Skip to content

Commit 877a9a6

Browse files
authored
update model versions (#6329)
* update model versions * change from CDF to calibration
1 parent 2209844 commit 877a9a6

File tree

2 files changed

+47
-54
lines changed

2 files changed

+47
-54
lines changed

Orchestrator/docs/NLRModels.md

Lines changed: 27 additions & 40 deletions
Original file line numberDiff line numberDiff line change
@@ -1,58 +1,40 @@
11
# Prebuilt Language Models
22

3-
Prebuilt language models have been trained towards more sophisticated tasks for both monolingual as well as multilingual scenarios, including intent prediction and entity extraction.
4-
Entity extraction is currently experimental and not yet readt for production use.
3+
Prebuilt language models have been trained towards more sophisticated tasks for both monolingual as well as multilingual scenarios, including intent prediction and entity extraction. Entity extraction is currently experimental and not yet ready for production use.
54

65
The following prebuilt language models are now available in [versions repository][2].
76

87
See the [References](#references) section for technical descriptions of the AI technology behind the models.
98

10-
See the [References](#references) section for technical descriptions of the AI technology behind the models .
11-
129
## Default Models
1310

1411
### pretrained.20200924.microsoft.dte.00.06.en.onnx
15-
This is a high quality EN-only base model for intent detection that strikes the balance between size,
16-
speed and predictive performance.
17-
It is a 6-layer pretrained [Transformer][7] model optimized for conversation.
18-
Its architecture is pretrained for example-based use ([KNN][3]),
19-
thus it can be used out of box. This is the default model used if none explicitly specified.
12+
This is a high quality EN-only base model for intent detection that strikes the balance between size, speed and predictive performance. It is a 6-layer pretrained [Transformer][7] model optimized for conversation. Its architecture is pretrained for example-based use ([KNN][3]), thus it can be used out of box. This is the default model used if none explicitly specified.
2013

2114
### pretrained.20210205.microsoft.dte.00.06.unicoder_multilingual.onnx
22-
This is a high quality multilingual base model for intent detection. It's smaller and faster than its 12-layer alternative.
23-
It is a 6-layer pretrained pretrained [Transformer][7] model optimized for conversation.
24-
Its architecture is pretrained for example-based use ([KNN][3]), thus it can be used out of box. The model supports in total 100 languages (full list can be found at [XLMR Supported Languages][8]). 8 languages (EN, ES, DE, FR, IT, JA, PT, and ZH) are fine-tuned with additional data (performance can be found [here](#multilingual-intent-detection-models-evaluation)).
15+
This is a high quality multilingual base model for intent detection. It's smaller and faster than its 12-layer alternative. It is a 6-layer pretrained [Transformer][7] model optimized for conversation. Its architecture is pretrained for example-based use ([KNN][3]), thus it can be used out of box. The model supports in total 100 languages (full list can be found at [XLMR Supported Languages][8]). 8 languages (EN, ES, DE, FR, IT, JA, PT, and ZH) are fine-tuned with additional data (performance can be found [here](#multilingual-intent-detection-models-evaluation)).
2516

26-
### pretrained.20210218.microsoft.dte.00.06.bert_example_ner.en.onnx (experimental)
27-
This is a high quality EN-only base model for entity extraction. It's smaller and faster than its 12-layer alternative.
28-
It is a 6-layer pretrained pretrained [Transformer][7] model optimized for conversation.
29-
Its architecture is pretrained for example-based use ([KNN][3]), thus it can be used out of box.
17+
### pretrained.20210401.microsoft.dte.00.06.bert_example_ner_addon_free.en.onnx (experimental)
18+
This is a high quality EN-only base model for entity extraction. It is a 6-layer pretrained [Transformer][7] model optimized for conversation. Its architecture is pretrained for example-based use ([KNN][3]), thus it can be used out of box.
3019

