Access ZEO Med 2

ZEO Med 2, including the LoRA adapter weights, is released under the Apache License 2.0. Commercial use, modification and redistribution are permitted under that license. Access is granted automatically once you share your contact information.

By requesting access you confirm that you have read the Apache License 2.0 and the medical-use limitations described in this model card.

Log in or Sign Up to review the conditions and access this model content.

ZeroOne AI

Open in AInode ZeroOne AI Evaluation Code License

1. Model Introduction

ZEO = Zero ยท Expert ยท One. ZEO is ZeroOne AI's model family that brings expert-level intelligence into the journey from zero to one. ZEO Med is the medical-specialized ZEO model, designed to deliver expert-level medical reasoning in secure enterprise and hospital environments.

ZEO Med 2 is a medical language model for Korean healthcare licensing examinations and English medical question answering. On the Korean doctor licensing exam it is state of the art among open-weight models measured under the same evaluation setup, and on the nurse, pharmacist and dentist exams and on MedQA-USMLE it stays in the top tier.

ZEO Med 2 was developed by ZeroOne AI together with Samsung Medical Center, under the Advanced GPU Utilization Support Program of the Ministry of Science and ICT, Republic of Korea. See Acknowledgements.

2. Model Summary

Model ZEO Med 2
Base model google/gemma-4-31B-it
Adaptation LoRA adapter (medical post-training)
Domain Medical question answering
Languages Korean, English
Answering Single chat-completion call; internal reasoning, final answer scored
Decoding temperature 0.7, top-p 0.95, 5 samples, majority vote
Max Tokens 8,192
Weights LoRA adapter included (open weights)
License Apache License 2.0

3. Evaluation Results

Every score below was measured under the same conditions. The model is never shown worked examples with their answers beforehand (0-shot), and each sample uses one prompt and one call in which the model reasons and then gives its final answer in the same response โ€” there is no second call asking for the answer. Each question is sampled five times and the most frequent answer is taken as the answer for that question. Blank cells were not measured under these conditions and are not filled in from other setups.

3.1 Open-weight models measured under the same setup

ZEO Med 2 leaderboard

Doctor, Nurse, Pharmacist and Dentist are the Korean national licensing examinations (KorMedMCQA). Every model in this table was measured by us under the contract final-sc5-medical-private-cot-final-answer-0shot-onecall-max8192-v1.

Model Doctor Nurse Pharmacist Dentist 4-exam average MedQA-USMLE
ZEO Med 2 97.70% (425/435) 96.36% (846/878) 94.69% (838/885) 87.05% (706/811) 93.95% 94.82% (1207/1273)
Gemma-4-31B-it 97.01% (422/435) 96.24% (845/878) 95.03% (841/885) 87.42% (709/811) 93.93% 94.74% (1206/1273)
Qwen3.6-27B 93.56% (407/435) 95.22% (836/878) 93.90% (831/885) 83.23% (675/811) 91.48% 94.11% (1198/1273)
Qwen3.6-35B-A3B 93.10% (405/435) 93.28% (819/878) 92.77% (821/885) 82.00% (665/811) 90.29% 94.34% (1201/1273)
Mistral Medium 3.5 ยถ 90.80% (395/435) 93.51% (821/878) 92.99% (823/885) 80.39% (652/811) 89.42% 91.99% (1171/1273)
HARI-Q2.5-Thinking * 89.20% 90.99% 90.94% 72.96% 86.02% 88.36%
gpt-oss-120b (native/default thinking) ยถ 86.67% (377/435) 89.07% (782/878) 89.15% (789/885) 75.83% (615/811) 85.18% 92.38% (1176/1273)
Nemotron 3 Nano 83.22% (362/435) 86.67% (761/878) 85.76% (759/885) 66.46% (539/811) 80.53% 88.45% (1126/1273)
Nemotron 3 Nano Omni 81.84% (356/435) 83.71% (735/878) 88.02% (779/885) 66.09% (536/811) 79.92% 83.03% (1057/1273)
MedGemma 1.0 72.41% (315/435) 78.70% (691/878) 75.37% (667/885) 60.79% (493/811) 71.82% 86.65% (1103/1273)
MedGemma 1.5 67.59% (294/435) 67.65% (594/878) 67.23% (595/885) 49.08% (398/811) 62.89% 72.43% (922/1273)
Footnotes
  • No mark: results we evaluated directly on open-weight/adapter models under the 0-shot SC@5 family of contracts. Per-model differences in runtime and parser are published in PROTOCOL.md.
  • * HARI: values published on its official model card. The four KorMedMCQA exams are 5-shot and MedQA-USMLE is 0-shot, and KorMed4 macro is the simple average of the published Doctor/Nurse/Pharmacist/Dentist scores.
  • ยถ Mistral ยท gpt-oss: reference values from our own evaluation, run with a different reasoning setting from the rest of the table. Mistral Medium 3.5 ran with explicit high reasoning. gpt-oss-120b ran with its native/default thinking setting, which is why it is not labelled gpt-oss-120b (high).

3.2 KMed.ai

KMed.ai reported an average of 96.4 points on the 2025 Korean national doctor licensing examination, in the official announcement by Seoul National University Hospital and NAVER. That is a score on the actual national examination, not on the KorMedMCQA benchmark used throughout this model card. The two are different examinations with different items, so the figure is not converted to %, no gap against ZEO Med 2 is derived from it, and it is not placed in the table above.

3.3 Frontier API models

GPT-5.2 and Gemini 3.1 Flash-Lite are closed models reachable only through an API, so they are listed separately from the open-weight models in the Section 3.1 table. We measured them ourselves through OpenRouter under the same evaluation conditions as the Section 3.1 table. Cells marked โ€” were not measured.

Model Doctor Nurse Pharmacist Dentist 4-exam average MedQA-USMLE
ZEO Med 2 97.70% (425/435) 96.36% (846/878) 94.69% (838/885) 87.05% (706/811) 93.95% 94.82% (1207/1273)
GPT-5.2 97.47% (424/435) โ€” โ€” โ€” โ€” 95.99% (1222/1273)
Gemini 3.1 Flash-Lite 96.09% (418/435) 96.47% 96.05% 90.26% 94.72% 94.42% (1202/1273)

The ZEO Med 2 row repeats the scores from the Section 3.1 table. ZEO Med 2 leads on the Korean doctor licensing examination. GPT-5.2 leads on MedQA-USMLE, and Gemini 3.1 Flash-Lite leads on the nurse, pharmacist and dentist examinations and on the four-exam average. This is a controlled comparison under matching evaluation conditions.

