gomyk commited on
Commit
8b65e38
·
verified ·
1 Parent(s): d5fc3ce

Upload L4_uniform_distilled (distilled) from MiniLM-L12

Browse files
Files changed (1) hide show
  1. README.md +80 -0
README.md CHANGED
@@ -104,6 +104,86 @@ embeddings = model.encode(sentences)
104
  print(embeddings.shape) # (3, 384)
105
  ```
106
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
107
 
108
 
109
  ## Training
 
104
  print(embeddings.shape) # (3, 384)
105
  ```
106
 
107
+ ## MTEB Evaluation Results
108
+
109
+ **Overall Average: 54.62%**
110
+
111
+ | Task Group | Average |
112
+ |---|---|
113
+ | Classification | 58.95% |
114
+ | Clustering | 36.16% |
115
+ | STS | 67.19% |
116
+
117
+ ### Classification
118
+
119
+ | Task | Average | Details |
120
+ |---|---|---|
121
+ | AmazonCounterfactualClassification | 65.7% | en: 68.34%, de: 67.09%, en-ext: 66.57% |
122
+ | Banking77Classification | 78.43% | default: 78.43% |
123
+ | ImdbClassification | 60.43% | default: 60.43% |
124
+ | MTOPDomainClassification | 74.26% | en: 83.96%, es: 78.09%, fr: 76.42% |
125
+ | MassiveIntentClassification | 34.72% | en: 65.57%, zh-CN: 62.56%, ja: 62.19% |
126
+ | MassiveScenarioClassification | 40.37% | en: 71.61%, zh-CN: 69.22%, ja: 67.48% |
127
+ | ToxicConversationsClassification | 61.37% | default: 61.37% |
128
+ | TweetSentimentExtractionClassification | 56.29% | default: 56.29% |
129
+
130
+ ### Clustering
131
+
132
+ | Task | Average | Details |
133
+ |---|---|---|
134
+ | ArXivHierarchicalClusteringP2P | 51.36% | default: 51.36% |
135
+ | ArXivHierarchicalClusteringS2S | 46.3% | default: 46.3% |
136
+ | BiorxivClusteringP2P.v2 | 26.63% | default: 26.63% |
137
+ | MedrxivClusteringP2P.v2 | 30.15% | default: 30.15% |
138
+ | MedrxivClusteringS2S.v2 | 25.7% | default: 25.7% |
139
+ | StackExchangeClustering.v2 | 46.56% | default: 46.56% |
140
+ | StackExchangeClusteringP2P.v2 | 35.85% | default: 35.85% |
141
+ | TwentyNewsgroupsClustering.v2 | 26.74% | default: 26.74% |
142
+
143
+ ### STS
144
+
145
+ | Task | Average | Details |
146
+ |---|---|---|
147
+ | BIOSSES | 53.2% | default: 53.2% |
148
+ | SICK-R | 73.61% | default: 73.61% |
149
+ | STS12 | 72.11% | default: 72.11% |
150
+ | STS13 | 75.62% | default: 75.62% |
151
+ | STS14 | 72.2% | default: 72.2% |
152
+ | STS15 | 80.86% | default: 80.86% |
153
+ | STS17 | 48.49% | en-en: 81.73%, es-es: 75.65%, ar-ar: 57.76% |
154
+ | STS22.v2 | 49.17% | fr: 69.99%, es-en: 65.46%, zh: 63.87% |
155
+ | STSBenchmark | 79.48% | default: 79.48% |
156
+
157
+
158
+ ## Distillation Impact
159
+
160
+ | Task | Before | After | Delta |
161
+ |---|---|---|---|
162
+ | AmazonCounterfactualClassification | 67.02% | 65.7% | -1.32%p |
163
+ | ArXivHierarchicalClusteringP2P | 49.93% | 51.36% | +1.43%p |
164
+ | ArXivHierarchicalClusteringS2S | 46.08% | 46.3% | +0.22%p |
165
+ | Banking77Classification | 69.18% | 78.43% | +9.25%p |
166
+ | BiorxivClusteringP2P.v2 | 21.47% | 26.63% | +5.16%p |
167
+ | BIOSSES | 45.64% | 53.2% | +7.56%p |
168
+ | ImdbClassification | 59.38% | 60.43% | +1.05%p |
169
+ | MassiveIntentClassification | 36.9% | 34.72% | -2.18%p |
170
+ | MassiveScenarioClassification | 39.51% | 40.37% | +0.86%p |
171
+ | MedrxivClusteringP2P.v2 | 26.05% | 30.15% | +4.1%p |
172
+ | MedrxivClusteringS2S.v2 | 22.94% | 25.7% | +2.76%p |
173
+ | MTOPDomainClassification | 71.48% | 74.26% | +2.78%p |
174
+ | SICK-R | 62.01% | 73.61% | +11.6%p |
175
+ | StackExchangeClustering.v2 | 41.23% | 46.56% | +5.33%p |
176
+ | StackExchangeClusteringP2P.v2 | 32.19% | 35.85% | +3.66%p |
177
+ | STS12 | 57.85% | 72.11% | +14.26%p |
178
+ | STS13 | 65.48% | 75.62% | +10.14%p |
179
+ | STS14 | 60.39% | 72.2% | +11.81%p |
180
+ | STS15 | 73.93% | 80.86% | +6.93%p |
181
+ | STS17 | 46.29% | 48.49% | +2.2%p |
182
+ | STS22.v2 | 37.34% | 49.17% | +11.83%p |
183
+ | STSBenchmark | 65.38% | 79.48% | +14.1%p |
184
+ | ToxicConversationsClassification | 62.02% | 61.37% | -0.65%p |
185
+ | TweetSentimentExtractionClassification | 49.43% | 56.29% | +6.86%p |
186
+ | TwentyNewsgroupsClustering.v2 | 16.43% | 26.74% | +10.31%p |
187
 
188
 
189
  ## Training