3120
## Alternate Models
3221

3322
### pretrained.20200924.microsoft.dte.00.03.en.onnx
34-
This is a fast and small EN-only base model for intent detection with sufficient prediction performance.
35-
We suggest using this model if speed and memory size is critical to your deployment environment,
36-
otherwise consider other options. It is a generic 3-layer pretrained
37-
[Transformer][7] model optimized for conversation.
38-
Its architecture is pretrained for example-based use ([KNN][3]), thus it can be used out of box.
23+
This is a fast and small EN-only base model for intent detection with sufficient prediction performance. We suggest using this model if speed and memory size is critical to your deployment environment, otherwise consider other options. It is a generic 3-layer pretrained [Transformer][7] model optimized for conversation. Its architecture is pretrained for example-based use ([KNN][3]), thus it can be used out of box.
3924

4025
### pretrained.20200924.microsoft.dte.00.12.en.onnx
41-
This is a high quality EN-only base model for intent detection, but is larger and slower than other options.
42-
It is a 12-layer pretrained pretrained [Transformer][7] model optimized for conversation.
43-
Its architecture is pretrained for example-based use ([KNN][3]), thus it can be used out of box.
26+
This is a high quality EN-only base model for intent detection, but is larger and slower than other options. It is a 12-layer pretrained [Transformer][7] model optimized for conversation. Its architecture is pretrained for example-based use ([KNN][3]), thus it can be used out of box.
27+
28+
### pretrained.20210521.microsoft.dte.01.06.int.en.onnx
29+
This is a high quality quantized EN-only base model for intent detection, and it is smaller and faster than other options. It is a 6-layer pretrained [Transformer][7] model optimized for conversation. Its architecture is pretrained for example-based use ([KNN][3]), thus it can be used out of box.
4430

4531
### pretrained.20201210.microsoft.dte.00.12.unicoder_multilingual.onnx
46-
This is a high quality multilingual base model for intent detection.
47-
It is a 12-layer pretrained pretrained [Transformer][7] model optimized for conversation.
32+
This is a high quality multilingual base model for intent detection. It is a 12-layer pretrained [Transformer][7] model optimized for conversation.
4833
Its architecture is pretrained for example-based use ([KNN][3]), thus it can be used out of box. The model supports in total 100 languages (full list can be found at [XLMR Supported Languages][8]). 8 languages (EN, ES, DE, FR, IT, JA, PT, and ZH) are fine-tuned with additional data (performance can be found [here](#multilingual-intent-detection-models-evaluation)).
4934

50-
## Experimental Models
35+
### pretrained.20210608.microsoft.dte.01.06.int.unicoder_multilingual.onnx
36+
This is a high quality quantized multilingual base model for intent detection. It is a 6-layer pretrained [Transformer][7] model optimized for conversation. Its architecture is pretrained for example-based use ([KNN][3]), thus it can be used out of box. The model supports in total 100 languages (full list can be found at [XLMR Supported Languages][8]). 8 languages (EN, ES, DE, FR, IT, JA, PT, and ZH) are fine-tuned with additional data (performance can be found [here](#multilingual-intent-detection-models-evaluation)).
5137

52-
### pretrained.20210218.microsoft.dte.00.12.bert_example_ner.en.onnx (experimental)
53-
This is a yet another high quality EN-only base model for entity extraction.
54-
It is a 12-layer pretrained pretrained [Transformer][7] model optimized for conversation.
55-
Its architecture is pretrained for example-based use ([KNN][3]), thus it can be used out of box.
5638

5739
## Models Evaluation
5840
For a more quantitative comparison analysis of the different models see the following performance characteristics.
@@ -64,9 +46,10 @@ For a more quantitative comparison analysis of the different models see the foll
6446

