Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from datasets import load_dataset, Dataset
|
3 |
+
from llama_index.core import PromptTemplate
|
4 |
+
from llama_index.core.prompts import ChatMessage
|
5 |
+
from llama_index.llms.openai import OpenAI
|
6 |
+
from pydantic import BaseModel, Field
|
7 |
+
import asyncio
|
8 |
+
import numpy as np
|
9 |
+
import pandas as pd
|
10 |
+
from chromadb import Client
|
11 |
+
from chromadb.utils.embedding_functions import SentenceTransformerEmbeddingFunction
|
12 |
+
import structlog
|
13 |
+
|
14 |
+
logger = structlog.get_logger()
|
15 |
+
|
16 |
+
logger.info('Loading embedding model')
|
17 |
+
embed_fn = SentenceTransformerEmbeddingFunction('BAAI/bge-small-en-v1.5')
|
18 |
+
|
19 |
+
|
20 |
+
def load_train_data_and_vectorstore():
|
21 |
+
logger.info("Loading dataset")
|
22 |
+
ds = load_dataset('SetFit/amazon_reviews_multi_en')
|
23 |
+
train_samples_per_class = 50
|
24 |
+
eval_test_samples_per_class = 10
|
25 |
+
train = Dataset.from_pandas(ds['train'].to_pandas().groupby('label').sample(train_samples_per_class, random_state=1234).reset_index(drop=True))
|
26 |
+
reviews = Client().create_collection(
|
27 |
+
name='reviews',
|
28 |
+
embedding_function=embed_fn,
|
29 |
+
get_or_create=True
|
30 |
+
)
|
31 |
+
logger.info("Adding documents to vector store")
|
32 |
+
reviews.add(documents=train['text'], metadatas=[{'rating': x} for x in train['label']], ids=train['id'])
|
33 |
+
return train, reviews
|
34 |
+
|
35 |
+
train, reviews = load_train_data_and_vectorstore()
|
36 |
+
|
37 |
+
class Rating(BaseModel):
|
38 |
+
rating: int = Field(..., description="Rating of the review", enum=[0, 1, 2, 3, 4])
|
39 |
+
|
40 |
+
llm = OpenAI(model="gpt-4o-mini")
|
41 |
+
structured_llm = llm.as_structured_llm(Rating)
|
42 |
+
|
43 |
+
prompt_tmpl_str = """\
|
44 |
+
The review text is below.
|
45 |
+
---------------------
|
46 |
+
{review}
|
47 |
+
---------------------
|
48 |
+
Given the review text and not prior knowledge, \
|
49 |
+
please attempt to predict the score of the review.
|
50 |
+
|
51 |
+
Query: What is the rating of this review?
|
52 |
+
Answer: \
|
53 |
+
"""
|
54 |
+
|
55 |
+
prompt_tmpl = PromptTemplate(
|
56 |
+
prompt_tmpl_str,
|
57 |
+
)
|
58 |
+
|
59 |
+
async def zero_shot_predict(text):
|
60 |
+
messages = [
|
61 |
+
ChatMessage.from_str(prompt_tmpl.format(review=text))
|
62 |
+
]
|
63 |
+
response = await structured_llm.achat(messages)
|
64 |
+
return response.raw.rating
|
65 |
+
|
66 |
+
few_shot_prompt_tmpl_str = """\
|
67 |
+
The review text is below.
|
68 |
+
---------------------
|
69 |
+
{review}
|
70 |
+
---------------------
|
71 |
+
Given the review text and not prior knowledge, \
|
72 |
+
please attempt to predict the review score of the context. \
|
73 |
+
Here are several examples of reviews and their ratings:
|
74 |
+
|
75 |
+
{random_few_shot_examples}
|
76 |
+
|
77 |
+
Query: What is the rating of this review?