Gemini 3.1 Flash-Lite's nurse, pharmacist and dentist scores are recorded as percentages only, so no item counts are shown for those three cells.

4. Evaluation Protocol

Each question is answered in one chat-completion call. The model performs its reasoning inside that single call and then emits a final answer line; the scorer reads only that line. There is no separate rationale request and no second call, and hidden reasoning is neither scored nor released.

Prompting is template-driven. A system template fixes the answering format and a user template carries the question, the option list and the option labels through {question}, {choices} and {labels} placeholders. The exact templates used for these scores are published here: KorMedMCQA system ยท KorMedMCQA user ยท MedQA system ยท MedQA user. They hold placeholders only and carry no benchmark question text. Filled-in prompts are not published.

Five samples are taken per question and the answer is the one most of them agree on. If a sample runs out of tokens before it finishes, that is recorded and published together with the score instead of being hidden.

The full setup is written out in PROTOCOL.md.

5. Run the Model

ZEO Med 2 is a LoRA adapter over google/gemma-4-31B-it. Serving loads the base model and applies the adapter โ€” you do not merge weights. The adapter files are in this repository under evaluation-code/artifacts/adapter/ZEO-Med-31B-Adapter-v1/ (adapter_model.safetensors, adapter_config.json); access is gated with automatic approval (see Section 10, License).

Serve with vLLM

The served model id is ZEO-Med-31B-Adapter-v1 (the identifier used throughout the evaluation scripts); on Hugging Face the model is shown as ZEO Med 2.

python -m vllm.entrypoints.openai.api_server \
  --model google/gemma-4-31B-it \
  --enable-lora \
  --lora-modules ZEO-Med-31B-Adapter-v1=evaluation-code/artifacts/adapter/ZEO-Med-31B-Adapter-v1 \
  --max-lora-rank 8 \
  --dtype bfloat16 \
  --language-model-only \
  --reasoning-parser gemma4 \
  --max-model-len 16384

serve_vllm.sh (Section 6, Reproduce the Evaluation) wraps this exact command with the pinned base revision and the adapter hash check.

Chat

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
resp = client.chat.completions.create(
    model="ZEO-Med-31B-Adapter-v1",
    messages=[
        {"role": "user", "content": "๊ณ ํ˜ˆ์•• 1์ฐจ ์•ฝ์ œ ์„ ํƒ์˜ ์ผ๋ฐ˜์  ์›์น™์„ ์„ค๋ช…ํ•ด์ค˜."},
    ],
    temperature=0.7,
    top_p=0.95,
    max_tokens=8192,
)
print(resp.choices[0].message.content)

The model reasons internally within the single call and returns the final answer in content. The base model's tokenizer and chat template are used; the adapter does not replace them.

The model is for research and non-clinical use. See Sections 7 and 8.

6. Reproduce the Evaluation

This is separate from Section 5, Run the Model: here you re-run the exact 0-shot SC@5 contract to reproduce the published scores. Benchmark question text is not redistributed โ€” bring your own licensed copy of the datasets. Full step-by-step instructions, including the CPU-only package self-test, are in the evaluation-code guide.

run_full5_reproduction.sh does not start vLLM itself: you first serve the five task endpoints (with serve_vllm.sh zeo), then pass their URLs, PIDs, and the input/manifest directories as environment variables. It requires all of the following to be set:

export PACKAGE_ROOT=/path/to/evaluation-code
export MODEL_PACKAGE_ROOT=/path/to/evaluation-code
export SERVE_PY=/path/to/vllm-0.23.0/bin/python
export SERVED_MODEL=ZEO-Med-31B-Adapter-v1
export SERVING_MODE=dynamic_lora
export FULL5_INPUT_DIR=/path/to/full5-inputs        # your licensed benchmark rows
export FULL5_MANIFEST_DIR=/path/to/full5-manifests
export OUTPUT_ROOT=/path/to/new-full5-output
export NUM_WORKERS=48
export REQUEST_SEED=42
# one URL + PID per task, from the endpoints you started with serve_vllm.sh:
export DOCTOR_URL=... NURSE_URL=... PHARMACIST_URL=... DENTIST_URL=... MEDQA_URL=...
export DOCTOR_SERVE_PID=... NURSE_SERVE_PID=... PHARMACIST_SERVE_PID=... DENTIST_SERVE_PID=... MEDQA_SERVE_PID=...

bash evaluation-code/scripts/run_full5_reproduction.sh

A full-5 macro delta of zero and an all-task match indicate a successful reproduction.

7. Evaluation Datasets

The scores above were measured on the public test sets below. Because of the licenses of those test sets, no question text or dataset file is redistributed here; anyone reproducing the measurement obtains them under their own terms.

  • KorMedMCQA โ€” doctor, nurse, pharmacist and dentist configurations, test split, revision 79efd6f91edfc8036330d7a4daa88b9f2deb9a82. The dataset card states CC-BY-NC-2.0.
  • MedQA-USMLE-4-options โ€” test split, revision 0fb93dd23a7339b6dcd27e241cb9b5eca62d4d18. The dataset card states CC-BY-4.0.

Question text is excluded from this repository by publication policy. This is a release-control decision, not legal advice.

8. Intended Use and Limitations

ZEO Med 2 is released for research, education, evaluation, and product development. It is not a medical device and is not approved for autonomous diagnosis, prescription, treatment, or other independent clinical decision-making, and may not be used as the sole basis for a clinical decision. Any healthcare deployment must include qualified professional oversight, institution-specific validation, and compliance with applicable laws.

The published scores measure exam-style multiple-choice answer selection under one fixed setup and do not establish clinical safety or diagnostic accuracy.

9. Base Model and Modifications

ZEO Med 2 is a modified and fine-tuned derivative of:

  • Base model: google/gemma-4-31B-it (revision 3548789868c5356dbf307c98e6f609007b82b3eb)
  • Base model license: Apache License 2.0 (Developer: Google)
  • Modifications by ZeroOne AI: medical-domain LoRA fine-tuning, Korean/English medical instruction tuning, response alignment, and configuration changes

ZEO Med 2 is not affiliated with, endorsed by, or sponsored by Google. The full license text is in LICENSE; attribution and modification notices are in NOTICE.

10. License

ZEO Med 2 is released under the Apache License 2.0, and that includes the LoRA adapter weights. Commercial use, modification, and redistribution are permitted under that license.

