Upload folder using huggingface_hub

README.md

@@ -0,0 +1,153 @@
+---
+license: mit
+datasets:
+- hpprc/emb
+- hotchpotch/hpprc_emb-scores
+- microsoft/ms_marco
+language:
+- ja
+base_model:
+- tohoku-nlp/bert-base-japanese-v3
+---
+
+高性能な日本語 [SPLADE](https://github.com/naver/splade) (Sparse Lexical and Expansion Model) モデルです。[テキストからスパースベクトルへの変換デモ](https://huggingface.co/spaces/hotchpotch/japanese-splade-demo-streamlit)で、どのようなスパースベクトルに変換できるか、WebUI から気軽にお試しいただけます。
+
+なお、テクニカルレポートは後日公開予定です。
+
+
+# 利用方法
+
+## [YASEM (Yet Another Splade|Sparse Embedder)](https://github.com/hotchpotch/yasem)
+
+```bash
+pip install yasem
+```
+
+```python
+from yasem import SpladeEmbedder
+
+model_name = "hotchpotch/japanese-splade-base-v1"
+embedder = SpladeEmbedder(model_name)
+
+sentences = [
+  "車の燃費を向上させる方法は？",
+  "急発進や急ブレーキを避け、一定速度で走行することで燃費が向上します。",
+  "車を長持ちさせるには、消耗品を適切なタイミングで交換することが重要です。",
+]
+
+embeddings = embedder.encode(sentences)
+similarity = embedder.similarity(embeddings, embeddings)
+
+print(similarity)
+# [[21.49299249 10.48868281  6.25582337]
+#  [10.48868281 12.90587398  3.19429791]
+#  [ 6.25582337  3.19429791 12.89678271]]
+```
+
+```python
+token_values = embedder.get_token_values(embeddings[0])
+
+print(token_values)
+
+#{
+#    '車': 2.1796875,
+#    '燃費': 2.146484375,
+#    '向上': 1.7353515625,
+#    '方法': 1.55859375,
+#    '燃料': 1.3291015625,
+#    '効果': 1.1376953125,
+#    '良い': 0.873046875,
+#    '改善': 0.8466796875,
+#    'アップ': 0.833984375,
+#    'いう': 0.70849609375,
+#    '理由': 0.64453125,
+# ...
+```
+
+## transformers
+
+```python
+
+from transformers import AutoModelForMaskedLM, AutoTokenizer
+import torch
+
+model = AutoModelForMaskedLM.from_pretrained(model_name)
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+def splade_max_pooling(logits, attention_mask):
+    relu_log = torch.log(1 + torch.relu(logits))
+    weighted_log = relu_log * attention_mask.unsqueeze(-1)
+    max_val, _ = torch.max(weighted_log, dim=1)
+    return max_val
+
+tokens = tokenizer(
+    sentences, return_tensors="pt", padding=True, truncation=True, max_length=512
+)
+tokens = {k: v.to(model.device) for k, v in tokens.items()}
+
+with torch.no_grad():
+    outputs = model(**tokens)
+embeddings = splade_max_pooling(outputs.logits, tokens["attention_mask"])
+
+similarity = torch.matmul(embeddings.unsqueeze(0), embeddings.T).squeeze(0)
+print(similarity)
+
+# tensor([[21.4943, 10.4816,  6.2540],
+#         [10.4816, 12.9024,  3.1939],
+#         [ 6.2540,  3.1939, 12.8919]])
+```
+
+# ベンチマークスコア
+
+## retrieval (JMTEB)
+
+[JMTEB](https://github.com/sbintuitions/JMTEB) の評価結果です。japanese-splade-base-v1 は [JMTEB をスパースベクトルで評価できるように変更したコード](https://github.com/hotchpotch/JMTEB/tree/add_splade)での評価となっています。 
+なお、japanese-splade-base-v1 は jaqket, mrtydi のドメインを学習(testのデータ以外)しています。
+
+| model_name                                                                 |  Avg.   | jagovfaqs_22k | jaqket | mrtydi | nlp_journal_abs_intro | nlp_journal_title_abs | nlp_journal_title_intro |
+| :------------------------------------------------------------------------- | ------: | ------------: | -----: | -----: | ---------------------: | ---------------------: | -----------------------: |
+| [japanese-splade-base-v1](https://huggingface.co/hotchpotch/japanese-splade-base-v1) | **0.7465** | 0.6499        | **0.6992** | **0.4365** | 0.8967                | **0.9766**            | 0.8203                  |
+| [text-embedding-3-large](https://huggingface.co/OpenAI/text-embedding-3-large) | 0.7448 | 0.7241        | 0.4821 | 0.3488 | **0.9933**            | 0.9655                | **0.9547**              |
