andreaceto commited on
Commit
c236fd6
·
verified ·
1 Parent(s): fe0e2d5

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +188 -97
README.md CHANGED
@@ -7,104 +7,195 @@ tags:
7
  model-index:
8
  - name: schedulebot-nlu-engine
9
  results: []
 
 
 
 
10
  ---
11
-
12
  <!-- This model card has been generated automatically according to the information the Trainer had access to. You
13
  should probably proofread and complete it, then remove this comment. -->
14
 
15
- # schedulebot-nlu-engine
16
-
17
- This model is a fine-tuned version of [distilbert-base-uncased](https://huggingface.co/distilbert-base-uncased) on an unknown dataset.
18
- It achieves the following results on the evaluation set:
19
- - Loss: 0.3576
20
- - Intent Accuracy: 0.9110
21
- - Intent F1: 0.9109
22
- - Ner F1: 0.6897
23
-
24
- ## Model description
25
-
26
- More information needed
27
-
28
- ## Intended uses & limitations
29
-
30
- More information needed
31
-
32
- ## Training and evaluation data
33
-
34
- More information needed
35
-
36
- ## Training procedure
37
-
38
- ### Training hyperparameters
39
-
40
- The following hyperparameters were used during training:
41
- - learning_rate: 1e-06
42
- - train_batch_size: 32
43
- - eval_batch_size: 32
44
- - seed: 42
45
- - optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
46
- - lr_scheduler_type: linear
47
- - num_epochs: 50
48
-
49
- ### Training results
50
-
51
- | Training Loss | Epoch | Step | Validation Loss | Intent Accuracy | Intent F1 | Ner F1 |
52
- |:-------------:|:-----:|:----:|:---------------:|:---------------:|:---------:|:------:|
53
- | 1.115 | 1.0 | 64 | 0.7159 | 0.8059 | 0.8059 | 0.6490 |
54
- | 1.0456 | 2.0 | 128 | 0.6909 | 0.8105 | 0.8104 | 0.6545 |
55
- | 0.966 | 3.0 | 192 | 0.6603 | 0.8151 | 0.8141 | 0.6656 |
56
- | 0.924 | 4.0 | 256 | 0.6358 | 0.8265 | 0.8260 | 0.6711 |
57
- | 0.9042 | 5.0 | 320 | 0.6084 | 0.8242 | 0.8249 | 0.6765 |
58
- | 0.8602 | 6.0 | 384 | 0.5821 | 0.8379 | 0.8383 | 0.6743 |
59
- | 0.811 | 7.0 | 448 | 0.5633 | 0.8516 | 0.8508 | 0.6765 |
60
- | 0.8068 | 8.0 | 512 | 0.5532 | 0.8447 | 0.8457 | 0.6711 |
61
- | 0.772 | 9.0 | 576 | 0.5326 | 0.8516 | 0.8523 | 0.6765 |
62
- | 0.7349 | 10.0 | 640 | 0.5132 | 0.8676 | 0.8671 | 0.6765 |
63
- | 0.7047 | 11.0 | 704 | 0.5053 | 0.8630 | 0.8622 | 0.6798 |
64
- | 0.6994 | 12.0 | 768 | 0.4976 | 0.8630 | 0.8625 | 0.6809 |
65
- | 0.6831 | 13.0 | 832 | 0.4860 | 0.8630 | 0.8635 | 0.6820 |
66
- | 0.6509 | 14.0 | 896 | 0.4741 | 0.8858 | 0.8858 | 0.6809 |
67
- | 0.643 | 15.0 | 960 | 0.4616 | 0.8836 | 0.8833 | 0.6820 |
68
- | 0.6392 | 16.0 | 1024 | 0.4562 | 0.8858 | 0.8860 | 0.6832 |
69
- | 0.5852 | 17.0 | 1088 | 0.4452 | 0.8836 | 0.8833 | 0.6874 |
70
- | 0.5854 | 18.0 | 1152 | 0.4374 | 0.8813 | 0.8812 | 0.6885 |
71
- | 0.5917 | 19.0 | 1216 | 0.4358 | 0.8858 | 0.8851 | 0.6897 |
72
- | 0.5678 | 20.0 | 1280 | 0.4298 | 0.8927 | 0.8922 | 0.6875 |
73
- | 0.5684 | 21.0 | 1344 | 0.4229 | 0.9064 | 0.9061 | 0.6865 |
74
- | 0.5594 | 22.0 | 1408 | 0.4111 | 0.9041 | 0.9042 | 0.6853 |
75