The base model google/gemma-4-31B-it is under the same license. Attribution and the list of modifications made by ZeroOne AI are in NOTICE, which you must retain when redistributing the model or a derivative of it, as required by Section 4 of the license.

Access to the repository files is gated with automatic approval: you share your contact information once and access is granted immediately.

11. Contact

12. Acknowledgements

ZEO Med 2 was developed under the Advanced GPU Utilization Support Program of the Ministry of Science and ICT, Republic of Korea, project no. 02-26-01-0282, awarded to Samsung Medical Center.

We thank the research team of the Department of Emergency Medicine, Samsung Medical Center โ€” Prof. Won Chul Cha, principal investigator, and Prof. Meong Hi Son โ€” for the collaboration on this project.

13. Citation

@misc{zeo-med-2-2026,
    title={ZEO Med 2},
    author={ZeroOne AI},
    year={2026},
    url={https://huggingface.co/ZeroOneAI/ZEO-Med-2}
}

ํ•œ๊ตญ์–ด

ZeroOne AI

Open in AInode ZeroOne AI Evaluation Code License

English | ํ•œ๊ตญ์–ด

1. ๋ชจ๋ธ ์†Œ๊ฐœ

ZEO = Zero ยท Expert ยท One. ZEO๋Š” (์ฃผ)์ œ๋กœ์›์—์ด์•„์ด(ZeroOne AI)์˜ ์ž์ฒด AI ๋ชจ๋ธ ํŒจ๋ฐ€๋ฆฌ๋กœ, Zero์—์„œ One์œผ๋กœ ์ด์–ด์ง€๋Š” ๋ฌธ์ œ ํ•ด๊ฒฐ ๊ณผ์ •์— ๋„๋ฉ”์ธ ์ „๋ฌธ๊ฐ€ ์ˆ˜์ค€์˜ ์ง€๋Šฅ์„ ๊ฒฐํ•ฉํ•œ๋‹ค๋Š” ์˜๋ฏธ๋ฅผ ๋‹ด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ZEO Med๋Š” ์˜๋ฃŒ ๋ถ„์•ผ์— ํŠนํ™”๋œ ZEO ๋ชจ๋ธ๋กœ, ์•ˆ์ „ํ•œ ๊ธฐ์—… ๋ฐ ๋ณ‘์› ํ™˜๊ฒฝ์—์„œ ์ „๋ฌธ๊ฐ€ ์ˆ˜์ค€์˜ ์˜ํ•™์  ์ถ”๋ก ์„ ์ œ๊ณตํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

์ง€์˜ค๋ฉ”๋“œ2(ZEO Med 2)๋Š” ํ•œ๊ตญ ์˜๋ฃŒ ๊ตญ๊ฐ€๊ณ ์‹œ์™€ ์˜๋ฌธ ์˜๋ฃŒ ๋ฌธ์ œ ํ’€์ด๋ฅผ ์œ„ํ•œ ์˜๋ฃŒ ์–ธ์–ด ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์˜์‚ฌ ๊ตญ๊ฐ€๊ณ ์‹œ์—์„œ๋Š” ๋™์ผํ•œ ํ‰๊ฐ€ ํ™˜๊ฒฝ์œผ๋กœ ์ธก์ •ํ•œ open-weight ๋ชจ๋ธ ์ค‘ ์ตœ๊ณ  ์„ฑ๋Šฅ์ด๋ฉฐ, ๊ฐ„ํ˜ธ์‚ฌยท์•ฝ์‚ฌยท์น˜๊ณผ์˜์‚ฌ ๊ตญ๊ฐ€๊ณ ์‹œ์™€ MedQA-USMLE์—์„œ๋„ ์ตœ์ƒ์œ„๊ถŒ ์„ฑ์ ์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค.

ZEO Med 2๋Š” ๊ณผํ•™๊ธฐ์ˆ ์ •๋ณดํ†ต์‹ ๋ถ€ ์ฒจ๋‹จ GPU ํ™œ์šฉ ์ง€์› ์‚ฌ์—…์˜ ์ง€์›์„ ๋ฐ›์•„, (์ฃผ)์ œ๋กœ์›์—์ด์•„์ด(ZeroOne AI)๊ฐ€ ์‚ผ์„ฑ์„œ์šธ๋ณ‘์›๊ณผ ํ•จ๊ป˜ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๊ฐ์‚ฌ์˜ ๊ธ€์— ์žˆ์Šต๋‹ˆ๋‹ค.

2. ๋ชจ๋ธ ๊ฐœ์š”

๋ชจ๋ธ ์ง€์˜ค๋ฉ”๋“œ2(ZEO Med 2)
๊ธฐ๋ฐ˜ ๋ชจ๋ธ google/gemma-4-31B-it
์ ์‘ ๋ฐฉ์‹ LoRA ์–ด๋Œ‘ํ„ฐ (์˜๋ฃŒ ์‚ฌํ›„ํ•™์Šต)
๋ถ„์•ผ ์˜๋ฃŒ ๋ฌธ์ œ ํ’€์ด
์–ธ์–ด ํ•œ๊ตญ์–ด, ์˜์–ด
์‘๋‹ต ๋ฐฉ์‹ ํ•œ ๋ฒˆ์˜ ํ˜ธ์ถœ๋กœ ์ถ”๋ก ๊ณผ ์ตœ์ข… ๋‹ต์„ ํ•จ๊ป˜ ์ƒ์„ฑ, ์ตœ์ข… ๋‹ต๋งŒ ์ฑ„์ 
๋””์ฝ”๋”ฉ temperature 0.7, top-p 0.95, 5ํšŒ ์ƒ˜ํ”Œ๋ง ํ›„ ๋‹ค์ˆ˜๊ฒฐ
์ตœ๋Œ€ ํ† ํฐ 8,192
๊ฐ€์ค‘์น˜ LoRA ์–ด๋Œ‘ํ„ฐ ํฌํ•จ (์˜คํ”ˆ ์›จ์ดํŠธ)
๋ผ์ด์„ผ์Šค Apache License 2.0

