youngseok12/HyperCLOVA-X-SEED-Think-14B-minimal-sft-71875-71949-correct-540

HyperCLOVA X SEED Think-14B Minimal SFT + 71949 Correct 540

This repository contains a standalone BF16 model created by continuing the Minimal SFT model `youngseok12/HyperCLOVA-X-SEED-Think-14B-minimal-sft-71875` with a small correct-only LoRA on AI Hub 71949 data.

The direct starting checkpoint was the locally verified merged Minimal SFT model. Its original foundation is naver-hyperclovax/HyperCLOVAX-SEED-Think-14B, revision 9b74e35d4c7e4ffec489f4171273caca8948a2b9.

Data

• —AI Hub dataset 71949: <https://www.aihub.or.kr/aihubdata/data/view.do?currMenu=115&topMenu=100&aihubDataSe=realm&dataSetSn=71949>
• —Stratified private-training probe pool: 1,000 rows
• —Correct-only rows used for training: 540
• —Excluded: 166 wrong and 294 uncertain rows
• —Target format: one correct choice letter, A/B/C/D
• —No duplication or resampling
• —Public benchmark questions, answers, and evaluation artifacts were not used
• —Correct-pool input SHA-256: 29791fa19d412d3c48d7080381079ab3ccf3190059a5ba04e94be37f479feef8

Training

The same low-impact Minimal SFT recipe was used:

• —LoRA rank 4, alpha 8, dropout 0
• —Target modules: q_proj, v_proj
• —Learning rate 1e-6, constant scheduler, no warmup, weight decay 0
• —BF16, maximum sequence length 1024
• —Per-device batch 1, gradient accumulation 16
• —Assistant-only loss, packing disabled
• —One epoch, 34 optimizer steps
• —Seed and data seed: 42

The final train loss was 3.9535395538. The adapter was merged into the Minimal SFT checkpoint using BF16 safetensors.

Validation

• —Adapter tensors: 152; NaN/Inf: 0
• —Merged model: 6 safetensors shards, 421 tensors; NaN/Inf: 0
• —Generation smoke: 2/2 passed
• —Local canonical benchmark and official K-AI leaderboard score: not measured

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "youngseok12/HyperCLOVA-X-SEED-Think-14B-minimal-sft-71875-71949-correct-540"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

License and limitations

This derivative model is subject to the HyperCLOVA X SEED 14B Think Model License Agreement in LICENSE, with the required attribution in NOTICE. AI Hub source-data terms also apply. This is an experimental research and evaluation model; outputs may be incorrect and it is not a substitute for professional medical, legal, financial, or other advice.