Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
@@ -24,10 +24,10 @@ model-index:
|
|
24 |
metrics:
|
25 |
- name: Test RMSE
|
26 |
type: rmse
|
27 |
-
value: 0.
|
28 |
- name: Test R²
|
29 |
type: r2
|
30 |
-
value: 0.
|
31 |
- name: Test Loss
|
32 |
type: loss
|
33 |
value: 0.0002
|
@@ -35,35 +35,83 @@ model-index:
|
|
35 |
|
36 |
# Topic Drift Detector Model
|
37 |
|
38 |
-
## Version:
|
39 |
|
40 |
-
This model detects topic drift in conversations using an enhanced attention-based architecture. Trained on the [leonvanbokhorst/topic-drift-v2](https://huggingface.co/datasets/leonvanbokhorst/topic-drift-v2) dataset.
|
41 |
|
42 |
## Model Architecture
|
43 |
-
- Multi-head attention mechanism (4 heads)
|
44 |
-
-
|
45 |
-
-
|
46 |
-
-
|
47 |
-
-
|
48 |
-
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
49 |
|
50 |
## Performance Metrics
|
51 |
```txt
|
52 |
=== Full Training Results ===
|
53 |
-
Best Validation RMSE: 0.
|
54 |
-
Best Validation R²: 0.
|
55 |
|
56 |
=== Test Set Results ===
|
57 |
Loss: 0.0002
|
58 |
-
RMSE: 0.
|
59 |
-
R²: 0.
|
60 |
-
|
61 |
```
|
62 |
|
63 |
-
## Training
|
64 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
65 |
|
66 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
67 |
```python
|
68 |
import torch
|
69 |
from transformers import AutoModel, AutoTokenizer
|
@@ -73,7 +121,7 @@ base_model = AutoModel.from_pretrained('BAAI/bge-m3')
|
|
73 |
tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-m3')
|
74 |
|
75 |
# Load topic drift detector
|
76 |
-
model = torch.load('models/
|
77 |
model.eval()
|
78 |
|
79 |
# Prepare conversation window (8 turns)
|
@@ -103,18 +151,22 @@ print(f"Topic drift score: {drift_scores.item():.4f}")
|
|
103 |
# Higher scores indicate more topic drift
|
104 |
```
|
105 |
|
106 |
-
##
|
107 |
-
|
108 |
-
-
|
109 |
-
-
|
110 |
-
-
|
111 |
-
|
112 |
-
|
113 |
-
-
|
114 |
-
-
|
115 |
|
116 |
## Limitations
|
117 |
- Works best with English conversations
|
118 |
- Requires exactly 8 turns of conversation
|
119 |
- Each turn should be between 1-512 tokens
|
120 |
- Relies on BAAI/bge-m3 embeddings
|
|
|
|
|
|
|
|
|
|
24 |
metrics:
|
25 |
- name: Test RMSE
|
26 |
type: rmse
|
27 |
+
value: 0.0144
|
28 |
- name: Test R²
|
29 |
type: r2
|
30 |
+
value: 0.8666
|
31 |
- name: Test Loss
|
32 |
type: loss
|
33 |
value: 0.0002
|
|
|
35 |
|
36 |
# Topic Drift Detector Model
|
37 |
|
38 |
+
## Version: v20241225_184257
|
39 |
|
40 |
+
This model detects topic drift in conversations using an enhanced hierarchical attention-based architecture. Trained on the [leonvanbokhorst/topic-drift-v2](https://huggingface.co/datasets/leonvanbokhorst/topic-drift-v2) dataset.