3. ํ‰๊ฐ€ ๊ฒฐ๊ณผ

์•„๋ž˜ ์ ์ˆ˜๋Š” ๋ชจ๋‘ ๊ฐ™์€ ์กฐ๊ฑด์—์„œ ์ธก์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ์— ์˜ˆ์‹œ ๋ฌธํ•ญ๊ณผ ์ •๋‹ต์„ ๋ฏธ๋ฆฌ ๋ณด์—ฌ์ฃผ์ง€ ์•Š์œผ๋ฉฐ(0-shot), ํ•˜๋‚˜์˜ ํ”„๋กฌํ”„ํŠธ๋กœ ํ•œ ๋ฒˆ ํ˜ธ์ถœํ•ด ๋ชจ๋ธ์ด ์ถ”๋ก ์„ ๊ฑฐ์นœ ๋’ค ์ตœ์ข… ๋‹ต์„ ๊ฐ™์€ ์‘๋‹ต ์•ˆ์—์„œ ์ด์–ด์„œ ๋‚ด๋†“๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๋‹ต์„ ๋‹ค์‹œ ๋ฌป๋Š” ๋‘ ๋ฒˆ์งธ ํ˜ธ์ถœ์€ ์—†์Šต๋‹ˆ๋‹ค. ๋ฌธํ•ญ๋งˆ๋‹ค 5ํšŒ ์ƒ˜ํ”Œ๋งํ•ด ๊ฐ€์žฅ ๋งŽ์ด ๋‚˜์˜จ ๋‹ต์„ ๊ทธ ๋ฌธํ•ญ์˜ ๋‹ต์œผ๋กœ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ๋นˆ์นธ์€ ์ด ์กฐ๊ฑด์œผ๋กœ ์ธก์ •ํ•˜์ง€ ์•Š์€ ๊ฐ’์ด๋ฉฐ, ๋‹ค๋ฅธ ์กฐ๊ฑด์˜ ๊ฐ’์œผ๋กœ ์ฑ„์šฐ์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.

3.1 ๊ฐ™์€ ์กฐ๊ฑด์œผ๋กœ ์ธก์ •ํ•œ open-weight ๋ชจ๋ธ

ZEO Med 2 leaderboard

DoctorยทNurseยทPharmacistยทDentist๋Š” ํ•œ๊ตญ ์˜๋ฃŒ ๊ตญ๊ฐ€๊ณ ์‹œ(KorMedMCQA)์ž…๋‹ˆ๋‹ค. ์ด ํ‘œ์˜ ๋ชจ๋“  ๋ชจ๋ธ์€ ๋‹น์‚ฌ๊ฐ€ final-sc5-medical-private-cot-final-answer-0shot-onecall-max8192-v1 ๊ณ„์•ฝ์œผ๋กœ ์ง์ ‘ ์ธก์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.

๋ชจ๋ธ Doctor Nurse Pharmacist Dentist 4๊ฐœ ์ง๊ตฐ ํ‰๊ท  MedQA-USMLE
ZEO Med 2 97.70% (425/435) 96.36% (846/878) 94.69% (838/885) 87.05% (706/811) 93.95% 94.82% (1207/1273)
Gemma-4-31B-it 97.01% (422/435) 96.24% (845/878) 95.03% (841/885) 87.42% (709/811) 93.93% 94.74% (1206/1273)
Qwen3.6-27B 93.56% (407/435) 95.22% (836/878) 93.90% (831/885) 83.23% (675/811) 91.48% 94.11% (1198/1273)
Qwen3.6-35B-A3B 93.10% (405/435) 93.28% (819/878) 92.77% (821/885) 82.00% (665/811) 90.29% 94.34% (1201/1273)
Mistral Medium 3.5 ยถ 90.80% (395/435) 93.51% (821/878) 92.99% (823/885) 80.39% (652/811) 89.42% 91.99% (1171/1273)
HARI-Q2.5-Thinking * 89.20% 90.99% 90.94% 72.96% 86.02% 88.36%
gpt-oss-120b (native/default thinking) ยถ 86.67% (377/435) 89.07% (782/878) 89.15% (789/885) 75.83% (615/811) 85.18% 92.38% (1176/1273)
Nemotron 3 Nano 83.22% (362/435) 86.67% (761/878) 85.76% (759/885) 66.46% (539/811) 80.53% 88.45% (1126/1273)
Nemotron 3 Nano Omni 81.84% (356/435) 83.71% (735/878) 88.02% (779/885) 66.09% (536/811) 79.92% 83.03% (1057/1273)
MedGemma 1.0 72.41% (315/435) 78.70% (691/878) 75.37% (667/885) 60.79% (493/811) 71.82% 86.65% (1103/1273)
MedGemma 1.5 67.59% (294/435) 67.65% (594/878) 67.23% (595/885) 49.08% (398/811) 62.89% 72.43% (922/1273)
ํ‘œ์‹๊ณผ ๊ฐ์ฃผ
  • ํ‘œ์‹ ์—†์Œ: ๋‹น์‚ฌ๊ฐ€ 0-shot SC@5 ๊ณ„์—ด ๊ณ„์•ฝ์œผ๋กœ ์ง์ ‘ ํ‰๊ฐ€ํ•œ open-weight/adapter ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ๋ณ„ runtimeยทparser ์ฐจ์ด๋Š” PROTOCOL.md์—์„œ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค.
  • * HARI: ๊ณต์‹ ๋ชจ๋ธ์นด๋“œ ๊ณต๊ฐœ๊ฐ’์ž…๋‹ˆ๋‹ค. KorMedMCQA 4๊ฐœ ์‹œํ—˜์€ 5-shot, MedQA-USMLE๋Š” 0-shot์ด๋ฉฐ, KorMed4 macro๋Š” ๊ณต๊ฐœ๋œ Doctor/Nurse/Pharmacist/Dentist ์ ์ˆ˜์˜ ๋‹จ์ˆœ ํ‰๊ท ์ž…๋‹ˆ๋‹ค.
  • ยถ Mistralยทgpt-oss: ๋‹น์‚ฌ ์ง์ ‘ ํ‰๊ฐ€ ์ฐธ๊ณ ๊ฐ’์ด๋ฉฐ, ํ‘œ์˜ ๋‚˜๋จธ์ง€ ๋ชจ๋ธ๊ณผ ์ถ”๋ก  ์„ค์ •์ด ๋‹ค๋ฆ…๋‹ˆ๋‹ค. Mistral Medium 3.5๋Š” explicit high reasoning์œผ๋กœ, gpt-oss-120b๋Š” native/default thinking ์„ค์ •์œผ๋กœ ์ธก์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ ํ˜„์žฌ ๊ฐ’์„ gpt-oss-120b (high)๋กœ ํ‘œ๊ธฐํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.