- | 0.5556 | 23.0 | 1472 | 0.4058 | 0.8995 | 0.8995 | 0.6853 |
76
- | 0.521 | 24.0 | 1536 | 0.4058 | 0.8904 | 0.8902 | 0.6853 |
77
- | 0.5224 | 25.0 | 1600 | 0.3999 | 0.9041 | 0.9041 | 0.6853 |
78
- | 0.5269 | 26.0 | 1664 | 0.3923 | 0.9110 | 0.9109 | 0.6875 |
79
- | 0.4832 | 27.0 | 1728 | 0.3916 | 0.9064 | 0.9066 | 0.6853 |
80
- | 0.4938 | 28.0 | 1792 | 0.3908 | 0.9041 | 0.9041 | 0.6853 |
81
- | 0.4819 | 29.0 | 1856 | 0.3851 | 0.9064 | 0.9061 | 0.6853 |
82
- | 0.4781 | 30.0 | 1920 | 0.3916 | 0.9041 | 0.9041 | 0.6853 |
83
- | 0.4671 | 31.0 | 1984 | 0.3799 | 0.9041 | 0.9042 | 0.6875 |
84
- | 0.4689 | 32.0 | 2048 | 0.3810 | 0.9041 | 0.9042 | 0.6875 |
85
- | 0.4799 | 33.0 | 2112 | 0.3759 | 0.9064 | 0.9066 | 0.6897 |
86
- | 0.4691 | 34.0 | 2176 | 0.3699 | 0.9110 | 0.9107 | 0.6897 |
87
- | 0.4752 | 35.0 | 2240 | 0.3699 | 0.9087 | 0.9082 | 0.6918 |
88
- | 0.4574 | 36.0 | 2304 | 0.3676 | 0.9132 | 0.9132 | 0.6897 |
89
- | 0.4654 | 37.0 | 2368 | 0.3668 | 0.9132 | 0.9134 | 0.6918 |
90
- | 0.4416 | 38.0 | 2432 | 0.3663 | 0.9110 | 0.9112 | 0.6918 |
91
- | 0.4433 | 39.0 | 2496 | 0.3685 | 0.9110 | 0.9111 | 0.6897 |
92
- | 0.4202 | 40.0 | 2560 | 0.3625 | 0.9155 | 0.9156 | 0.6897 |
93
- | 0.4523 | 41.0 | 2624 | 0.3595 | 0.9178 | 0.9176 | 0.6897 |
94
- | 0.4325 | 42.0 | 2688 | 0.3589 | 0.9178 | 0.9176 | 0.6897 |
95
- | 0.4406 | 43.0 | 2752 | 0.3600 | 0.9087 | 0.9085 | 0.6897 |
96
- | 0.4348 | 44.0 | 2816 | 0.3594 | 0.9132 | 0.9130 | 0.6875 |
97
- | 0.4339 | 45.0 | 2880 | 0.3573 | 0.9178 | 0.9176 | 0.6897 |
98
- | 0.4259 | 46.0 | 2944 | 0.3576 | 0.9132 | 0.9131 | 0.6897 |
99
- | 0.4272 | 47.0 | 3008 | 0.3568 | 0.9155 | 0.9153 | 0.6897 |
100
- | 0.4276 | 48.0 | 3072 | 0.3568 | 0.9155 | 0.9153 | 0.6897 |
101
- | 0.4297 | 49.0 | 3136 | 0.3575 | 0.9132 | 0.9131 | 0.6897 |
102
- | 0.4223 | 50.0 | 3200 | 0.3576 | 0.9110 | 0.9109 | 0.6897 |
103
-
104
-
105
- ### Framework versions
106
-
107
- - Transformers 4.53.2
108
- - Pytorch 2.6.0+cu124
109
- - Datasets 4.0.0
110
- - Tokenizers 0.21.2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7
  model-index:
8
  - name: schedulebot-nlu-engine
9
  results: []
10
+ datasets:
11
+ - andreaceto/hasd
12
+ language:
13
+ - en
14
  ---
 
15
  <!-- This model card has been generated automatically according to the information the Trainer had access to. You
16
  should probably proofread and complete it, then remove this comment. -->
17
 
18
+ # Schedulebot-nlu-engine
19
+
20
+ ## Model Description
21
+
22
+ This model is a multi-task Natural Language Understanding (NLU) engine designed specifically for an appointment scheduling chatbot. It is fine-tuned from a **`distilbert-base-uncased`** backbone and is capable of performing two tasks simultaneously:
23
+
24
+ - **Intent Classification**: Identifying the user's primary goal (e.g., `schedule`, `cancel`).
25
+ - **Named Entity Recognition (NER)**: Extracting custom, domain-specific entities (e.g., `appointment_type`).
26
+
27
+ This model stands out due to its custom classification heads, which use a more complex architecture to improve performance on nuanced tasks.