6547
| Model |Base Model |Layers |Encoding time per query | Disk Allocation |
6648
| ------------ | ------------ | ------------ | ------------ | ------------ |
67-
|pretrained.20200924.microsoft.dte.00.03.en.onnx | BERT | 3 | ~ 7 ms | 164M |
68-
|pretrained.20200924.microsoft.dte.00.06.en.onnx | BERT | 6 | ~ 16 ms | 261M |
69-
|pretrained.20200924.microsoft.dte.00.12.en.onnx | BERT | 12 | ~ 26 ms | 427M |
49+
|pretrained.20200924.microsoft.dte.00.03.en.onnx | BERT | 3 | ~ 7 ms | 164M |
50+
|pretrained.20200924.microsoft.dte.00.06.en.onnx | BERT | 6 | ~ 14 ms | 261M |
51+
|pretrained.20200924.microsoft.dte.00.12.en.onnx | BERT | 12 | ~ 26 ms | 427M |
52+
|pretrained.20210521.microsoft.dte.01.06.int.en.onnx | BERT | 6 | ~ 6 ms | 65M |
7053

7154
-
7255
The following table shows how accurate is each model relative to provided training sample size using [Snips NLU][4] system, evaluated by **micro-average-accuracy**.
@@ -77,44 +60,48 @@ For a more quantitative comparison analysis of the different models see the foll
7760
|pretrained.20200924.microsoft.dte.00.03.en.onnx | 0.756 | 0.839 | 0.904 | 0.929 | 0.943 | 0.951 |
7861
|pretrained.20200924.microsoft.dte.00.06.en.onnx | 0.924 | 0.940 | 0.957 | 0.960 | 0.966 | 0.969 |
7962
|pretrained.20200924.microsoft.dte.00.12.en.onnx | 0.902 | 0.931 | 0.951 | 0.960 | 0.964 | 0.969 |
63+
|pretrained.20210521.microsoft.dte.01.06.int.en.onnx | 0.917 | 0.939 | 0.951 | 0.958 | 0.963 | 0.965 |
64+
8065

8166
### Multilingual Intent Detection Models Evaluation
8267
- The following table shows the size & speed performance attributes.
8368

8469
| Model | Base Model | Layers | Encoding time per query | Disk Allocation |
8570
| ------------------------------------------------------------ | ---------- | ------ | ----------------------- | --------------- |
86-
| pretrained.20210205.microsoft.dte.00.06.unicoder_multilingual.onnx | Unicoder | 6 | ~ 16 ms | 896M |
87-
| pretrained.20201210.microsoft.dte.00.12.unicoder_multilingual.onnx | Unicoder | 12 | ~ 30 ms | 1.08G |
71+
| pretrained.20210205.microsoft.dte.00.06.unicoder_multilingual.onnx | Unicoder | 6 | ~ 9 ms | 918M |
72+
| pretrained.20201210.microsoft.dte.00.12.unicoder_multilingual.onnx | Unicoder | 12 | ~ 16 ms | 1.08G |
73+
| pretrained.20210608.microsoft.dte.01.06.int.unicoder_multilingual.onnx | Unicoder | 6 | ~ 4 ms | 230M |
8874

8975
- The following table shows how accurate is each model by training and testing on the same language, evaluated by **micro-average-accuracy** on an internal dataset.
9076

9177
| Model | de-de | en-us | es-es | es-mx | fr-ca | fr-fr | it-it | ja-jp | pt-br | zh-cn |
9278
| ------------------------------------------------------------ | ----- | ----- | ----- | ----- | ----- | ----- | ----- | ----- | ----- | ----- |
9379
| pretrained.20210205.microsoft.dte.00.06.unicoder_multilingual.onnx | 0.638 | 0.785 | 0.662 | 0.760 | 0.723 | 0.661 | 0.701 | 0.786 | 0.735 | 0.805 |
9480
| pretrained.20201210.microsoft.dte.00.12.unicoder_multilingual.onnx | 0.642 | 0.764 | 0.646 | 0.754 | 0.722 | 0.636 | 0.689 | 0.789 | 0.725 | 0.809 |
81+
| pretrained.20210608.microsoft.dte.01.06.int.unicoder_multilingual.onnx | 0.634 | 0.765 | 0.657 | 0.743 | 0.715 | 0.646 | 0.697 | 0.780 | 0.743 | 0.799 |
9582

