Update README.md
Browse files
README.md
CHANGED
|
@@ -1,6 +1,7 @@
|
|
| 1 |
---
|
| 2 |
language:
|
| 3 |
- en
|
|
|
|
| 4 |
license: apache-2.0
|
| 5 |
tags:
|
| 6 |
- text-classification
|
|
@@ -23,19 +24,21 @@ base_model:
|
|
| 23 |
- distilbert/distilbert-base-uncased
|
| 24 |
---
|
| 25 |
|
| 26 |
-
# 🎯
|
| 27 |
|
| 28 |
-
|
| 29 |
|
| 30 |
-
|
|
|
|
|
|
|
| 31 |
|
| 32 |
---
|
| 33 |
|
| 34 |
-
## 📊
|
| 35 |
|
| 36 |
-
|
| 37 |
|
| 38 |
-
|
|
| 39 |
| :--- | :--- |
|
| 40 |
| Test Macro F1 | 0.9175 |
|
| 41 |
| Test Accuracy | 0.9175 |
|
|
@@ -43,19 +46,21 @@ Evaluated on the official Banking77 test set (3,080 samples):
|
|
| 43 |
| Test Recall (Macro) | 0.9175 |
|
| 44 |
| Test Loss | 0.3478 |
|
| 45 |
|
| 46 |
-
###
|
| 47 |
|
| 48 |
-
|
|
| 49 |
| :--- | :--- | :--- | :--- | :--- |
|
| 50 |
-
| Baseline (TF-IDF + LogReg + Bayes) |
|
| 51 |
| DistilBERT + LoRA (r=32) | ~2.6% (~1.8M) | 0.9132 | 0.9133 | 0.3189 |
|
| 52 |
| DistilBERT (Full Fine-Tuning) | 100% (66.9M) | 0.9175 | 0.9175 | 0.3478 |
|
| 53 |
|
|
|
|
|
|
|
| 54 |
---
|
| 55 |
|
| 56 |
-
## 🚀
|
| 57 |
|
| 58 |
-
### 1.
|
| 59 |
|
| 60 |
```python
|
| 61 |
from transformers import pipeline
|
|
@@ -65,9 +70,41 @@ classifier = pipeline(
|
|
| 65 |
model="nzgl21/ticket-routing-distilbert"
|
| 66 |
)
|
| 67 |
|
| 68 |
-
#
|
| 69 |
ticket = "I noticed an unknown fee charged on my credit card statement."
|
| 70 |
prediction = classifier(ticket)
|
| 71 |
|
| 72 |
print(prediction)
|
| 73 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
language:
|
| 3 |
- en
|
| 4 |
+
- ru
|
| 5 |
license: apache-2.0
|
| 6 |
tags:
|
| 7 |
- text-classification
|
|
|
|
| 24 |
- distilbert/distilbert-base-uncased
|
| 25 |
---
|
| 26 |
|
| 27 |
+
# 🎯 Маршрутизация обращений и классификация интентов (DistilBERT)
|
| 28 |
|
| 29 |
+
Данная модель представляет собой дообученную версию [distilbert-base-uncased](https://huggingface.co/distilbert-base-uncased) для **автоматической классификации и маршрутизации клиентских обращений** службы поддержки по **77 категориям** банковского датасета [Banking77](https://huggingface.co/datasets/PolyAI/banking77).
|
| 30 |
|
| 31 |
+
Модель разработана в рамках микросервиса для автоматического распределения входящих тикетов по профильным отделам банка.