28
+
29
+ ## Model Architecture
30
+
31
+ The model uses a standard `distilbert-base-uncased` model as its core feature extractor. Two custom classification "heads" are placed on top of this base to perform the downstream tasks.
32
+
33
+ - **Base Model**: `distilbert-base-uncased`
34
+ - **Classifier Heads**: each head is a Multi-Layer Perceptron (MLP) with the following structure to allow for more complex feature interpretation:
35
+ 1. A Linear layer projecting the transformer's output dimension (768) to an intermediate size (384).
36
+ 2. A GELU activation function.
37
+ 3. A Dropout layer with a rate of 0.3 for regularization.
38
+ 4. A final Linear layer projecting the intermediate size to the number of output labels for the specific task (intent or NER).
39
+
40
+ ## Intended Use
41
+
42
+ This model is intended to be the core NLU component of a conversational AI system for managing appointments.
43
+
44
+ For instructions on how to use the model check the [dedicated file](./how_to_use.md).
45
+
46
+ ## Training Data
47
+
48
+ The model was trained on the **HASD (Hybrid Appointment Scheduling Dataset)**, a custom dataset built specifically for this task.
49
+
50
+ - **Source**: The dataset is a hybrid of real-world conversational examples from `clinc/clinc_oos` (for simple intents) and synthetically generated, template-based examples for complex scheduling intents.
51
+ - **Balancing**: To combat class imbalance, intents sourced from `clinc/clinc_oos` were **down-sampled** to a maximum of **150 examples** each.
52
+ - **Augmentation**: To increase data diversity for complex intents (`schedule`, `reschedule`, etc.), **Contextual Word Replacement** was used. A `distilbert-base-uncased` model augmented the templates by replacing non-placeholder words with contextually relevant synonyms.
53
+
54
+ The dataset is available [here](https://huggingface.co/datasets/andreaceto/hasd).
55
+
56
+ ### Intents
57
+
58
+ The model is trained to recognize the following intents:
59
+ `schedule`, `reschedule`, `cancel`, `query_avail`, `greeting`, `positive_reply`, `negative_reply`, `bye`, `oos` (out-of-scope).
60
+
61
+ ### Entities
62
+
63
+ The model is trained to recognize the following custom named entities:
64
+ `practitioner_name`, `appointment_type`, `appointment_id`.
65
+
66
+ ## Training Procedure
67
+
68
+ The model was trained using a two-stage fine-tuning strategy to ensure stability and performance.
69
+
70
+ ### Stage 1: Training the Classifier Heads
71
+
72
+ - The `distilbert-base-uncased` base model was entirely **frozen**.
73
+ - Only the randomly initialized MLP heads for intent and NER classification were trained.
74
+
75
+ **Setup**:
76
+
77
+ ```python
78
+ # Define a data collator to handle padding for token classification
79
+ data_collator = DataCollatorForTokenClassification(tokenizer=tokenizer)
80
+ # Define Training Arguments
81
+ training_args = TrainingArguments(
82
+ output_dir="path/to/output_dir",
83
+ overwrite_output_dir=True,
84
+ num_train_epochs=200, # Training epochs
85
+ per_device_train_batch_size=32,
86
+ per_device_eval_batch_size=32,
87
+ learning_rate=1e-4, # Learning Rate
88
+ weight_decay=1e-5, # AdamW weight decay
89
+ logging_dir="path/to/logging_dir",
90
+ logging_strategy="epoch",
91
+ eval_strategy="epoch",
92
+ save_strategy="epoch",
93
+ load_best_model_at_end=True,
94
+ metric_for_best_model="eval_loss", # Focus on validation loss as the key metric
95
+ # --- Hub Arguments ---
96
+ push_to_hub=True,
97
+ hub_model_id=hub_model_id,
98
+ hub_strategy="end",
99
+ hub_token=hf_token,
100
+ report_to="tensorboard" # Tensorboard to monitor training
101
+ )
102
+ # Create the Trainer
103
+ trainer = Trainer(
104
+ model=model,
105
+ args=training_args,
106
+ train_dataset=processed_datasets["train"],
107
+ eval_dataset=processed_datasets["validation"],
108
+ processing_class=tokenizer,
109
+ data_collator=data_collator,
110
+ compute_metrics=compute_metrics, # Custom function (check how_to_use.md)
111
+ callbacks=[EarlyStoppingCallback(early_stopping_patience=10)]
112
+ )
113
+ ```
114
+
115
+ ### Stage 2: Fine-Tuning
116
+
117
+ - The DistilBERT backbone was entirely **unfrozen**.
118
+ - Using a very low LR allows the model to adapt even better to the new data while preserving the powerful, general-purpose knowledge.