+| [GLuCoSE-base-ja-v2](https://huggingface.co/pkshatech/GLuCoSE-base-ja-v2) | 0.7336 | 0.6979        | 0.6729 | 0.4186 | 0.9029                | 0.9511                | 0.7580                  |
+| [multilingual-e5-large](https://huggingface.co/intfloat/multilingual-e5-large) | 0.7098 | 0.7030        | 0.5878 | 0.4363 | 0.8600                | 0.9470                | 0.7248                  |
+| [multilingual-e5-small](https://huggingface.co/intfloat/multilingual-e5-small) | 0.6727 | 0.6411        | 0.4997 | 0.3605 | 0.8521                | 0.9526                | 0.7299                  |
+| [ruri-large](https://huggingface.co/cl-nagoya/ruri-large)                       | 0.7302 | **0.7668**    | 0.6174 | 0.3803 | 0.8712                | 0.9658                | 0.7797                  |
+
+
+## reranking
+
+### [JaCWIR](https://huggingface.co/datasets/hotchpotch/JaCWIR)
+
+なお、japanese-splade-base-v1 は **JaCWIR のドメインを学習していません**。
+
+| model_names                                                                     | map@10 | hit_rate@10 |
+| :------------------------------------------------------------------------------ | -----: | ----------: |
+| [japanese-splade-base-v1](https://huggingface.co/hotchpotch/japanese-splade-base-v1) |   **0.9122** |        **0.9854** |
+| [text-embedding-3-small](https://platform.openai.com/docs/guides/embeddings)    | 0.8168 |      0.9506 |
+| [GLuCoSE-base-ja-v2](https://huggingface.co/pkshatech/GLuCoSE-base-ja-v2)       | 0.8567 |      0.9676 |
+| [bge-m3+dense](https://huggingface.co/BAAI/bge-m3)                              | 0.8642 |      0.9684 |
+| [multilingual-e5-large](https://huggingface.co/intfloat/multilingual-e5-large)  | 0.8759 |      0.9726 |
+| [multilingual-e5-small](https://huggingface.co/intfloat/multilingual-e5-small)  |  0.869 |        0.97 |
+| [ruri-large](https://huggingface.co/cl-nagoya/ruri-large)                       | 0.8291 |      0.9594 |
+
+### [JQaRA](https://github.com/hotchpotch/JQaRA)
+なお、japanese-splade-base-v1 は JQaRA のドメイン(test以外)を学習したものとなっています。
+
+| model_names                                                                     | ndcg@10 | mrr@10 |
+| :------------------------------------------------------------------------------ | ------: | -----: |
+| [japanese-splade-base-v1](https://huggingface.co/hotchpotch/japanese-splade-base-v1) |    **0.6441** |   **0.8616** |
+| [text-embedding-3-small](https://platform.openai.com/docs/guides/embeddings)    |  0.3881 | 0.6107 |
+| [bge-m3+dense](https://huggingface.co/BAAI/bge-m3)                              |   0.539 | 0.7854 |
+| [multilingual-e5-large](https://huggingface.co/intfloat/multilingual-e5-large)  |   0.554 | 0.7988 |
+| [multilingual-e5-small](https://huggingface.co/intfloat/multilingual-e5-small)  |  0.4917 | 0.7291 |
+| [GLuCoSE-base-ja-v2](https://huggingface.co/pkshatech/GLuCoSE-base-ja-v2)   |    0.606  |   0.8359 |
+| [ruri-large](https://huggingface.co/cl-nagoya/ruri-large)                   |    0.6287 |   0.8418 |
+
+## 学習元データセット
+
+ [hpprc/emb](https://huggingface.co/datasets/hpprc/emb) から、auto-wiki-qa, mmarco, jsquad jaquad, auto-wiki-qa-nemotron, quiz-works quiz-no-mori, miracl, jqara mr-tydi, baobab-wiki-retrieval, mkqa データセットを利用しています。
+ また英語データセットとして、MS Marcoを利用しています。
+
+## 注意事項
+
+`tokenizer.json` ファイルを同梱していますが、このファイルは text-embeddings-inference を動かすためのダミーファイルです。詳細は、[text-embeddings-inference で日本語トークナイザーモデルの推論をする](https://secon.dev/entry/2024/09/30/160000/)をご覧ください。