|
41 |
|
42 |
## Model Architecture
|
43 |
+
- Multi-head attention mechanism (4 heads, head dimension 128)
|
44 |
+
- Hierarchical pattern detection with multi-scale analysis
|
45 |
+
- Explicit transition point detection with linguistic markers
|
46 |
+
- Pattern-aware self-attention mechanism
|
47 |
+
- Dynamic window augmentation
|
48 |
+
- Contrastive learning with pattern-aware sampling
|
49 |
+
- Adversarial training with pattern-aware perturbations
|
50 |
+
|
51 |
+
### Key Components:
|
52 |
+
1. **Embedding Processor**:
|
53 |
+
- Input dimension: 1024
|
54 |
+
- Hidden dimension: 512
|
55 |
+
- Dropout rate: 0.35
|
56 |
+
- PreNorm layers with residual connections
|
57 |
+
|
58 |
+
2. **Attention Blocks**:
|
59 |
+
- 3 layers of attention
|
60 |
+
- 4 attention heads
|
61 |
+
- Feed-forward dimension: 2048
|
62 |
+
- Learned position encodings
|
63 |
+
|
64 |
+
3. **Pattern Detection**:
|
65 |
+
- Hierarchical LSTM layers
|
66 |
+
- Bidirectional processing
|
67 |
+
- Multi-scale pattern analysis
|
68 |
+
- Pattern classification with 7 types
|
69 |
+
|
70 |
+
4. **Transition Detection**:
|
71 |
+
- Linguistic marker attention
|
72 |
+
- Explicit transition scoring
|
73 |
+
- Marker-based context integration
|
74 |
|
75 |
## Performance Metrics
|
76 |
```txt
|
77 |
=== Full Training Results ===
|
78 |
+
Best Validation RMSE: 0.0142
|
79 |
+
Best Validation R²: 0.8711
|
80 |
|
81 |
=== Test Set Results ===
|
82 |
Loss: 0.0002
|
83 |
+
RMSE: 0.0144
|
84 |
+
R²: 0.8666
|
|
|
85 |
```
|
86 |
|
87 |
+
## Training Details
|
88 |
+
- Dataset: 6400 conversations (5120 train, 640 val, 640 test)
|
89 |
+
- Window size: 8 turns
|
90 |
+
- Batch size: 32
|
91 |
+
- Learning rate: 0.0001 with cosine decay
|
92 |
+
- Warmup steps: 100
|
93 |
+
- Early stopping patience: 15
|
94 |
+
- Max gradient norm: 1.0
|
95 |
+
- Mixed precision training (AMP)
|
96 |
+
- Base embeddings: BAAI/bge-m3
|
97 |
|
98 |
+
### Training Enhancements:
|
99 |
+
1. **Dynamic Window Augmentation**:
|
100 |
+
- Adaptive window sizes
|
101 |
+
- Interpolation-based resizing
|
102 |
+
- Maintains temporal consistency
|
103 |
+
|
104 |
+
2. **Contrastive Learning**:
|
105 |
+
- Pattern-aware positive/negative sampling
|
106 |
+
- Temperature-scaled similarities
|
107 |
+
- Weighted combination of embeddings
|
108 |
+
|
109 |
+
3. **Adversarial Training**:
|
110 |
+
- Pattern-aware perturbations
|
111 |
+
- Self-distillation loss
|
112 |
+
- Epsilon ball projection
|
113 |
+
|
114 |
+
## Usage Example
|
115 |
```python
|
116 |
import torch
|
117 |
from transformers import AutoModel, AutoTokenizer
|
|
|
121 |
tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-m3')
|
122 |
|
123 |
# Load topic drift detector
|
124 |
+
model = torch.load('models/v20241225_184257/topic_drift_model.pt')
|
125 |
model.eval()
|
126 |
|
127 |
# Prepare conversation window (8 turns)
|
|
|
151 |
# Higher scores indicate more topic drift
|
152 |
```
|
153 |
|
154 |
+
## Pattern Types
|
155 |
+
The model detects 7 distinct pattern types:
|
156 |
+
1. "maintain" - No significant drift
|
157 |
+
2. "gentle_wave" - Subtle topic evolution
|
158 |
+
3. "single_peak" - One clear transition
|
159 |
+
4. "multi_peak" - Multiple transitions
|
160 |
+
5. "ascending" - Gradually increasing drift
|
161 |
+
6. "descending" - Gradually decreasing drift
|
162 |
+
7. "abrupt" - Sudden topic change
|
163 |
|
164 |
## Limitations
|
165 |
- Works best with English conversations
|
166 |
- Requires exactly 8 turns of conversation
|
167 |
- Each turn should be between 1-512 tokens
|
168 |
- Relies on BAAI/bge-m3 embeddings
|
169 |
+
- May be sensitive to conversation style variations
|
170 |
+
|
171 |
+
## Training Curves
|
172 |
+
![Training Curves](plots/v20241225_184257/training_curves.png)
|