119
+
120
+ **Setup**:
121
+
122
+ ```python
123
+ # Define Training Arguments
124
+ training_args = TrainingArguments(
125
+ output_dir="path/to/output_dir",
126
+ overwrite_output_dir=True,
127
+ num_train_epochs=50, # Fine-tuning epochs
128
+ per_device_train_batch_size=32,
129
+ per_device_eval_batch_size=32,
130
+ learning_rate=1e-6, # Learning Rate
131
+ weight_decay=1e-3, # AdamW weight decay
132
+ logging_dir="path/to/logging_dir",
133
+ logging_strategy="epoch",
134
+ eval_strategy="epoch",
135
+ save_strategy="epoch",
136
+ load_best_model_at_end=True,
137
+ metric_for_best_model="eval_loss", # Focus on NER F1 as the key metric
138
+ # --- Hub Arguments ---
139
+ push_to_hub=True,
140
+ hub_model_id=hub_model_id,
141
+ hub_strategy="end",
142
+ hub_token=hf_token,
143
+ report_to="tensorboard" # Tensorboard to monitor training
144
+ )
145
+ # Create the Trainer
146
+ trainer = Trainer(
147
+ model=model,
148
+ args=training_args,
149
+ train_dataset=processed_datasets["train"],
150
+ eval_dataset=processed_datasets["validation"],
151
+ processing_class=tokenizer,
152
+ data_collator=data_collator,
153
+ compute_metrics=compute_metrics, # Custom function (check how_to_use.md)
154
+ callbacks=[EarlyStoppingCallback(early_stopping_patience=5)]
155
+ )
156
+ ```
157
+ ## Evaluation
158
+
159
+ The model was evaluated on a held-out test set, and its performance was measured for both tasks.
160
+
161
+ ### Intent Classification Performance
162
+
163
+ | Intent | Precision | Recall | F1-Score | Support |
164
+ | --- | --- | --- | --- | --- |
165
+ | bye | 0.9500 | 0.8261 | 0.8837 | 23 |
166
+ | cancel | 0.9211 | 0.8434 | 0.8805 | 83 |
167
+ | greeting | 0.9545 | 0.9545 | 0.9545 | 22 |
168
+ |negative_reply | 0.9091 | 0.9091 | 0.9091 | 22 |
169
+ | oos | 1.0000 | 0.8696 | 0.9302 | 23 |
170
+ |positive_reply | 0.7407 | 0.9091 | 0.8163 | 22 |
171
+ | query_avail | 0.9620 | 0.9383 | 0.9500 | 81 |
172
+ | reschedule | 0.8506 | 0.8916 | 0.8706 | 83 |
173
+ | schedule | 0.8488 | 0.9125 | 0.8795 | 80 |
174
+ | --- | --- | --- | --- | ---- |
175
+ | **Accuracy** | | | **0.8952** | 439 |
176
+ | **Macro Avg** | **0.9041** | **0.8949** | **0.8972** | 439 |
177
+ | **Weighted Avg** | **0.8998** | **0.8952** | **0.8960** | 439 |
178
+
179
+ ### NER (Token Classification) Performance
180
+
181
+ | Entity | Precision | Recall | F1-Score | Support |
182
+ | --- | --- | --- | --- | --- |
183
+ | B-appointment_id | 1.0000 | 1.0000 | 1.0000 | 61 |
184
+ | B-appointment_type | 0.8646 | 0.7477 | 0.8019 | 111 |
185
+ | B-practitioner_name | 0.9161 | 0.9467 | 0.9311 | 150 |
186
+ | B-appointment_id | 0.9667 | 0.9667 | 0.9667 | 210 |
187
+ | I-appointment_type | 0.8182 | 0.7368 | 0.7754 | 171 |
188
+ | I-practitioner_name | 0.9540 | 0.8941 | 0.9231 | 255 |
189
+ | O | 0.9782 | 0.9892 | 0.9837 | 3813 |
190
+ | --- | --- | --- | --- | ---- |
191
+ | **Accuracy** | | | 0.9673 | 4771 |
192
+ | **Macro Avg** | 0.9283 | 0.8973 | 0.9117 | 4771 |
193
+ | **Weighted Avg** | 0.9664 | 0.9673 | 0.9666 | 4771 |
194
+
195
+ The model achieves near-perfect results on the NER task and excellent results on the intent classification task for this specific dataset.
196
+
197
+ ## Limitations and Bias
198
+
199
+ - The model's performance is highly dependent on the quality and scope of the **HASD dataset**. It may not generalize well to phrasing or appointment types significantly different from what it was trained on.
200
+ - The dataset was primarily generated from templates, which may not capture the full diversity of real human language.
201
+ - The model inherits any biases present in the `distilbert-base-uncased` model and the `clinc/clinc_oos` dataset.