config.json

@@ -0,0 +1,25 @@
+{
+  "_name_or_path": "./",
+  "architectures": [
+    "BertForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "classifier_dropout": null,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "position_embedding_type": "absolute",
+  "torch_dtype": "float16",
+  "transformers_version": "4.43.0.dev0",
+  "type_vocab_size": 2,
+  "use_cache": true,
+  "vocab_size": 32768
+}

generation_config.json

@@ -0,0 +1,5 @@
+{
+  "_from_model_config": true,
+  "pad_token_id": 0,
+  "transformers_version": "4.43.0.dev0"
+}

model_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bfbc21a3e0df40bbb2536a2ee9f50a688c70754b63af7c854de0f86ffbbfdd90
+size 1004

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79b9e3c40771abae754a234c6bef00ac531c0e7d903efb478090dd03f9b12ff1
+size 222563762

special_tokens_map.json

@@ -0,0 +1,7 @@
+{
+  "cls_token": "[CLS]",
+  "mask_token": "[MASK]",
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "unk_token": "[UNK]"
+}

tokenizer.json

@@ -0,0 +1,522 @@
+{
+  "version": "1.0",
+  "truncation": null,
+  "padding": null,
+  "added_tokens": [
+    {
+      "id": 0,
+      "content": "[PAD]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 1,
+      "content": "[UNK]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 2,
+      "content": "[CLS]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 3,
+      "content": "[SEP]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 4,
+      "content": "[MASK]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    }
+  ],
+  "normalizer": null,
+  "pre_tokenizer": {
+    "type": "Whitespace"
+  },
+  "post_processor": {
+    "type": "TemplateProcessing",
+    "single": [
+      {
+        "SpecialToken": {
+          "id": "[CLS]",
+          "type_id": 0
+        }
+      },
+      {
+        "Sequence": {
+          "id": "A",
+          "type_id": 0
+        }
+      },
+      {
+        "SpecialToken": {
+          "id": "[SEP]",
+          "type_id": 0
+        }
+      }
+    ],
+    "pair": [
+      {
+        "SpecialToken": {
+          "id": "[CLS]",
+          "type_id": 0
+        }
+      },
+      {
+        "Sequence": {
+          "id": "A",
+          "type_id": 0
+        }
+      },
+      {
+        "SpecialToken": {
+          "id": "[SEP]",
+          "type_id": 0
+        }
+      },
+      {
+        "SpecialToken": {
+          "id": "[CLS]",
+          "type_id": 0
+        }
+      },
+      {
+        "Sequence": {
+          "id": "B",
+          "type_id": 0
+        }
+      },
+      {
+        "SpecialToken": {
+          "id": "[SEP]",
+          "type_id": 0
+        }
+      }
+    ],
+    "special_tokens": {
+      "[CLS]": {
+        "id": "[CLS]",
+        "ids": [
+          2
+        ],
+        "tokens": [
+          "[CLS]"
+        ]
+      },
+      "[SEP]": {
+        "id": "[SEP]",
+        "ids": [
+          3
+        ],
+        "tokens": [
+          "[SEP]"
+        ]
+      }
+    }
+  },
+  "decoder": {
+    "type": "BPEDecoder",
+    "suffix": "</w>"
+  },
+  "model": {
+    "type": "BPE",
+    "dropout": null,
+    "unk_token": null,
+    "continuing_subword_prefix": null,
+    "end_of_word_suffix": null,
+    "fuse_unk": false,
+    "byte_fallback": false,
+    "ignore_merges": false,
+    "vocab": {
+      "[PAD]": 0,
+      "[UNK]": 1,
+      "[CLS]": 2,
+      "[SEP]": 3,
+      "[MASK]": 4,
+      "!": 5,
+      "\"": 6,
