Commit
•
ec6ecbd
1
Parent(s):
e510cdb
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,70 @@
|
|
| 1 |
---
|
|
|
|
|
|
|
| 2 |
license: apache-2.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
language:
|
| 3 |
+
- en
|
| 4 |
license: apache-2.0
|
| 5 |
+
tags:
|
| 6 |
+
- roberta
|
| 7 |
+
- classification
|
| 8 |
+
- dialog state tracking
|
| 9 |
+
- natural language understanding
|
| 10 |
+
- uncertainty
|
| 11 |
+
- conversational system
|
| 12 |
+
- task-oriented dialog
|
| 13 |
+
datasets:
|
| 14 |
+
- ConvLab/multiwoz21
|
| 15 |
+
metrics:
|
| 16 |
+
- Joint Goal Accuracy
|
| 17 |
+
- Slot F1
|
| 18 |
+
- Joint Goal Expected Calibration Error
|
| 19 |
+
|
| 20 |
+
model-index:
|
| 21 |
+
- name: setsumbt-dst-nlu-multiwoz21
|
| 22 |
+
results:
|
| 23 |
+
- task:
|
| 24 |
+
type: classification
|
| 25 |
+
name: dialog state tracking
|
| 26 |
+
dataset:
|
| 27 |
+
type: ConvLab/multiwoz21
|
| 28 |
+
name: MultiWOZ21
|
| 29 |
+
split: test
|
| 30 |
+
metrics:
|
| 31 |
+
- type: Joint Goal Accuracy
|
| 32 |
+
value: 51.8
|
| 33 |
+
name: JGA
|
| 34 |
+
- type: Slot F1
|
| 35 |
+
value: 91.1
|
| 36 |
+
name: Slot F1
|
| 37 |
+
- type: Joint Goal Expected Calibration Error
|
| 38 |
+
value: 12.7
|
| 39 |
+
name: JECE
|
| 40 |
+
|
| 41 |
---
|
| 42 |
+
|
| 43 |
+
# SetSUMBT-dst-nlu-multiwoz21
|
| 44 |
+
|
| 45 |
+
This model is a fine-tuned version [SetSUMBT](https://github.com/ConvLab/ConvLab-3/tree/master/convlab/dst/setsumbt) of [roberta-base](https://huggingface.co/roberta-base) on [MultiWOZ2.1](https://huggingface.co/datasets/ConvLab/multiwoz21).
|
| 46 |
+
This model is a combined DST and NLU model and is a distribution distilled version of a ensemble of 5 models. This model should be used to produce uncertainty estimates for the dialogue belief state.
|
| 47 |
+
|
| 48 |
+
Refer to [ConvLab-3](https://github.com/ConvLab/ConvLab-3) for model description and usage.
|
| 49 |
+
|
| 50 |
+
## Training procedure
|
| 51 |
+
|
| 52 |
+
### Training hyperparameters
|
| 53 |
+
|
| 54 |
+
The following hyperparameters were used during training:
|
| 55 |
+
- learning_rate: 0.00001
|
| 56 |
+
- train_batch_size: 3
|
| 57 |
+
- eval_batch_size: 16
|
| 58 |
+
- seed: 0
|
| 59 |
+
- gradient_accumulation_steps: 1
|
| 60 |
+
- optimizer: AdamW
|
| 61 |
+
- loss: Ensemble Distribution Distillation Loss
|
| 62 |
+
- lr_scheduler_type: linear
|
| 63 |
+
- num_epochs: 50.0
|
| 64 |
+
|
| 65 |
+
### Framework versions
|
| 66 |
+
|
| 67 |
+
- Transformers 4.17.0
|
| 68 |
+
- Pytorch 1.8.0+cu110
|
| 69 |
+
- Datasets 2.3.2
|
| 70 |
+
- Tokenizers 0.12.1
|