Full fine-tuned variant of openai/gpt-oss-20b on a Korean, domain-specific corpus covering the Zeroin fund-evaluation methodology (fund classification principles, etc.).

This release is distributed under CC BY-NC 4.0 — free for non-commercial research, evaluation, and educational use. See LICENSE and NOTICE.

Model summary

Field	Value
Base model	`openai/gpt-oss-20b` (Apache 2.0)
Architecture	`GptOssForCausalLM` (MoE, 32 experts, 4 active)
Parameters	~21B (full fine-tune)
Precision	bfloat16
Context length	131072 (base) / 4096 (training)
Chat template	Harmony response format (preserved from base)
Training framework	transformers + TRL + DeepSpeed ZeRO-3
Training hardware	3x NVIDIA H200
Training steps	1000
Max length	4096
Per-device batch size	1 (effective batch = 24 with grad accum 8)
Learning rate	1e-5 (cosine, warmup)
Final training loss	~0.0025

Intended use

Korean-language Q&A and reasoning about the Zeroin fund-classification methodology (Sufficiency / Comparability / Continuity principles, etc.).
Research, documentation generation, and agent workflows in non-commercial contexts.

Out-of-scope use

Commercial deployments of any kind (prohibited by CC BY-NC 4.0).
Public-facing deployments without additional safety filtering.
Medical, legal, financial advice to end-users.
Any use not permitted by the base model's gpt-oss usage policy.

Usage (transformers)

python

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "KG-ZEROIN/gpt-oss-20b-zeroin-v8-r4"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "user", "content": "<your Korean prompt here, e.g. a question about Zeroin fund-evaluation methodology>"}
]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

output = model.generate(
    inputs,
    max_new_tokens=1024,
    do_sample=False,
)
print(tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=False))

Training data

The training corpus is published as a companion dataset: KG-ZEROIN/zeroin-v8-r4-dataset (CC BY-NC 4.0). The corpus is derived from the Zeroin fund-evaluation methodology Korean domain document and is organized as a Harmony-ready chat-messages JSONL.

Limitations

The model may reproduce fragments of the training data when prompted with closely matching questions.
Specialized to Korean financial / fund-classification vocabulary; general-domain performance is close to the base model's behavior but has not been independently re-evaluated.
Reasoning traces in the Harmony analysis channel are NOT intended for end-user display.

License and attribution

Base model components: openai/gpt-oss-20b, Apache License 2.0, Copyright 2025 OpenAI. Apache 2.0 attribution is preserved in NOTICE.
Use is additionally governed by OpenAI's gpt-oss usage policy.

See LICENSE (CC BY-NC 4.0) and NOTICE (upstream attribution) in this repository.

Citation (base model)

bibtex

@misc{openai2025gptoss,
  title={gpt-oss-120b & gpt-oss-20b Model Card},
  author={OpenAI},
  year={2025},
  eprint={2508.10925},
  archivePrefix={arXiv}
}

사사 (Acknowledgment)

사사번호 (Project ID): RQT-25-090024

이 성과는 정부(과학기술정보통신부)의 재원으로 AI컴퓨팅자원활용기반강화(GPU 임차 지원) 사용자 지원사업의 지원을 받아 수행된 연구임.

This research was supported by the AI Computing Infrastructure Enhancement (GPU Rental Support) User Support Program funded by the Ministry of Science and ICT (MSIT), Republic of Korea.