3.2 KMed.ai

KMed.ai๋Š” ์„œ์šธ๋Œ€๋ณ‘์›ยท๋„ค์ด๋ฒ„ ๊ณต์‹ ๋ฐœํ‘œ์—์„œ 2025๋…„๋„ ์˜์‚ฌ ๊ตญ๊ฐ€๊ณ ์‹œ ํ‰๊ท  96.4์ ์„ ๊ณต๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์‹ค์ œ ๊ตญ๊ฐ€๊ณ ์‹œ์—์„œ ๋ฐ›์€ ์ ์ˆ˜์ด๋ฉฐ, ์ด ๋ชจ๋ธ์นด๋“œ ์ „๋ฐ˜์—์„œ ์“ฐ๋Š” KorMedMCQA ๋ฒค์น˜๋งˆํฌ ์ ์ˆ˜๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค. ๋‘ ์‹œํ—˜์€ ๋ฌธํ•ญ์ด ๋‹ค๋ฅธ ๋ณ„๊ฐœ์˜ ์‹œํ—˜์ด๋ฏ€๋กœ ์ด ๊ฐ’์„ %๋กœ ๋ฐ”๊พธ์ง€ ์•Š๊ณ , ZEO Med 2์™€์˜ ์ ์ˆ˜ ์ฐจ์ด๋„ ๊ณ„์‚ฐํ•˜์ง€ ์•Š์œผ๋ฉฐ, ์œ„ ํ‘œ์—๋„ ๋„ฃ์ง€ ์•Š์Šต๋‹ˆ๋‹ค.

3.3 ํ”„๋ก ํ‹ฐ์–ด API ๋ชจ๋ธ

GPT-5.2์™€ Gemini 3.1 Flash-Lite๋Š” API๋กœ๋งŒ ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๋Š” ๋น„๊ณต๊ฐœ ๋ชจ๋ธ์ด๋ผ, 3.1์ ˆ ํ‘œ์˜ open-weight ๋ชจ๋ธ๊ณผ ๋ถ„๋ฆฌํ•ด ์‹ฃ์Šต๋‹ˆ๋‹ค. ๋‹น์‚ฌ๊ฐ€ OpenRouter๋ฅผ ํ†ตํ•ด 3.1์ ˆ ํ‘œ์™€ ๊ฐ™์€ ํ‰๊ฐ€ ์กฐ๊ฑด์œผ๋กœ ์ง์ ‘ ์ธก์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. โ€”๋Š” ์ธก์ •ํ•˜์ง€ ์•Š์€ ๊ฐ’์ž…๋‹ˆ๋‹ค.

๋ชจ๋ธ Doctor Nurse Pharmacist Dentist 4๊ฐœ ์ง๊ตฐ ํ‰๊ท  MedQA-USMLE
ZEO Med 2 97.70% (425/435) 96.36% (846/878) 94.69% (838/885) 87.05% (706/811) 93.95% 94.82% (1207/1273)
GPT-5.2 97.47% (424/435) โ€” โ€” โ€” โ€” 95.99% (1222/1273)
Gemini 3.1 Flash-Lite 96.09% (418/435) 96.47% 96.05% 90.26% 94.72% 94.42% (1202/1273)

ZEO Med 2 ํ–‰์€ 3.1์ ˆ ํ‘œ์˜ ์ ์ˆ˜๋ฅผ ๋‹ค์‹œ ์˜ฎ๊ธด ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์˜์‚ฌ ๊ตญ๊ฐ€๊ณ ์‹œ๋Š” ZEO Med 2๊ฐ€, MedQA-USMLE๋Š” GPT-5.2๊ฐ€, ๊ฐ„ํ˜ธ์‚ฌยท์•ฝ์‚ฌยท์น˜๊ณผ์˜์‚ฌ ๊ตญ๊ฐ€๊ณ ์‹œ์™€ 4๊ฐœ ์ง๊ตฐ ํ‰๊ท ์€ Gemini 3.1 Flash-Lite๊ฐ€ ๊ฐ€์žฅ ๋†’์Šต๋‹ˆ๋‹ค. ํ‰๊ฐ€ ์กฐ๊ฑด์ด ์„œ๋กœ ๊ฐ™์€ ํ†ต์ œ๋œ ๋น„๊ต ์ง€ํ‘œ์ž…๋‹ˆ๋‹ค.

Gemini 3.1 Flash-Lite์˜ ๊ฐ„ํ˜ธ์‚ฌยท์•ฝ์‚ฌยท์น˜๊ณผ์˜์‚ฌ ์ ์ˆ˜๋Š” ๋ฐฑ๋ถ„์œจ๋กœ๋งŒ ๊ธฐ๋ก๋˜์–ด ์žˆ์–ด ๊ทธ ์„ธ ์นธ์—๋Š” ๋ฌธํ•ญ ์ˆ˜๋ฅผ ํ‘œ๊ธฐํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.

4. ํ‰๊ฐ€ ๋ฐฉ๋ฒ•

๋ฌธํ•ญ๋งˆ๋‹ค ํ•œ ๋ฒˆ์˜ chat-completion ํ˜ธ์ถœ๋กœ ๋‹ตํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์€ ๊ทธ ํ˜ธ์ถœ ์•ˆ์—์„œ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•œ ๋’ค ์ตœ์ข… ๋‹ต ํ•œ ์ค„์„ ๋‚ด๋†“๊ณ , ์ฑ„์ ๊ธฐ๋Š” ๊ทธ ์ค„๋งŒ ์ฝ์Šต๋‹ˆ๋‹ค. ๊ทผ๊ฑฐ๋ฅผ ๋”ฐ๋กœ ์š”์ฒญํ•˜๋Š” ์ ˆ์ฐจ๋‚˜ ๋‘ ๋ฒˆ์งธ ํ˜ธ์ถœ์€ ์—†์œผ๋ฉฐ, ๋‚ด๋ถ€ ์ถ”๋ก ์€ ์ฑ„์ ํ•˜์ง€๋„ ๊ณต๊ฐœํ•˜์ง€๋„ ์•Š์Šต๋‹ˆ๋‹ค.

