youngseok12/AX-3.1-Light-sft_A1_paperguided4k

A.X-3.1-Light SFT A1 (Paper-guided 4K)

This is a full BF16 model obtained by merging a LoRA SFT adapter into `skt/A.X-3.1-Light`. No separate adapter is required for inference.

Training

• —Method: LoRA supervised fine-tuning with assistant-only loss
• —Data: 4,000 selected examples from five AI Hub sources
• —Selection: A1 paper-guided selection with source/context/document balancing
• —Base model: skt/A.X-3.1-Light
• —Precision: BF16
• —Epochs: 1
• —Learning rate: 5e-5
• —Maximum sequence length: 2,048
• —Effective batch size: 8
• —Seed: 42
• —LoRA rank / alpha / dropout: 16 / 32 / 0.05
• —LoRA target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

AI Hub sources

The source data are referenced by dataset ID only; no benchmark data or training examples are included in this repository. Users must follow the applicable AI Hub terms of use.

• —569 — 행정 문서 대상 기계독해: AI Hub
• —71610 — 금융·법률 문서 기계독해: AI Hub
• —71857 — 국어 교과 지문형 문제: AI Hub
• —71874 — 전문 의학지식: AI Hub
• —71949 — 인과 관계 기반 추론: AI Hub

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "youngseok12/AX-3.1-Light-sft_A1_paperguided4k"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

Local evaluation

These are local proxy results on the KDS canonical suite, not official K-AI leaderboard scores. The evaluation used the text-only HLE subset and reported Original MuSR separately in the suite.

The model is derived from the Apache-2.0 licensed base model. See the `LICENSE` file and the base model's license for terms and notices.

Limitations

This model was fine-tuned for Korean text-generation and benchmark-style question answering. It may produce incorrect or poorly formatted answers and should not be used as the sole basis for medical, legal, financial, or other high-impact decisions.