9683
- The following table shows how accurate is each model by training on **en-us** and testing on the different languages, evaluated by **micro-average-accuracy** on an internal dataset.
9784

9885
| Model | de-de | en-us | es-es | es-mx | fr-ca | fr-fr | it-it | ja-jp | pt-br | zh-cn |
9986
| ------------------------------------------------------------ | ----- | ----- | ----- | ----- | ----- | ----- | ----- | ----- | ----- | ----- |
10087
| pretrained.20210205.microsoft.dte.00.06.unicoder_multilingual.onnx | 0.495 | 0.785 | 0.530 | 0.621 | 0.560 | 0.518 | 0.546 | 0.663 | 0.568 | 0.687 |
10188
| pretrained.20201210.microsoft.dte.00.12.unicoder_multilingual.onnx | 0.499 | 0.764 | 0.529 | 0.604 | 0.562 | 0.515 | 0.547 | 0.646 | 0.555 | 0.681 |
89+
| pretrained.20210608.microsoft.dte.01.06.int.unicoder_multilingual.onnx | 0.496 | 0.765 | 0.529 | 0.623 | 0.562 | 0.511 | 0.540 | 0.670 | 0.579 | 0.692 |
10290

10391
### English Entity Extraction Models Evaluation
10492

10593
- The following table shows the size & speed performance attributes.
10694

10795
| Model | Base Model | Layers | Encoding time per query | Disk Allocation |
10896
| ------------------------------------------------------------ | ---------- | ------ | ----------------------- | --------------- |
109-
| pretrained.20210218.microsoft.dte.00.06.bert_example_ner.en.onnx | BERT | 6 | ~ 23 ms | 259M |
110-
| pretrained.20210218.microsoft.dte.00.12.bert_example_ner.en.onnx | BERT | 12 | ~ 40 ms | 425M |
97+
| pretrained.20210401.microsoft.dte.00.06.bert_example_ner_addon_free.en.onnx | TNLR | 6 | ~ 29 ms | 253M |
11198

11299
- The following table shows how accurate is each model relative to provided training sample size using [Snips NLU][4] system, evaluated by **macro-average-F1**.
113100

114101
| Training samples per entity type | 10 | 20 | 50 | 100 | 200 |
115102
| ------------------------------------------------------------ | ----- | ----- | ----- | ----- | ----- |
116-
| pretrained.20210218.microsoft.dte.00.06.bert_example_ner.en.onnx | 0.637 | 0.658 | 0.673 | 0.686 | 0.684 |
117-
| pretrained.20210218.microsoft.dte.00.12.bert_example_ner.en.onnx | 0.661 | 0.664 | 0.670 | 0.685 | 0.681 |
103+
| pretrained.20210401.microsoft.dte.00.06.bert_example_ner_addon_free.en.onnx | 0.702 | 0.712 | 0.731 | 0.752 | 0.739 |
104+
118105

