.80	90.46	38.51	49.74	64.15
gpt-oss-20b-Ja-Fin-Thinking (Ours)	72.50	91.89	94.24	45.51	62.71	68.15

+5.57 points improvement over the official instruction-tuned model.

pfmt-bench-fin-ja

Model	Avg.	turn1	turn2
gpt-oss-20b (official)	7.883	7.858	7.908
gpt-oss-20b-Ja-Fin-Thinking (Ours)	8.209	7.992	8.425

Training

Supervised Fine-Tuning

Fine-tuned on our synthetic instruction dataset with reasoning traces:

Dataset: nri-fin-reasoning + supplementary data
Total samples: ~1.44M
Total tokens: ~9.5B
Epochs: 2

Training Infrastructure:

Hardware: AWS p5en.48xlarge (NVIDIA H200 Tensor Core GPU x 8)
Training time: ~240 hours

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "nri-ai/gpt-oss-20b-Ja-Fin-Thinking"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

messages = [
    {"role": "user", "content": "分散投資のメリットとデメリットを説明してください。"}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=8192)

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True)
print(response)

Intended Use

Primary Use Cases

Financial question answering in Japanese
Financial document analysis and summarization
Financial reasoning and calculation tasks
Multi-turn financial advisory conversations

Out-of-Scope Uses

Production deployment without additional safety evaluation
Professional financial advice (this is a research model)
Non-financial domain applications

Limitations

Domain specificity: Optimized for Japanese financial domain; performance on other domains may vary
Synthetic training data: May contain hallucinations despite quality filtering
Language coverage: Primarily Japanese and English

Ethical Considerations

Financial information generated by this model should not be used as professional financial advice without review by qualified experts
Users should verify important financial information against authoritative sources and professional guidance before making decisions
The model may reflect biases present in training data

License

This model is released under the Apache 2.0 license.

Privacy Notice

For details on how personal information is handled, please see the Privacy Notice (日本語).

Citation

@inproceedings{okochiDomainSpecificLLM2026,
  author    = {大河内 悠磨 and Sim, Fabio Milentiansen and 岡田 智靖},
  title     = {ドメイン特化LLMの推論能力向上を目的とした合成指示データセットの構築と金融ドメインにおける評価},
  booktitle = {言語処理学会第32回年次大会 (NLP2026) },
  year      = {2026},
  month     = mar,
  address   = {Utsunomiya, Tochigi, Japan},
  publisher = {言語処理学会},
  note      = {Paper ID: C7-2},
  url       = {https://www.anlp.jp/proceedings/annual_meeting/2026/pdf_dir/C7-2.pdf}
}

@misc{okochi2026constructingsyntheticinstructiondatasets,
  title     = {Constructing Synthetic Instruction Datasets for Improving Reasoning in Domain-Specific LLMs: A Case Study in the Japanese Financial Domain},
  author    = {Yuma Okochi and Fabio Milentiansen Sim and Tomoyasu Okada},
  year      = {2026},
  eprint    = {2603.01353},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url       = {https://arxiv.org/abs/2603.01353}
}

Acknowledgments

This model was developed with the support of the "GENIAC (Generative AI Accelerator Challenge)" project, implemented by the Ministry of Economy, Trade and Industry (METI) and the New Energy and Industrial Technology Development Organization (NEDO), with the aim of strengthening Japan's development capabilities in generative AI.