Spaces:
Sleeping
Sleeping
Commit
·
9fc1cf3
1
Parent(s):
386b3f5
add init app without context
Browse files- app.py +68 -0
- requirements.txt +4 -0
app.py
ADDED
@@ -0,0 +1,68 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import BlenderbotTokenizer, BlenderbotForConditionalGeneration, Trainer, TrainingArguments
|
2 |
+
from datasets import load_dataset
|
3 |
+
import torch
|
4 |
+
import gradio as gr
|
5 |
+
|
6 |
+
# ✅ Charger le modèle et le tokenizer
|
7 |
+
model_name = "facebook/blenderbot-400M-distill"
|
8 |
+
tokenizer = BlenderbotTokenizer.from_pretrained(model_name)
|
9 |
+
model = BlenderbotForConditionalGeneration.from_pretrained(model_name)
|
10 |
+
|
11 |
+
# ✅ Charger le dataset "fka/awesome-chatgpt-prompts"
|
12 |
+
dataset = load_dataset("fka/awesome-chatgpt-prompts")
|
13 |
+
|
14 |
+
# ✅ Préparer les données en adaptant les colonnes disponibles
|
15 |
+
def preprocess_function(examples):
|
16 |
+
inputs = examples["act"]
|
17 |
+
targets = examples["prompt"]
|
18 |
+
|
19 |
+
# Tokenisation avec padding et tronquage
|
20 |
+
model_inputs = tokenizer(inputs, max_length=128, truncation=True, padding="max_length", return_tensors="pt")
|
21 |
+
labels = tokenizer(targets, max_length=128, truncation=True, padding="max_length", return_tensors="pt")
|
22 |
+
|
23 |
+
model_inputs["labels"] = labels["input_ids"]
|
24 |
+
return model_inputs
|
25 |
+
|
26 |
+
# Appliquer le prétraitement en retirant les colonnes inutiles
|
27 |
+
tokenized_dataset = dataset.map(preprocess_function, batched=True, remove_columns=["act", "prompt"])
|
28 |
+
|
29 |
+
# ✅ Vérifier si un fine-tuning est nécessaire
|
30 |
+
do_training = False # Change à True pour entraîner le modèle
|
31 |
+
|
32 |
+
if do_training:
|
33 |
+
training_args = TrainingArguments(
|
34 |
+
output_dir="./results",
|
35 |
+
evaluation_strategy="epoch",
|
36 |
+
learning_rate=5e-5,
|
37 |
+
per_device_train_batch_size=8,
|
38 |
+
per_device_eval_batch_size=8,
|
39 |
+
num_train_epochs=3,
|
40 |
+
weight_decay=0.01,
|
41 |
+
save_total_limit=2,
|
42 |
+
push_to_hub=False,
|
43 |
+
)
|
44 |
+
|
45 |
+
trainer = Trainer(
|
46 |
+
model=model,
|
47 |
+
args=training_args,
|
48 |
+
train_dataset=tokenized_dataset["train"],
|
49 |
+
eval_dataset=tokenized_dataset["test"],
|
50 |
+
)
|
51 |
+
|
52 |
+
trainer.train()
|
53 |
+
|
54 |
+
# ✅ Pipeline d'inférence pour le chatbot
|
55 |
+
def chatbot_response(user_input, history=None): # Accepte l'historique pour éviter l'erreur
|
56 |
+
inputs = tokenizer(user_input, return_tensors="pt", max_length=128, truncation=True)
|
57 |
+
with torch.no_grad():
|
58 |
+
output = model.generate(**inputs, max_length=128, num_beams=5, early_stopping=True)
|
59 |
+
return tokenizer.decode(output[0], skip_special_tokens=True)
|
60 |
+
|
61 |
+
# ✅ Interface Gradio pour dialoguer avec le chatbot
|
62 |
+
chatbot = gr.ChatInterface(
|
63 |
+
fn=chatbot_response,
|
64 |
+
title="Chatbot Service Après-Vente",
|
65 |
+
description="Posez vos questions sur le service après-vente et obtenez des réponses instantanées.",
|
66 |
+
)
|
67 |
+
|
68 |
+
chatbot.launch()
|
requirements.txt
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
transformers
|
2 |
+
gradio
|
3 |
+
datasets
|
4 |
+
torch
|