Add dataset card with curriculum learning information
Browse files
README.md
CHANGED
@@ -1,73 +1,39 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
- name: valence_icc
|
41 |
-
dtype: float64
|
42 |
-
- name: arousal_icc
|
43 |
-
dtype: float64
|
44 |
-
- name: dominance_icc
|
45 |
-
dtype: float64
|
46 |
-
- name: n_categorical_evaluators
|
47 |
-
dtype: int64
|
48 |
-
- name: n_dimensional_evaluators
|
49 |
-
dtype: int64
|
50 |
-
- name: consensus_valence
|
51 |
-
dtype: float64
|
52 |
-
- name: consensus_arousal
|
53 |
-
dtype: float64
|
54 |
-
- name: consensus_dominance
|
55 |
-
dtype: float64
|
56 |
-
- name: naturalness_std
|
57 |
-
dtype: float64
|
58 |
-
- name: naturalness_icc
|
59 |
-
dtype: float64
|
60 |
-
- name: consensus_naturalness
|
61 |
-
dtype: float64
|
62 |
-
splits:
|
63 |
-
- name: train
|
64 |
-
num_bytes: 2803902170.0
|
65 |
-
num_examples: 7798
|
66 |
-
download_size: 2739152711
|
67 |
-
dataset_size: 2803902170.0
|
68 |
-
configs:
|
69 |
-
- config_name: default
|
70 |
-
data_files:
|
71 |
-
- split: train
|
72 |
-
path: data/train-*
|
73 |
-
---
|
|
|
1 |
+
|
2 |
+
# IEMOCAP with Curriculum Learning Metrics
|
3 |
+
|
4 |
+
This dataset enhances the original IEMO_WAV_Diff_2 dataset with inter-evaluator agreement metrics
|
5 |
+
for curriculum learning following Lotfian & Busso (2019).
|
6 |
+
|
7 |
+
## Additional Columns
|
8 |
+
|
9 |
+
- `curriculum_order`: Training order (1=highest agreement, train first)
|
10 |
+
- `overall_agreement`: Combined agreement score (0-1, higher is better)
|
11 |
+
- `fleiss_kappa`: Categorical agreement (-1 to 1, higher is better)
|
12 |
+
- `krippendorff_alpha`: Krippendorff's alpha for categorical reliability
|
13 |
+
- `valence_std`, `arousal_std`, `dominance_std`: Standard deviation of dimensional ratings (lower is better)
|
14 |
+
- `valence_icc`, `arousal_icc`, `dominance_icc`: Intraclass correlation coefficients (0-1, higher is better)
|
15 |
+
- `n_categorical_evaluators`, `n_dimensional_evaluators`: Number of evaluators
|
16 |
+
- `consensus_valence`, `consensus_arousal`, `consensus_dominance`: Consensus dimensional ratings
|
17 |
+
|
18 |
+
## Usage for Curriculum Learning
|
19 |
+
|
20 |
+
Sort samples by `curriculum_order` and train on high-agreement samples first:
|
21 |
+
|
22 |
+
```python
|
23 |
+
from datasets import load_dataset
|
24 |
+
|
25 |
+
dataset = load_dataset("cairocode/MSPI_WAV_Diff_Curriculum")
|
26 |
+
train_data = dataset["train"].sort("curriculum_order")
|
27 |
+
|
28 |
+
# Start with high agreement samples
|
29 |
+
easy_samples = train_data.filter(lambda x: x["overall_agreement"] > 0.5)
|
30 |
+
hard_samples = train_data.filter(lambda x: x["overall_agreement"] < 0.5)
|
31 |
+
```
|
32 |
+
|
33 |
+
## Citation
|
34 |
+
|
35 |
+
If you use this dataset, please cite:
|
36 |
+
|
37 |
+
- Original IEMOCAP: Busso et al. (2008)
|
38 |
+
- Curriculum learning approach: Lotfian & Busso (2019)
|
39 |
+
- Original dataset: cairocode/IEMO_WAV_Diff_2
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|