|
| 32 |
+
|
| 33 |
+
* 🔗 **Репозиторий проекта на GitHub:** [nzgl21/ticket-routing-service](https://github.com/nzgl21/ticket-routing-service)
|
| 34 |
|
| 35 |
---
|
| 36 |
|
| 37 |
+
## 📊 Результаты бенчмарка (Test Set)
|
| 38 |
|
| 39 |
+
Оценка качества проводилась на официальной отложенной тестовой выборке `Banking77` (3 080 примеров):
|
| 40 |
|
| 41 |
+
| Метрика | Значение |
|
| 42 |
| :--- | :--- |
|
| 43 |
| Test Macro F1 | 0.9175 |
|
| 44 |
| Test Accuracy | 0.9175 |
|
|
|
|
| 46 |
| Test Recall (Macro) | 0.9175 |
|
| 47 |
| Test Loss | 0.3478 |
|
| 48 |
|
| 49 |
+
### Сравнение с бейзлайнами и подходами
|
| 50 |
|
| 51 |
+
| Подход | Обучаемые параметры | Test Macro F1 | Test Accuracy | Test Loss |
|
| 52 |
| :--- | :--- | :--- | :--- | :--- |
|
| 53 |
+
| Baseline (TF-IDF + LogReg + Bayes) | — | 0.8901 | — | — |
|
| 54 |
| DistilBERT + LoRA (r=32) | ~2.6% (~1.8M) | 0.9132 | 0.9133 | 0.3189 |
|
| 55 |
| DistilBERT (Full Fine-Tuning) | 100% (66.9M) | 0.9175 | 0.9175 | 0.3478 |
|
| 56 |
|
| 57 |
+
*Полные отчеты по экспериментам и коду обучения доступны в [репозитории проекта](https://github.com/nzgl21/ticket-routing-service).*
|
| 58 |
+
|
| 59 |
---
|
| 60 |
|
| 61 |
+
## 🚀 Как использовать модель
|
| 62 |
|
| 63 |
+
### 1. Быстрый инференс через pipeline
|
| 64 |
|
| 65 |
```python
|
| 66 |
from transformers import pipeline
|
|
|
|
| 70 |
model="nzgl21/ticket-routing-distilbert"
|
| 71 |
)
|
| 72 |
|
| 73 |
+
# Пример клиентского тикета
|
| 74 |
ticket = "I noticed an unknown fee charged on my credit card statement."
|
| 75 |
prediction = classifier(ticket)
|
| 76 |
|
| 77 |
print(prediction)
|
| 78 |
+
# Вывод: [{'label': 'card_payment_fee_charged', 'score': 0.964}]
|
| 79 |
+
```
|
| 80 |
+
|
| 81 |
+
### 2. Загрузка через PyTorch / AutoModel
|
| 82 |
+
|
| 83 |
+
```python
|
| 84 |
+
import torch
|
| 85 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
| 86 |
+
|
| 87 |
+
model_name = "nzgl21/ticket-routing-distilbert"
|
| 88 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 89 |
+
model = AutoModelForSequenceClassification.from_pretrained(model_name)
|
| 90 |
+
|
| 91 |
+
inputs = tokenizer("How can I change my PIN?", return_tensors="pt")
|
| 92 |
+
with torch.no_grad():
|
| 93 |
+
logits = model(**inputs).logits
|
| 94 |
+
|
| 95 |
+
predicted_class_id = logits.argmax().item()
|
| 96 |
+
print("Предсказанный интент:", model.config.id2label[predicted_class_id])
|
| 97 |
+
```
|
| 98 |
+
|
| 99 |
+
---
|
| 100 |
+
|
| 101 |
+
## ⚙️ Гиперпараметры обучения
|
| 102 |
+
|
| 103 |
+
Модель обучалась с помощью Hugging Face `Trainer` со следующей конфигурацией:
|
| 104 |
+
|
| 105 |
+
* **Базовая модель:** `distilbert-base-uncased`
|
| 106 |
+
* **Количество эпох:** 4
|
| 107 |
+
* **Learning Rate:** `3e-5` (оптимизатор AdamW)
|
| 108 |
+
* **Weight Decay:** 0.01
|
| 109 |
+
* **Размер батча:** 32 (Train)
|
| 110 |
+
* **Максимальная длина последовательности:** 64 токена
|