Qwen3-14B SequenceClassification (multi-task) Prompt: base + choices + hints EMA, label smoothing, R-Drop monsaraida’s pipeline 2nd stage (inference on 50% of the test set) [Masaya-model-1] LoRA SFT: Qwen2.5-32B-Instruct CausalLM (Question-wise label pruning) Prompt: base + choices [monsaraida-model-1-a] LoRA SFT: Qwen3-14B SequenceClassification (multi-task) Prompt: base + choices + hints EMA, label smoothing, R-Drop, AWP Private = 0.945 Public = 0.950 CV = 0.949 Private = 0.945 Public = 0.949 CV = 0.948 [Masaya-model-2] QLoRA SFT: Qwen2.5-72B-Instruct quantized with AutoRound (GPTQ) CausalLM (Question-wise label pruning) Prompt: base + choices Private = 0.946 Public = 0.950 Private = 0.944 Public = 0.949 CV = 0.949 [monsaraida-model-2-a] LoRA SFT: Qwen3-14B SequenceClassification (multi-task) Prompt: base + choices + hints EMA, label smoothing, R-Drop Private = 0.944 Public = 0.950 CV = 0.948 [Masaya+monsaraida-model-1] LoRA SFT: Qwen2.5-32B-Instruct CausalLM (Question-wise label pruning) Prompt: base + choices EMA Private = 0.945 Public = 0.951 CV = 0.950 Masaya’s pipeline 1st stage (inference on the entire test set) Final Results Private = 0.948 (3rd) Public = 0.954 (1st) Ensemble https://www.kaggle.com/competitions/map-charting-student-math-misunderstandings/writeups/3rd-place-solution • プロンプトエンジニアリング ◦ 「他の選択肢」や「ヒント」を追加→たっ た数行の変更で効果抜群 • 検証戦略 ◦ (1) fold 0のみ (2) 5 fold平均 (3) 全学習で 複数シード平均 → 過学習を回避 • LoRA SFT マルチタスク学習 ◦ Qwen3-14B LoRA SFT (SeqCls) ◦ メイン65分類+補助2/3/36分類 • 汎化性向上 ◦ R-Drop / AWP → 精度改善 ◦ EMA → CV安定+精度改善 • 多段階推論 ◦ (1) test全体を推論 (2)確信度の低い50% のみ推論 → 推論高速化