ํ”„๋กฌํ”„ํŠธ๋Š” ํ…œํ”Œ๋ฆฟ์œผ๋กœ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. system ํ…œํ”Œ๋ฆฟ์ด ๋‹ต๋ณ€ ํ˜•์‹์„ ๊ณ ์ •ํ•˜๊ณ , user ํ…œํ”Œ๋ฆฟ์ด {question}, {choices}, {labels} ์ž๋ฆฌํ‘œ์‹œ์ž๋กœ ๋ฌธํ•ญ๊ณผ ์„ ํƒ์ง€, ์„ ํƒ์ง€ ๊ธฐํ˜ธ๋ฅผ ์ „๋‹ฌํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ์ˆ˜๋ฅผ ๋‚ผ ๋•Œ ์‚ฌ์šฉํ•œ ์‹ค์ œ ํ…œํ”Œ๋ฆฟ์„ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค: KorMedMCQA system ยท KorMedMCQA user ยท MedQA system ยท MedQA user. ์ž๋ฆฌํ‘œ์‹œ์ž๋งŒ ๋“ค์–ด ์žˆ๊ณ  ๋ฒค์น˜๋งˆํฌ ๋ฌธํ•ญ ์›๋ฌธ์€ ํฌํ•จํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ฐ’์ด ์ฑ„์›Œ์ง„ ํ”„๋กฌํ”„ํŠธ๋Š” ๊ณต๊ฐœํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.

๋ฌธํ•ญ๋งˆ๋‹ค 5ํšŒ๋ฅผ ์ƒ˜ํ”Œ๋งํ•ด ๊ฐ€์žฅ ๋งŽ์ด ๋‚˜์˜จ ๋‹ต์„ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ์–ด๋–ค ์ƒ˜ํ”Œ์ด ํ† ํฐ์„ ๋‹ค ์“ฐ๊ณ  ๋๋‚˜๋ฉด ๊ทธ ์‚ฌ์‹ค๋„ ์ ์ˆ˜์™€ ํ•จ๊ป˜ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค.

์ „์ฒด ์„ค์ •์€ PROTOCOL.md์— ์ •๋ฆฌ๋ผ ์žˆ์Šต๋‹ˆ๋‹ค.

5. ๋ชจ๋ธ ์‚ฌ์šฉ

ZEO Med 2๋Š” google/gemma-4-31B-it ์œ„์— ์–น๋Š” LoRA ์–ด๋Œ‘ํ„ฐ์ž…๋‹ˆ๋‹ค. ์„œ๋น™ ์‹œ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์„ ๋กœ๋“œํ•˜๊ณ  ์–ด๋Œ‘ํ„ฐ๋ฅผ ์ ์šฉํ•˜๋ฉฐ, ๊ฐ€์ค‘์น˜๋ฅผ ๋ณ‘ํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์–ด๋Œ‘ํ„ฐ ํŒŒ์ผ์€ ์ด ์ €์žฅ์†Œ์˜ evaluation-code/artifacts/adapter/ZEO-Med-31B-Adapter-v1/ (adapter_model.safetensors, adapter_config.json)์— ์žˆ๊ณ , ์ ‘๊ทผ์€ ์ž๋™ ์Šน์ธ ๋ฐฉ์‹์˜ gated์ž…๋‹ˆ๋‹ค(10์ ˆ ๋ผ์ด์„ผ์Šค ์ฐธ๊ณ ).

vLLM ์„œ๋น™

์„œ๋น™ ๋ชจ๋ธ id๋Š” ZEO-Med-31B-Adapter-v1์ž…๋‹ˆ๋‹ค(ํ‰๊ฐ€ ์Šคํฌ๋ฆฝํŠธ ์ „๋ฐ˜์—์„œ ์“ฐ๋Š” ์‹๋ณ„์ž). Hugging Face ํŽ˜์ด์ง€์—๋Š” ZEO Med 2๋กœ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค.

python -m vllm.entrypoints.openai.api_server \
  --model google/gemma-4-31B-it \
  --enable-lora \
  --lora-modules ZEO-Med-31B-Adapter-v1=evaluation-code/artifacts/adapter/ZEO-Med-31B-Adapter-v1 \
  --max-lora-rank 8 \
  --dtype bfloat16 \
  --language-model-only \
  --reasoning-parser gemma4 \
  --max-model-len 16384

serve_vllm.sh(6์ ˆ ํ‰๊ฐ€ ์žฌํ˜„)๊ฐ€ ๊ธฐ๋ฐ˜ revision ๊ณ ์ •๊ณผ ์–ด๋Œ‘ํ„ฐ ํ•ด์‹œ ๊ฒ€์ฆ๊นŒ์ง€ ํฌํ•จํ•ด ์œ„ ๋ช…๋ น์„ ๊ทธ๋Œ€๋กœ ๊ฐ์Œ‰๋‹ˆ๋‹ค.

๋Œ€ํ™”

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
resp = client.chat.completions.create(
    model="ZEO-Med-31B-Adapter-v1",
    messages=[
        {"role": "user", "content": "๊ณ ํ˜ˆ์•• 1์ฐจ ์•ฝ์ œ ์„ ํƒ์˜ ์ผ๋ฐ˜์  ์›์น™์„ ์„ค๋ช…ํ•ด์ค˜."},
    ],
    temperature=0.7,
    top_p=0.95,
    max_tokens=8192,
)
print(resp.choices[0].message.content)

๋ชจ๋ธ์€ ํ•œ ๋ฒˆ์˜ ํ˜ธ์ถœ ์•ˆ์—์„œ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ์ตœ์ข… ๋‹ต์„ content๋กœ ๋ฐ˜ํ™˜ํ•ฉ๋‹ˆ๋‹ค. ํ† ํฌ๋‚˜์ด์ €์™€ chat template์€ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์˜ ๊ฒƒ์„ ์‚ฌ์šฉํ•˜๋ฉฐ, ์–ด๋Œ‘ํ„ฐ๊ฐ€ ์ด๋ฅผ ๋ฐ”๊พธ์ง€ ์•Š์Šต๋‹ˆ๋‹ค.

์ด ๋ชจ๋ธ์€ ์—ฐ๊ตฌยท๋น„์ž„์ƒ ์šฉ๋„์ž…๋‹ˆ๋‹ค. 7์ ˆ ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ์…‹๊ณผ 8์ ˆ ์‚ฌ์šฉ ๋ชฉ์ ๊ณผ ํ•œ๊ณ„๋ฅผ ์ฐธ๊ณ ํ•˜์‹ญ์‹œ์˜ค.

