HyperCLOVA X SEED Think-14B — AI Hub 71875 Base-wrong Correction Minimal SFT
This repository contains a standalone BF16 model derived from naver-hyperclovax/HyperCLOVAX-SEED-Think-14B. A deliberately minimal LoRA adapter was trained on the Base-wrong correction pool from AI Hub dataset 71875 and merged into the pristine base weights.
Model details
- Base model: naver-hyperclovax/HyperCLOVAX-SEED-Think-14B
- Base revision: 9b74e35d4c7e4ffec489f4171273caca8948a2b9
- Architecture: HyperCLOVAXForCausalLM
- Weight format: BF16 safetensors, standalone merged full model
- Chat template: official HyperCLOVA X template, preserved byte-for-byte
- LoRA: rank 4, alpha 8, dropout 0, bias none
- Target modules: q_proj, v_proj
- Objective: assistant-token-only causal-language-model cross entropy
- Learning rate: 1e-6; scheduler: constant; warmup: 0; weight decay: 0
- Training: 606 examples, 1 epoch, 38 optimizer steps, effective batch size 16
- Per-device batch: 1; gradient accumulation: 16
- Maximum sequence length: 1024; precision: BF16; packing: false
- Seed: 42; data seed: 42
- Public benchmark data: not used
Training data
The training pool contains 606 rows selected from the 3,000-row AI Hub 71875 K_train source after a pristine Base label probe: the Base model produced an unambiguous wrong answer label on these rows. The pool contains 175 internal-medicine, 177 obstetrics-and-gynecology, 182 pediatrics, and 72 emergency-medicine rows. Uncertain rows (669) were excluded. Training targets are the gold A–D answer labels; the Base model's wrong output is not included as a training target.
The source file SHA-256 is ff13bf66d46102c773d28934239ec347117726c617df74b6443b0763991cfa60. The source dataset is AI Hub 71875 필수의료 의학지식 QA: https://www.aihub.or.kr/aihubdata/data/view.do?currMenu=115&topMenu=100&aihubDataSe=realm&dataSetSn=71875
This is an experimental wrong-answer correction arm; it should not be interpreted as a general benchmark result. No benchmark questions, answers, evaluation artifacts, logs, credentials, or .env files are included in this repository. AI Hub source-data terms remain applicable.
Usage
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "youngseok12/HyperCLOVA-X-SEED-Think-14B-sft-71875-wrong-correction-minimal"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id, dtype=torch.bfloat16, device_map="auto"
)
messages = [{"role": "user", "content": "대한민국의 수도는 어디인가요?"}]
inputs = tokenizer.apply_chat_template(
messages, add_generation_prompt=True, tokenize=True, return_tensors="pt"
).to(model.device)
with torch.inference_mode():
outputs = model.generate(**inputs, max_new_tokens=64, do_sample=False)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
Intended use and limitations
This is an experimental Korean-language fine-tuned model for research and controlled evaluation. It can produce factual or reasoning errors and is not a substitute for professional medical, legal, financial, or other advice. The HyperCLOVA X acceptable-use restrictions and all applicable laws continue to apply to this derivative model.
License and notices
The full HyperCLOVA X SEED 14B Think Model License Agreement is included in LICENSE, and the required NAVER attribution is in NOTICE. Redistribution must comply with that agreement, including its prohibited-use policy and attribution requirements. AI Hub source-data terms also remain applicable.