|
78 |
+
Answer: \
|
79 |
+
"""
|
80 |
+
|
81 |
+
few_shot_prompt_tmpl = PromptTemplate(
|
82 |
+
few_shot_prompt_tmpl_str,
|
83 |
+
function_mappings={"random_few_shot_examples": random_few_shot_examples_fn},
|
84 |
+
)
|
85 |
+
|
86 |
+
rng = np.random.Generator(np.random.PCG64(1234))
|
87 |
+
def random_few_shot_examples_fn(**kwargs):
|
88 |
+
if n_samples:=kwargs.get('n_samples'):
|
89 |
+
random_examples = train.shuffle(generator=rng)[:n_samples]
|
90 |
+
else:
|
91 |
+
random_examples = train.shuffle(generator=rng)[:5]
|
92 |
+
|
93 |
+
result_strs = []
|
94 |
+
for text, rating in zip(random_examples['text'], random_examples['label']):
|
95 |
+
result_strs.append(f"Text: {text}\nRating: {rating}")
|
96 |
+
return "\n\n".join(result_strs)
|
97 |
+
|
98 |
+
async def random_few_shot_predict(text, n_examples=5):
|
99 |
+
tasks = []
|
100 |
+
for _ in range(3):
|
101 |
+
messages = [
|
102 |
+
ChatMessage.from_str(few_shot_prompt_tmpl.format(review=text, n_samples=n_examples))
|
103 |
+
]
|
104 |
+
tasks.append(structured_llm.achat(messages, temperature=0.9))
|
105 |
+
results = await asyncio.gather(*tasks)
|
106 |
+
ratings = [r.raw.rating for r in results]
|
107 |
+
# print(ratings)
|
108 |
+
return pd.Series(ratings).mode()[0]
|
109 |
+
|
110 |
+
def dynamic_few_shot_examples_fn(**kwargs):
|
111 |
+
n_examples = kwargs.get('n_examples', 5)
|
112 |
+
retrievals = reviews.query(
|
113 |
+
query_texts=[kwargs['review']],
|
114 |
+
n_results=n_examples
|
115 |
+
)
|
116 |
+
result_strs = []
|
117 |
+
documents = retrievals['documents'][0]
|
118 |
+
metadatas = retrievals['metadatas'][0]
|
119 |
+
for document, metadata in zip(documents, metadatas):
|
120 |
+
result_strs.append(f"Text: {document}\nRating: {metadata.get('rating')}")
|
121 |
+
return "\n\n".join(result_strs)
|
122 |
+
|
123 |
+
dynamic_few_shot_prompt_tmpl_str = """\
|
124 |
+
The review text is below.
|
125 |
+
---------------------
|
126 |
+
{review}
|
127 |
+
---------------------
|
128 |
+
Given the review text and not prior knowledge, \
|
129 |
+
please attempt to predict the review score of the context. \
|
130 |
+
Here are several examples of reviews and their ratings:
|
131 |
+
|
132 |
+
{dynamic_few_shot_examples}
|
133 |
+
|
134 |
+
Query: What is the rating of this review?
|
135 |
+
Answer: \
|
136 |
+
"""
|
137 |
+
|
138 |
+
dynamic_few_shot_prompt_tmpl = PromptTemplate(
|
139 |
+
dynamic_few_shot_prompt_tmpl_str,
|
140 |
+
function_mappings={"dynamic_few_shot_examples": dynamic_few_shot_examples_fn},
|
141 |
+
)
|
142 |
+
|
143 |
+
async def dynamic_few_shot_predict(text, n_examples=5):
|
144 |
+
messages = [
|
145 |
+
ChatMessage.from_str(dynamic_few_shot_prompt_tmpl.format(review=text, n_examples=n_examples))
|
146 |
+
]
|
147 |
+
response = await dynamic_few_shot_structured_llm.achat(messages)
|
148 |
+
return response.raw.rating
|
149 |
+
|
150 |
+
def classify(review, num_examples, api_key):
|
151 |
+
llm = OpenAI(model="gpt-4o-mini", api_key=api_key).as_structured_llm(Rating)
|
152 |
+
zero_shot = asyncio.run(zero_shot_predict(review))
|
153 |
+
random_few_shot = asyncio.run(random_few_shot_predict(review, num_examples))
|
154 |
+
dynamic_few_shot = asyncio.run(dynamic_few_shot_predict(review, num_examples))
|
155 |
+
return zero_shot, random_few_shot, dynamic_few_shot
|
156 |
+
|
157 |
+
with gr.Blocks() as demo:
|
158 |
+
with gr.Row():
|
159 |
+
with gr.Column():
|
160 |
+
api_key = gr.Textbox(label='Openai API Key')
|
161 |
+
n_examples = gr.Slider(minimum=1, maximum=10, value=5, step=1, label='Number of examples to retrieve', interactive=True)
|
162 |
+
review = gr.Textbox(label='Review', interactive=True)
|
163 |
+
submit = gr.Button(value='Submit')
|
164 |
+
with gr.Column():
|
165 |
+
zero_shot_label = gr.Textbox(label='Zero shot', interactive=False)
|
166 |
+
random_few_shot_label = gr.Textbox(label='Random few shot', interactive=False)
|
167 |
+
dynamic_few_shot_label = gr.Textbox(label='Dynamic few shot', interactive=False)
|
168 |
+
|
169 |
+
submit.click(classify, [review, n_examples], [zero_shot_label, random_few_shot_label, dynamic_few_shot_label])
|
170 |
+
|
171 |
+
demo.queue().launch()
|