6. ํ‰๊ฐ€ ์žฌํ˜„

5์ ˆ ๋ชจ๋ธ ์‚ฌ์šฉ๊ณผ ๋ณ„๊ฐœ์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ๋Š” ๊ณต๊ฐœ๋œ ์ ์ˆ˜๋ฅผ ์žฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด 0-shot SC@5 ๊ณ„์•ฝ์„ ๊ทธ๋Œ€๋กœ ๋‹ค์‹œ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋ฒค์น˜๋งˆํฌ ๋ฌธํ•ญ ์›๋ฌธ์€ ์žฌ๋ฐฐํฌํ•˜์ง€ ์•Š์œผ๋‹ˆ ๊ฐ์ž ๋ผ์ด์„ผ์Šค์— ๋”ฐ๋ผ ๋ฐ์ดํ„ฐ์…‹์„ ์ง์ ‘ ๋ฐ›์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค. CPU ์ „์šฉ ํŒจํ‚ค์ง€ self-test๋ฅผ ํฌํ•จํ•œ ๋‹จ๊ณ„๋ณ„ ์ƒ์„ธ ์ ˆ์ฐจ๋Š” ํ‰๊ฐ€ ์ฝ”๋“œ ์•ˆ๋‚ด์— ์žˆ์Šต๋‹ˆ๋‹ค.

run_full5_reproduction.sh๋Š” vLLM์„ ์ง์ ‘ ๋„์šฐ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋จผ์ € 5๊ฐœ ์‹œํ—˜ ์—”๋“œํฌ์ธํŠธ๋ฅผ ์„œ๋น™(serve_vllm.sh zeo)ํ•œ ๋’ค, ๊ทธ URLยทPID์™€ ์ž…๋ ฅ/๋งค๋‹ˆํŽ˜์ŠคํŠธ ๋””๋ ‰ํ„ฐ๋ฆฌ๋ฅผ ํ™˜๊ฒฝ๋ณ€์ˆ˜๋กœ ๋„˜๊ฒจ์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์„ ๋ชจ๋‘ ์„ค์ •ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.

export PACKAGE_ROOT=/path/to/evaluation-code
export MODEL_PACKAGE_ROOT=/path/to/evaluation-code
export SERVE_PY=/path/to/vllm-0.23.0/bin/python
export SERVED_MODEL=ZEO-Med-31B-Adapter-v1
export SERVING_MODE=dynamic_lora
export FULL5_INPUT_DIR=/path/to/full5-inputs        # ๊ฐ์ž ๋ผ์ด์„ผ์Šค๋กœ ๋ฐ›์€ ๋ฒค์น˜๋งˆํฌ ๋ฌธํ•ญ
export FULL5_MANIFEST_DIR=/path/to/full5-manifests
export OUTPUT_ROOT=/path/to/new-full5-output
export NUM_WORKERS=48
export REQUEST_SEED=42
# serve_vllm.sh๋กœ ๋„์šด ์—”๋“œํฌ์ธํŠธ์˜ ์‹œํ—˜๋ณ„ URL + PID:
export DOCTOR_URL=... NURSE_URL=... PHARMACIST_URL=... DENTIST_URL=... MEDQA_URL=...
export DOCTOR_SERVE_PID=... NURSE_SERVE_PID=... PHARMACIST_SERVE_PID=... DENTIST_SERVE_PID=... MEDQA_SERVE_PID=...

bash evaluation-code/scripts/run_full5_reproduction.sh

5๊ฐœ ์‹œํ—˜ macro delta๊ฐ€ 0์ด๊ณ  ๋ชจ๋“  ์‹œํ—˜์ด ์ผ์น˜ํ•˜๋ฉด ์žฌํ˜„์— ์„ฑ๊ณตํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.

7. ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ์…‹

์œ„ ์ ์ˆ˜๋Š” ์•„๋ž˜ ๊ณต๊ฐœ ํ…Œ์ŠคํŠธ์…‹์œผ๋กœ ์ธก์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ํ…Œ์ŠคํŠธ์…‹์˜ ๋ผ์ด์„ผ์Šค ๋•Œ๋ฌธ์— ๋ฌธํ•ญ ์›๋ฌธ์ด๋‚˜ ๋ฐ์ดํ„ฐ์…‹ ํŒŒ์ผ์€ ์ด ์ €์žฅ์†Œ์— ์žฌ๋ฐฐํฌํ•˜์ง€ ์•Š์œผ๋ฉฐ, ์žฌํ˜„ํ•˜๋ ค๋Š” ์ชฝ์—์„œ ๊ฐ์ž์˜ ์ด์šฉ ์กฐ๊ฑด์— ๋”ฐ๋ผ ์ง์ ‘ ๋ฐ›์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค.

  • KorMedMCQA โ€” doctor, nurse, pharmacist, dentist ๊ตฌ์„ฑ, test split, revision 79efd6f91edfc8036330d7a4daa88b9f2deb9a82. ๋ฐ์ดํ„ฐ์…‹ ์นด๋“œ ํ‘œ๊ธฐ๋Š” CC-BY-NC-2.0์ž…๋‹ˆ๋‹ค.
  • MedQA-USMLE-4-options โ€” test split, revision 0fb93dd23a7339b6dcd27e241cb9b5eca62d4d18. ๋ฐ์ดํ„ฐ์…‹ ์นด๋“œ ํ‘œ๊ธฐ๋Š” CC-BY-4.0์ž…๋‹ˆ๋‹ค.

๋ฌธํ•ญ ์›๋ฌธ์€ ๊ณต๊ฐœ ์ •์ฑ…์ƒ ์ด ์ €์žฅ์†Œ์—์„œ ์ œ์™ธํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ฐฐํฌ ํ†ต์ œ ๊ฒฐ์ •์ด๋ฉฐ ๋ฒ•๋ฅ  ์ž๋ฌธ์ด ์•„๋‹™๋‹ˆ๋‹ค.

8. ์‚ฌ์šฉ ๋ชฉ์ ๊ณผ ํ•œ๊ณ„

ZEO Med 2๋Š” ์—ฐ๊ตฌยท๊ต์œกยทํ‰๊ฐ€ยท์ œํ’ˆ ๊ฐœ๋ฐœ์„ ์œ„ํ•ด ๊ณต๊ฐœํ•œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์˜๋ฃŒ๊ธฐ๊ธฐ๊ฐ€ ์•„๋‹ˆ๋ฉฐ ์ž์œจ์  ์ง„๋‹จยท์ฒ˜๋ฐฉยท์น˜๋ฃŒ ๋“ฑ ๋…๋ฆฝ์  ์ž„์ƒ ์˜์‚ฌ๊ฒฐ์ •์šฉ์œผ๋กœ ์Šน์ธ๋˜์ง€ ์•Š์•˜๊ณ , ์ž„์ƒ ํŒ๋‹จ์˜ ์œ ์ผํ•œ ๊ทผ๊ฑฐ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์˜๋ฃŒ ํ˜„์žฅ ๋„์ž… ์‹œ์—๋Š” ์ž๊ฒฉ ์žˆ๋Š” ์ „๋ฌธ๊ฐ€์˜ ๊ฐ๋…, ๊ธฐ๊ด€๋ณ„ ๊ฒ€์ฆ, ๊ด€๋ จ ๋ฒ•๊ทœ ์ค€์ˆ˜๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.

