--- language: - ja license: cc-by-sa-4.0 size_categories: - 1K= 4.0 フィルタが quality_score の borderline / weak positive / false negative を全て吸収するため、n-tuples-filtered の行数は n-tuples と完全に一致する (全 4,603 行が validated)。 ### label カラムの形式 List[float] 長さ 6 で [score(query, positive), score(query, neg_1), …, score(query, neg_5)]。label[0] が positive、label[1:] が negative_1〜negative_5 に 1:1 で順序通り対応する。スコアは Cross-Encoder の 生 logit (sigmoid 適用前)。 ## データセットの構築手順 ### pairs 1. JMTEB の miracl-retrieval-query (train, 2,433 行) と miracl-retrieval-corpus (corpus, 6,953,614 行) を取得する。 2. 各クエリの relevant_docs フィールドに含まれる docid を集合化する (4,589 件のユニーク docid)。 3. corpus から該当 docid の text を取り出し、(query, docid) を (query, positive) ペアに展開する。 4. 1 クエリに対し relevant_docs が複数ある場合は、複数行に展開する。 構築結果: | 項目 | 件数 | |---|---:| | 元クエリ数 (train) | 2,433 | | ユニーク docid 数 | 4,589 | | 生成された pairs | 4,932 | | 展開倍率 (pairs / queries) | 約 2.03 | ### triplets / n-tuples (Hard Negative Mining) 1. 候補プール構築 — pairs の全ユニーク positive (4,589 件) と、miracl-retrieval-corpus からランダムサンプリングした 2,000,000 件のテキスト (seed=42) を結合し、重複除去後の 2,000,605 件を候補プールとする。 2. 埋め込み生成 — 候補プールおよびユニーククエリ (2,433 件) を [cl-nagoya/ruri-v3-310m](https://huggingface.co/cl-nagoya/ruri-v3-310m) (768 次元、ModernBERT-Ja-310m ベース) でエンコード。検索クエリには "検索クエリ: "、文書側には "検索文書: " の prefix を付与した上で、L2 正規化された fp16 埋め込みを得る。 3. kNN 検索 — GPU matmul + chunked topk によりクエリ × 候補プールの cosine 類似度を計算、各クエリの top-100 を取得する (top-15 の運用に加え、後段の margin filter で不足する場合の拡張余地を確保。実測では top-up 不要だった)。 4. リランクスコア計算 — [cl-nagoya/ruri-v3-reranker-310m](https://huggingface.co/cl-nagoya/ruri-v3-reranker-310m) (Cross-Encoder) で、各ユニーク (query, positive) と (query, kNN top-50 候補) の組について 生 logit (sigmoid 適用前) を計算する (合計 121,714 ペア)。 5. Hard Negative 選定 — 行ごとに以下のステップで 5 件を選ぶ。 1. 各 pair の positive スコアが 2.0 未満なら、その行を丸ごと除外する (σ(2.0) ≈ 0.88 未満の弱い positive は学習雑音になりやすい)。 2. kNN top 候補から、同一クエリの他の positive と一致する pool_id を除外する (false negative 防止)。 3. positive スコアとの差が 4.0 未満 (pos - neg < 4.0) の候補を Hard Negative 候補から外す。 4. 残った候補のリランクスコア降順で上位 5 件を採用する。 5. 候補が 5 件に満たない場合は、kNN を top-100 まで拡張した中で margin filter を通過した候補から補い、それでも不足する場合のみ margin filter 落ち候補のうち最もスコアの高いものを補充する (今回のデータでは全行 margin filter のみで 5 件を満たし、top-up は発生しなかった)。 6. triplets と n-tuples 出力 — n-tuples は (query, positive, negative_1〜5, label)、triplets は同じ negative_1 を 1 件 negative とした (query, positive, negative)。 #### 統計 (HNM 結果) | 指標 | 値 | |---|---:| | 入力ペア数 | 4,932 | | positive スコア < 2.0 で除外 | 329 (6.7%) | | 候補不足で除外 | 0 | | margin filter のみで 5 件確保 | 4,603 (100%) | | top-up 使用行 | 0 | | 累計 negative (margin filter 通過) | 23,015 | | 累計 negative (top-up) | 0 | | 出力 triplets / n-tuples 行数 | 4,603 | #### label カラム統計 (n-tuples) | 項目 | min | median | mean | max | |---|---:|---:|---:|---:| | positive logit | 2.00 | 7.93 | 7.87 | 13.25 | | max(neg) logit | -7.62 | 0.99 | 1.21 | 8.75 | | mean(neg) logit | -8.29 | -0.62 | -0.42 | 7.72 | | margin (= pos - max_neg) | 4.00 | 5.51 | 6.66 | 19.16 | positive は σ ≈ 0.88〜0.99999 に集中しており、教師が確信を持って relevant と判定した行のみが残っている。max(neg) は σ ≈ 0.5 付近 (mean 1.21、σ(1.21) ≈ 0.77) で「relevant 寄りだが positive より弱い」典型的な Hard Negative の位置にある。margin は最小 4.0 (フィルタどおり) で、Goldilocks ゾーン (+3〜+8) に中央値が収まっている。 ### n-tuples-filtered (quality_score validated 行のみ) Cross-Encoder 蒸留スコア付与ガイドの quality_score 定義に従い、n-tuples の label から行ごとに 3 つの mask を計算する。 ```python import numpy as np POS_MIN = 2.0 # σ(2.0) ≈ 0.88: 確信のある positive のみ採用 MARGIN_MIN = 0.5 # 誤差レベルの分離は雑音と見なす MARGIN_PENALTY = 0.1 # 易しすぎる行を弱く後回しにする係数 def quality_score_vec(labels: np.ndarray) -> np.ndarray: p = labels[:, 0] neg = labels[:, 1:] max_n = neg.max(axis=1) mean_n = neg.mean(axis=1) margin = p - max_n fn_mask = margin <= 0 weak_mask = (~fn_mask) & (p < POS_MIN) border_mask = (~fn_mask) & (~weak_mask) & (margin < MARGIN_MIN) valid_mask = (~fn_mask) & (~weak_mask) & (~border_mask) raw = mean_n - MARGIN_PENALTY * margin return np.where(valid_mask, raw, -np.inf) ``` n-tuples-filtered は valid_mask が True の行のみを残し、quality_score 降順 で並べたもの。今回の HNM 設計では margin >= 4.0 のフィルタが quality_score の各 mask を全て吸収するため、n-tuples の全 4,603 行が validated 判定となり、行集合としては n-tuples と一致する (順序は quality_score 降順)。フィルタの内訳: | フィルタ | 条件 | 意味 | |---|---|---| | 偽 negative | max(neg) >= positive | negative 群に実は relevant な passage が混入 | | weak positive | positive < 2.0 | 教師が positive を「弱く relevant」としか見ていない | | borderline | 0 < margin < 0.5 | positive と最強 negative の差が誤差程度 | ## 蒸留スコア付けの設定 | 項目 | 値 | |---|---| | Teacher | cl-nagoya/ruri-v3-reranker-310m | | Max length | 512 | | 推論精度 | FP16 | | Batch size | 128 | | 入力フォーマット | 単一文ペア (query, document) | | 出力 | 生 logit (sigmoid 適用前) | | 観測 logit レンジ | min -10.94, max 13.26 (mean -4.62 / median -5.65 across 121,714 pairs) | KL 蒸留損失 (SparseDistillKLDivLoss / DistillKLDivLoss 等) は softmax(label / T) を target distribution として使うため、sigmoid 後の [0, 1] 値だと softmax target が一様分布に潰れて ranking 情報が失われる。これを防ぐため、教師の最終活性化を明示的に Identity に置き換えてスコアを出力した。 ```python import torch.nn as nn from sentence_transformers import CrossEncoder teacher = CrossEncoder("cl-nagoya/ruri-v3-reranker-310m", max_length=512, device="cuda") teacher.model.eval() teacher.default_activation_function = nn.Identity() scores = teacher.predict( [("query A", "passage A"), ("query A", "passage B")], batch_size=128, convert_to_numpy=True, activation_fct=nn.Identity(), ) assert abs(scores).max() > 1.0 ``` logit のスケール感: | logit | sigmoid(logit) | 解釈 | | ----: | --------------: | ------------------------------------- | | +8 | 0.9997 | 確信度高く relevant | | +5 | 0.993 | 強く relevant | | +2 | 0.881 | relevant 寄り | | 0 | 0.5 | 不確か | | -2 | 0.119 | irrelevant 寄り | | -5 | 0.007 | 強く irrelevant | | -8 | 0.0003 | 確信度高く irrelevant | ## KL 蒸留損失への組み込み ```python import torch.nn.functional as F def kl_distill_loss(student_logits, teacher_label, T=2.0): log_p_student = F.log_softmax(student_logits / T, dim=-1) p_teacher = F.softmax(teacher_label / T, dim=-1) return F.kl_div(log_p_student, p_teacher, reduction="batchmean") * (T * T) ``` sentence-transformers の SparseDistillKLDivLoss / DistillKLDivLoss を使う場合は、label カラムをそのまま teacher score として渡せばよい。 ## License CC BY-SA 4.0 直接のソースである sbintuitions/JMTEB が CC BY-SA 4.0 で配布されているため、本データセットも CC BY-SA 4.0 を継承する。元の MIRACL は Apache-2.0、corpus の Wikipedia 由来テキストは CC BY-SA に従う。 ## Citation ```bibtex @misc{sbintuitions_jmteb, author = {SB Intuitions}, title = {JMTEB: Japanese Massive Text Embedding Benchmark}, year = {2024}, url = {https://huggingface.co/datasets/sbintuitions/JMTEB}, } @article{miracl, author = {Xinyu Zhang and Nandan Thakur and Odunayo Ogundepo and Ehsan Kamalloo and David Alfonso-Hermelo and Xiaoguang Li and Qun Liu and Mehdi Rezagholizadeh and Jimmy Lin}, title = {{MIRACL}: A Multilingual Retrieval Dataset Covering 18 Diverse Languages}, journal = {Transactions of the Association for Computational Linguistics}, year = {2023}, } ``` ## Acknowledgements - 元データ: [sbintuitions/JMTEB](https://huggingface.co/datasets/sbintuitions/JMTEB) - 元タスク: [MIRACL](https://project-miracl.github.io/) - HNM 用 retriever: [cl-nagoya/ruri-v3-310m](https://huggingface.co/cl-nagoya/ruri-v3-310m) - 蒸留教師: [cl-nagoya/ruri-v3-reranker-310m](https://huggingface.co/cl-nagoya/ruri-v3-reranker-310m)