AX-3.1-Light-sft_v0_21_source_screen_numerical_300

AX-3.1-Light-sft_v0_21_source_screen_numerical_300

This is a standalone BF16 model obtained by fine-tuning skt/A.X-3.1-Light with a LoRA adapter and merging the adapter into the base weights. It is intended for Korean-language research and controlled evaluation. The repository contains no benchmark data, benchmark answers, training logs, or access credentials.

Model Details

• Base model: skt/A.X-3.1-Light
• Base revision used for training and merge: 9b41bb2406472634d8812c0b8931fa40fa9a6c3a
• Architecture: unchanged from the base model
• Weight format: BF16 safetensors
• Chat template: official A.X tokenizer chat template
• Custom Python model code: none
• Submission form: merged full model; no separate adapter is required
• Experiment condition: A v0.21-equivalent source-screening arm replacing 300 rows with numerical machine-reading examples.

Training Data

Each arm contains 5,801 training rows in a v0.21-equivalent source-screening mixture: 5,501 unchanged occurrences and 300 replacement occurrences. The replacement source for this model is AIHub-71568, 숫자연산 기계독해 데이터. The 300 examples were selected deterministically with seed 20260829; no quality, ranking, embedding, or model scoring was used for selection.

Public evaluation benchmarks such as KMMLU-Pro, CLIcK, HLE, SNU Ko-MuSR, Com2-main, and Original MuSR were not used as SFT data. The applicable terms of the AI Hub source data remain in force.

Training Procedure

• Objective: assistant-only causal-language-model cross entropy
• Epochs: 1
• Learning rate: 5e-5
• Optimizer: adamw_torch_fused
• Scheduler: linear, no warmup
• Weight decay: 0.0
• Maximum gradient norm: 1.0
• LoRA: rank 16, alpha 32, dropout 0.05
• LoRA target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Per-device batch size: 1
• Gradient accumulation: 8 (effective batch size 8)
• Maximum sequence length: 2048
• Precision: BF16
• Packing: disabled
• Random seed: 42

Local Evaluation

The full local canonical-suite evaluation used free and B1_constrained probes, 21,962 rows per model, and had zero generation errors. The following are the primary B1_constrained parsed-accuracy results; they are local evaluation results, not official K-AI leaderboard scores.

KMMLU-Pro 40.54%, CLIcK 64.31%, HLE(Ko) 4.31%, SNU Ko-MuSR 55.07%, Com2-main(Ko) 52.40%; five-axis mean 43.33%.

Usage

Load the repository with transformers AutoModelForCausalLM and AutoTokenizer, or directly with standard vLLM. The merged repository does not require a separate adapter or trust_remote_code.

Intended Use and Limitations

This model is an experimental Korean SFT model for research and controlled evaluation. It can produce factual errors and should not be used as a substitute for professional legal, accounting, medical, or financial advice.

License

The base model is distributed under the Apache License 2.0. The applicable terms of the AI Hub source data remain in force for use of the training data. See LICENSE for the base model license text.