youngseok12/HyperCLOVA-X-SEED-Think-14B-minimal-base-correct-only-ties-r4min

HyperCLOVA X SEED Think-14B Minimal-base Correct-only TIES

This repository contains a standalone BF16 model based on the Minimal SFT merged model derived from naver-hyperclovax/HyperCLOVAX-SEED-Think-14B. Three fresh correct-only LoRA adapters were trained from that Minimal SFT model and merged with TIES. The model is intended for controlled Korean-language research and evaluation.

Model details

• —Training base: local Minimal SFT merged model, derived from naver-hyperclovax/HyperCLOVAX-SEED-Think-14B
• —Base revision: 9b74e35d4c7e4ffec489f4171273caca8948a2b9
• —Weight format: BF16 safetensors, standalone merged model
• —LoRA: rank 4, alpha 8, dropout 0, bias none
• —Target modules: q_proj, v_proj
• —Objective: assistant-only causal-language-model loss with a one-character answer-label target (A/B/C/D)
• —Learning rate: 1e-6; constant scheduler; warmup 0; weight decay 0
• —Effective batch size: 16 (per-device batch 1 × gradient accumulation 16)
• —Maximum sequence length: 1024; precision: BF16; seed/data seed: 42
• —Adapter steps: 71875 35, 71568 9, 71949 34
• —Merge: TIES, weights [1, 1, 1], density 0.5
• —Official HyperCLOVA X chat template: preserved byte-for-byte

Training data

The rows were selected from 1,000-row private-training probe pools using the pristine HyperCLOVA base model's unambiguous correct judgments. The selected rows were then converted to the Minimal SFT label-only schema before training the three adapters on the Minimal SFT base. Public benchmark questions, answers, and evaluation files were not used.

• —AI Hub 71875 — 필수의료 의학지식 데이터: 557 correct rows
• —AI Hub 71568 — 경제·스포츠 숫자연산 MRC: 131 correct rows
• —AI Hub 71949 — 인과관계 기반 추론 데이터(업사이클링) / label_json: 535 correct rows

The AI Hub source-data terms remain applicable. This model card does not redistribute the source datasets.

Verification

The merged artifact passed an independent NaN/Inf scan over 421 floating-point tensors and a deterministic generation smoke test. No downstream benchmark or K-AI leaderboard score is claimed for this upload.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "youngseok12/HyperCLOVA-X-SEED-Think-14B-minimal-base-correct-only-ties-r4min"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_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))

License and limitations

The base and derivative model are subject to NAVER's HyperCLOVA X SEED 14B Think Model License Agreement. See LICENSE and NOTICE. This is an experimental model and may produce factual or reasoning errors; it is not a substitute for professional medical, legal, financial, or other advice.