๊ณต๊ฐœ๋œ ์ ์ˆ˜๋Š” ๊ณ ์ •๋œ ์กฐ๊ฑด์—์„œ์˜ ์‹œํ—˜ํ˜• ๊ฐ๊ด€์‹ ์ •๋‹ต ์„ ํƒ์„ ์ธก์ •ํ•œ ๊ฒƒ์œผ๋กœ, ์ž„์ƒ์  ์•ˆ์ „์„ฑ์ด๋‚˜ ์ง„๋‹จ ์ •ํ™•๋„๋ฅผ ์ž…์ฆํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.

9. ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๊ณผ ์ˆ˜์ • ์‚ฌํ•ญ

ZEO Med 2๋Š” ๋‹ค์Œ์„ ์ˆ˜์ •ยทํŒŒ์ธํŠœ๋‹ํ•œ ํŒŒ์ƒ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.

  • ๊ธฐ๋ฐ˜ ๋ชจ๋ธ: google/gemma-4-31B-it (revision 3548789868c5356dbf307c98e6f609007b82b3eb)
  • ๊ธฐ๋ฐ˜ ๋ชจ๋ธ ๋ผ์ด์„ผ์Šค: Apache License 2.0 (๊ฐœ๋ฐœ: Google)
  • (์ฃผ)์ œ๋กœ์›์—์ด์•„์ด(ZeroOne AI)์˜ ์ˆ˜์ •: ์˜๋ฃŒ ๋„๋ฉ”์ธ LoRA ํŒŒ์ธํŠœ๋‹, ํ•œ๊ตญ์–ดยท์˜์–ด ์˜๋ฃŒ instruction ํŠœ๋‹, ์‘๋‹ต ์ •๋ ฌ, ๊ตฌ์„ฑ ๋ณ€๊ฒฝ

ZEO Med 2๋Š” Google๊ณผ ์ œํœดยท๋ณด์ฆยทํ›„์› ๊ด€๊ณ„๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค. ๋ผ์ด์„ผ์Šค ์ „๋ฌธ์€ LICENSE์—, ๊ท€์†ยท์ˆ˜์ • ๊ณ ์ง€๋Š” NOTICE์— ์žˆ์Šต๋‹ˆ๋‹ค.

10. ๋ผ์ด์„ผ์Šค

ZEO Med 2๋Š” LoRA ์–ด๋Œ‘ํ„ฐ ๊ฐ€์ค‘์น˜๋ฅผ ํฌํ•จํ•ด Apache License 2.0์œผ๋กœ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ผ์ด์„ผ์Šค์— ๋”ฐ๋ผ ์ƒ์—…์  ์ด์šฉ, ์ˆ˜์ •, ์žฌ๋ฐฐํฌ๊ฐ€ ๋ชจ๋‘ ํ—ˆ์šฉ๋ฉ๋‹ˆ๋‹ค.

๊ธฐ๋ฐ˜ ๋ชจ๋ธ google/gemma-4-31B-it๋„ ๊ฐ™์€ ๋ผ์ด์„ผ์Šค์ž…๋‹ˆ๋‹ค. ๊ท€์† ํ‘œ๊ธฐ์™€ (์ฃผ)์ œ๋กœ์›์—์ด์•„์ด(ZeroOne AI)๊ฐ€ ๊ฐ€ํ•œ ์ˆ˜์ • ๋ชฉ๋ก์€ NOTICE์— ์žˆ์œผ๋ฉฐ, ๋ผ์ด์„ผ์Šค ์ œ4์กฐ์— ๋”ฐ๋ผ ์ด ๋ชจ๋ธ์ด๋‚˜ ํŒŒ์ƒ ๋ชจ๋ธ์„ ์žฌ๋ฐฐํฌํ•  ๋•Œ ํ•จ๊ป˜ ์œ ์ง€ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.

์ €์žฅ์†Œ ํŒŒ์ผ ์ ‘๊ทผ์€ ์ž๋™ ์Šน์ธ ๋ฐฉ์‹์˜ gated์ž…๋‹ˆ๋‹ค. ์—ฐ๋ฝ์ฒ˜๋ฅผ ํ•œ ๋ฒˆ ์ œ์ถœํ•˜๋ฉด ์ฆ‰์‹œ ์ ‘๊ทผ ๊ถŒํ•œ์ด ๋ถ€์—ฌ๋ฉ๋‹ˆ๋‹ค.

11. ๋ฌธ์˜

12. ๊ฐ์‚ฌ์˜ ๊ธ€

ZEO Med 2๋Š” ๊ณผํ•™๊ธฐ์ˆ ์ •๋ณดํ†ต์‹ ๋ถ€ ์ฒจ๋‹จ GPU ํ™œ์šฉ ์ง€์› ์‚ฌ์—…(๊ณผ์ œ๋ฒˆํ˜ธ 02-26-01-0282, ์ฃผ๊ด€ ์‚ผ์„ฑ์„œ์šธ๋ณ‘์›)์˜ ์ง€์›์„ ๋ฐ›์•„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค.

ํ•จ๊ป˜ ์—ฐ๊ตฌ๋ฅผ ์ง„ํ–‰ํ•ด ์ฃผ์‹  ์‚ผ์„ฑ์„œ์šธ๋ณ‘์› ์‘๊ธ‰์˜ํ•™๊ณผ ์ฐจ์›์ฒ  ๊ต์ˆ˜(์—ฐ๊ตฌ์ฑ…์ž„์ž)์™€ ์†๋ช…ํฌ ๊ต์ˆ˜ ์—ฐ๊ตฌํŒ€์— ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.

13. Citation

@misc{zeo-med-2-2026,
    title={ZEO Med 2},
    author={ZeroOne AI},
    year={2026},
    url={https://huggingface.co/ZeroOneAI/ZEO-Med-2}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ZeroOneAI/ZEO-Med-2

Adapter
(255)
this model