119106

120107

Orchestrator/v0.2/nlr_versions.json

Lines changed: 20 additions & 14 deletions
Original file line numberDiff line numberDiff line change
@@ -3,7 +3,7 @@
33
"defaults": {
44
"en_intent": "pretrained.20200924.microsoft.dte.00.06.en.onnx",
55
"multilingual_intent": "pretrained.20210205.microsoft.dte.00.06.unicoder_multilingual.onnx",
6-
"en_entity": "pretrained.20210218.microsoft.dte.00.06.bert_example_ner.en.onnx"
6+
"en_entity": "pretrained.20210401.microsoft.dte.00.06.bert_example_ner_addon_free.en.onnx"
77
},
88
"models": {
99
"pretrained.20200924.microsoft.dte.00.03.en.onnx": {
@@ -24,10 +24,16 @@
2424
"description": "Bot Framework SDK release 4.10 - English ONNX V1.4 12-layer per-token intent base model",
2525
"minSDKVersion": "4.10.0"
2626
},
27-
"pretrained.20210218.microsoft.dte.00.12.bert_example_ner.en.onnx": {
28-
"releaseDate": "02/18/2021",
29-
"modelUri": "https://aka.ms/pretrained.20210218.microsoft.dte.00.12.bert_example_ner.en.onnx",
30-
"description": "(experimental) Bot Framework SDK release 4.12 - English ONNX V1.4 12-layer per-token entity base model",
27+
"pretrained.20210521.microsoft.dte.01.06.int.en.onnx": {
28+
"releaseDate": "06/14/2021",
29+
"modelUri": "https://models.botframework.com/models/dte/onnx/pretrained.20210521.microsoft.dte.01.06.int.en.onnx.zip",
30+
"description": "Bot Framework SDK release 4.10 - English ONNX V1.4 6-layer int per-token intent base model with updated calibration thresholds",
31+
"minSDKVersion": "4.10.0"
32+
},
33+
"pretrained.20210205.microsoft.dte.00.06.unicoder_multilingual.onnx": {
34+
"releaseDate": "02/05/2021",
35+
"modelUri": "https://aka.ms/pretrained.20210205.microsoft.dte.00.06.unicoder_multilingual.onnx",
36+
"description": "Bot Framework SDK release 4.12 - Multilingual ONNX V1.4 6-layer per-token intent base model",
3137
"minSDKVersion": "4.12.0"
3238
},
3339
"pretrained.20201210.microsoft.dte.00.12.unicoder_multilingual.onnx": {
@@ -36,17 +42,17 @@
3642
"description": "Bot Framework SDK release 4.12 - Multilingual ONNX V1.4 12-layer per-token intent base model",
3743
"minSDKVersion": "4.12.0"
3844
},
39-
"pretrained.20210218.microsoft.dte.00.06.bert_example_ner.en.onnx": {
40-
"releaseDate": "02/18/2021",
41-
"modelUri": "https://aka.ms/pretrained.20210218.microsoft.dte.00.06.bert_example_ner.en.onnx",
42-
"description": "(experimental) Bot Framework SDK release 4.12 - English ONNX V1.4 6-layer per-token entity base model",
45+
"pretrained.20210608.microsoft.dte.01.06.int.unicoder_multilingual.onnx": {
46+
"releaseDate": "06/14/2021",
47+
"modelUri": "https://models.botframework.com/models/dte/onnx/pretrained.20210608.microsoft.dte.01.06.int.unicoder_multilingual.onnx.zip",
48+
"description": "Bot Framework SDK release 4.12 - Multilingual ONNX V1.4 6-layer int per-token intent base model with updated calibration thresholds",
4349
"minSDKVersion": "4.12.0"
4450
},
45-
"pretrained.20210205.microsoft.dte.00.06.unicoder_multilingual.onnx": {
46-
"releaseDate": "02/05/2021",
47-
"modelUri": "https://aka.ms/pretrained.20210205.microsoft.dte.00.06.unicoder_multilingual.onnx",
48-
"description": "Bot Framework SDK release 4.12 - Multilingual ONNX V1.4 6-layer per-token intent base model",
49-
"minSDKVersion": "4.12.0"
51+
"pretrained.20210401.microsoft.dte.00.06.bert_example_ner_addon_free.en.onnx": {
52+
"releaseDate": "04/01/2021",
53+
"modelUri": "https://models.botframework.com/models/dte/onnx/pretrained.20210401.microsoft.dte.00.06.bert_example_ner_addon_free.en.onnx.zip",
54+
"description": "(experimental) Bot Framework SDK release 4.14 - English ONNX V1.4 6-layer per-token entity base model",
55+
"minSDKVersion": "4.14.0"
5056
}
5157
}
5258
}

0 commit comments

Comments
 (0)