+      "%": 7,
+      "&": 8,
+      "'": 9,
+      "(": 10,
+      ")": 11,
+      "*": 12,
+      "+": 13,
+      ",": 14,
+      "-": 15,
+      ".": 16,
+      "/": 17,
+      "0": 18,
+      "1": 19,
+      "2": 20,
+      "3": 21,
+      "4": 22,
+      "5": 23,
+      "6": 24,
+      "7": 25,
+      "8": 26,
+      "9": 27,
+      ":": 28,
+      ";": 29,
+      "?": 30,
+      "A": 31,
+      "B": 32,
+      "C": 33,
+      "D": 34,
+      "E": 35,
+      "F": 36,
+      "G": 37,
+      "H": 38,
+      "I": 39,
+      "J": 40,
+      "K": 41,
+      "L": 42,
+      "M": 43,
+      "N": 44,
+      "O": 45,
+      "P": 46,
+      "Q": 47,
+      "R": 48,
+      "S": 49,
+      "T": 50,
+      "U": 51,
+      "V": 52,
+      "W": 53,
+      "X": 54,
+      "Y": 55,
+      "Z": 56,
+      "[": 57,
+      "]": 58,
+      "_": 59,
+      "a": 60,
+      "b": 61,
+      "c": 62,
+      "d": 63,
+      "e": 64,
+      "f": 65,
+      "g": 66,
+      "h": 67,
+      "i": 68,
+      "j": 69,
+      "k": 70,
+      "l": 71,
+      "m": 72,
+      "n": 73,
+      "o": 74,
+      "p": 75,
+      "q": 76,
+      "r": 77,
+      "s": 78,
+      "t": 79,
+      "u": 80,
+      "v": 81,
+      "w": 82,
+      "x": 83,
+      "y": 84,
+      "z": 85,
+      "|": 86,
+      "§": 87,
+      "Á": 88,
+      "Æ": 89,
+      "á": 90,
+      "æ": 91,
+      "ç": 92,
+      "è": 93,
+      "é": 94,
+      "í": 95,
+      "ð": 96,
+      "ö": 97,
+      "ú": 98,
+      "ü": 99,
+      "þ": 100,
+      "ā": 101,
+      "ē": 102,
+      "ŋ": 103,
+      "ƿ": 104,
+      "ɑ": 105,
+      "ɒ": 106,
+      "ɔ": 107,
+      "ɖ": 108,
+      "ə": 109,
+      "ɚ": 110,
+      "ɛ": 111,
+      "ɜ": 112,
+      "ɡ": 113,
+      "ɪ": 114,
+      "ɫ": 115,
+      "ɹ": 116,
+      "ɾ": 117,
+      "ʃ": 118,
+      "ʈ": 119,
+      "ʊ": 120,
+      "ʌ": 121,
+      "ʍ": 122,
+      "ʒ": 123,
+      "ʔ": 124,
+      "ʰ": 125,
+      "ʱ": 126,
+      "ʲ": 127,
+      "ʷ": 128,
+      "ˈ": 129,
+      "ː": 130,
+      "ˑ": 131,
+      "̚": 132,
+      "̥": 133,
+      "̩": 134,
+      "̪": 135,
+      "̯": 136,
+      "͡": 137,
+      "θ": 138,
+      "‑": 139,
+      "–": 140,
+      "—": 141,
+      "∅": 142,
+      "⟨": 143,
+      "⟩": 144,
+      "an": 145,
+      "th": 146,
+      "in": 147,
+      "on": 148,
+      "er": 149,
+      "is": 150,
+      "es": 151,
+      "or": 152,
+      "the": 153,
+      "ti": 154,
+      "ar": 155,
+      "al": 156,
+      "en": 157,
+      "ed": 158,
+      "of": 159,
+      "and": 160,
+      "gl": 161,
+      "ish": 162,
+      "ngl": 163,
+      "Engl": 164,
+      "English": 165,
+      "as": 166,
+      "ic": 167,
+      "ou": 168,
+      "20": 169,
+      "tion": 170,
+      "ing": 171,
+      "ec": 172,
+      "om": 173,
+      "at": 174,
+      "st": 175,
+      "it": 176,
+      "le": 177,
+      "ge": 178,
+      "re": 179,
+      "gu": 180,
+      "angu": 181,
+      "angua": 182,
+      "ch": 183,
+      "ent": 184,
+      "ve": 185,
+      "to": 186,
+      ").": 187,
+      "ation": 188,
+      "ri": 189,
+      "ly": 190,
+      "am": 191,
+      "oun": 192,
+      "ers": 193,
+      "anguage": 194,
+      "for": 195,
+      "fr": 196,
+      "ll": 197,
+      "us": 198,
+      "200": 199,
+      "he": 200,
+      "tic": 201,
+      "pr": 202,
+      "di": 203,
+      "ow": 204,
+      "et": 205,
+      "ig": 206,
+      "19": 207,
+      "pe": 208,
+      "ac": 209,
