AX-3.1-Light-sft_B12K_qcdd_1e

A.X-3.1-Light SFT B12K QCDD 1E

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

Model description

This model is a Korean question-answering and instruction-following model fine-tuned to place the answer first and optionally provide a short reason. The training set was selected from a frozen AI Hub candidate universe using quality, complexity, difficulty, and diversity-aware selection. The model is intended for research and evaluation of Korean text generation; its answers can be incorrect and it is not a substitute for professional judgment.

Training

• Base model: skt/A.X-3.1-Light
• Base revision: 9b41bb2406472634d8812c0b8931fa40fa9a6c3a
• Method: LoRA supervised fine-tuning with assistant-only causal-LM loss
• Precision: BF16
• Epochs: 1
• Learning rate: 5e-5
• Effective batch size: 8
• Maximum sequence length: 2,048
• Optimizer: adamw_torch_fused
• Scheduler: linear
• Warmup steps: 0
• Weight decay: 0
• 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
• Training examples: 12,000
• Training objective: 정답: <label or answer> (근거: 짧은 설명) answer-first format

The selected set contains 8,851 examples with a reasoning target and 3,149 examples without a reasoning target. The target contract limits an optional reason to a short explanation. Nine long-context examples required context trimming during serialization; no training target was truncated.

Training data

The examples were selected from 258,170 frozen clean candidates under the kds.selection.B12K_qcdd_scale.v1 policy, preserving the STEP 10 A1 source-independent greedy-selection semantics. Benchmark questions, answers, and evaluation artifacts are not included in this repository.

The source datasets are referenced by dataset ID only. Users must follow the applicable AI Hub terms of use.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "youngseok12/AX-3.1-Light-sft_B12K_qcdd_1e"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

Use the tokenizer's bundled chat template for conversational inference.

Local evaluation

These are local proxy results on the KDS canonical suite, not official K-AI leaderboard scores. The main average is the unweighted mean of the five main axes below, using the deterministic B1_constrained probe. HLE uses the text-only subset; Original MuSR is supplementary.

License and limitations

This model is derived from the Apache-2.0 licensed skt/A.X-3.1-Light model. The original base-model license, notices, and SK Telecom trademark terms also apply. AI Hub dataset terms apply to the source data. See LICENSE for the redistribution notice.

This model can produce inaccurate, incomplete, or poorly formatted answers. It should not be used as the sole basis for medical, legal, financial, or other high-impact decisions.