+      ".[": 210,
+      "ur": 211,
+      "wi": 212,
+      "201": 213,
+      "ect": 214,
+      "iv": 215,
+      "ess": 216,
+      "The": 217,
+      "ol": 218,
+      "ter": 219,
+      "de": 220,
+      "language": 221,
+      "wor": 222,
+      "from": 223,
+      "un": 224,
+      "In": 225,
+      "ver": 226,
+      "ir": 227,
+      "are": 228,
+      "cl": 229,
+      "ther": 230,
+      "ad": 231,
+      "man": 232,
+      "con": 233,
+      "ab": 234,
+      "ex": 235,
+      "with": 236,
+      "pp": 237,
+      "wh": 238,
+      "el": 239,
+      "97": 240,
+      "ary": 241,
+      "10": 242,
+      "su": 243,
+      "ph": 244,
+      "ul": 245,
+      "po": 246,
+      "978": 247,
+      "ld": 248,
+      "ak": 249,
+      "si": 250,
+      "ru": 251,
+      "tive": 252,
+      "ds": 253,
+      "oc": 254,
+      "enc": 255
+    },
+    "merges": [
+      "a n",
+      "t h",
+      "i n",
+      "o n",
+      "e r",
+      "i s",
+      "e s",
+      "o r",
+      "th e",
+      "t i",
+      "a r",
+      "a l",
+      "e n",
+      "e d",
+      "o f",
+      "an d",
+      "g l",
+      "is h",
+      "n gl",
+      "E ngl",
+      "Engl ish",
+      "a s",
+      "i c",
+      "o u",
+      "2 0",
+      "ti on",
+      "in g",
+      "e c",
+      "o m",
+      "a t",
+      "s t",
+      "i t",
+      "l e",
+      "g e",
+      "r e",
+      "g u",
+      "an gu",
+      "angu a",
+      "c h",
+      "en t",
+      "v e",
+      "t o",
+      ") .",
+      "a tion",
+      "r i",
+      "l y",
+      "a m",
+      "ou n",
+      "er s",
+      "angua ge",
+      "f or",
+      "f r",
+      "l l",
+      "u s",
+      "20 0",
+      "h e",
+      "ti c",
+      "p r",
+      "d i",
+      "o w",
+      "e t",
+      "i g",
+      "1 9",
+      "p e",
+      "a c",
+      ". [",
+      "u r",
+      "w i",
+      "20 1",
+      "ec t",
+      "i v",
+      "es s",
+      "T he",
+      "o l",
+      "t er",
+      "d e",
+      "l anguage",
+      "w or",
+      "fr om",
+      "u n",
+      "I n",
+      "v er",
+      "i r",
+      "ar e",
+      "c l",
+      "th er",
+      "a d",
+      "m an",
+      "c on",
+      "a b",
+      "e x",
+      "wi th",
+      "p p",
+      "w h",
+      "e l",
+      "9 7",
+      "ar y",
+      "1 0",
+      "s u",
+      "p h",
+      "u l",
+      "p o",
+      "97 8",
+      "l d",
+      "a k",
+      "s i",
+      "r u",
+      "ti ve",
+      "d s",
+      "o c",
+      "en c"
+    ]
+  }
+}

tokenizer_config.json

@@ -0,0 +1,63 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "[CLS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "[SEP]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "4": {
+      "content": "[MASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "clean_up_tokenization_spaces": true,
+  "cls_token": "[CLS]",
+  "do_lower_case": false,
+  "do_subword_tokenize": true,
+  "do_word_tokenize": true,
+  "jumanpp_kwargs": null,
+  "mask_token": "[MASK]",
+  "mecab_kwargs": {
+    "mecab_dic": "unidic_lite"
+  },
+  "model_max_length": 512,
+  "never_split": null,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "subword_tokenizer_type": "wordpiece",
+  "sudachi_kwargs": null,
+  "tokenizer_class": "BertJapaneseTokenizer",
+  "unk_token": "[UNK]",
+  "word_tokenizer_type": "mecab"
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1e9d53471fe8d498617107aa56613e484338c9a103459729aceae427b5c8dc1
+size 6776