|
|
import os |
|
|
os.system("pip install -q flash_attn==2.7.4.post1 transformers==4.49.0 accelerate>=0.26.0") |
|
|
import gradio as gr |
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline |
|
|
import re |
|
|
import os |
|
|
import torch |
|
|
|
|
|
hf_token = os.environ.get('hf_token') |
|
|
|
|
|
model_path='microsoft/Phi-4-mini-instruct' |
|
|
|
|
|
model = AutoModelForCausalLM.from_pretrained( |
|
|
model_path, |
|
|
|
|
|
|
|
|
trust_remote_code=True, |
|
|
) |
|
|
tokenizer = AutoTokenizer.from_pretrained(model_path) |
|
|
|
|
|
Examples_to_teach_model=""" |
|
|
Text: I hate apples |
|
|
Sentiment analysis: |
|
|
Sentiments: Negative |
|
|
PPrint Key words: hate, aples |
|
|
Text: I enjoy watching it |
|
|
Sentiment analysis: |
|
|
Sentiments: Positive |
|
|
PPrint Key words: enjoy |
|
|
Text: I'm tired of this long process |
|
|
Sentiment analysis: |
|
|
Sentiments: Negative |
|
|
PPrint Key words: tired, long process |
|
|
""" |
|
|
|
|
|
def make_prompt(sentence): |
|
|
prompt = (""" |
|
|
Given the below sentence(s) can you extract the sentiment and keywords for each sentence: |
|
|
|
|
|
""" + sentence |
|
|
) |
|
|
|
|
|
return prompt |
|
|
|
|
|
def split_conj(text): |
|
|
return re.sub('(but|yet|although|however|nevertheless|on the other hand|still|though)', "|", text).split('|') |
|
|
|
|
|
def get_sentiment_from_llm(review_text): |
|
|
|
|
|
""" |
|
|
segments=[] |
|
|
for sen in sentences: |
|
|
segments=segments+split_conj(sen) |
|
|
ls_outputs=[] |
|
|
|
|
|
segments= [x for x in segments if len(x)>=5] |
|
|
print(segments) |
|
|
""" |
|
|
|
|
|
pipe = pipeline( |
|
|
"text-generation", |
|
|
model=model, |
|
|
tokenizer=tokenizer, |
|
|
) |
|
|
|
|
|
|
|
|
|
|
|
generation_args = { |
|
|
"max_new_tokens": 500, |
|
|
"return_full_text": False, |
|
|
"temperature": 0.0, |
|
|
"do_sample": False, |
|
|
} |
|
|
|
|
|
question_and_background = make_prompt(review_text) |
|
|
messages = [ |
|
|
{"role": "system", "content": "You are a helpful AI assistant who helps to extract sentiments and keywords from given sentences."}, |
|
|
{"role": "user", "content": question_and_background} |
|
|
] |
|
|
output = pipe(messages, **generation_args) |
|
|
print(output) |
|
|
return output |
|
|
|
|
|
demo = gr.Blocks() |
|
|
sentiment_extr = gr.Interface( |
|
|
fn=get_sentiment_from_llm, |
|
|
inputs=gr.Textbox(label="Text input", type="text"), |
|
|
outputs=gr.Textbox(label="Sentiments", type="text"), |
|
|
title="Sentiment analysis and keywords extraction", |
|
|
description=""" |
|
|
Enter one or two sentences in the Text Input and click "Submit" to see the sentiments extracted. <br> |
|
|
For longer input, please allow 2-3 minutes as the model is running on small CPU. <br> |
|
|
Base model: Flan-t5 from Google. <br> |
|
|
Prompt tuned by Thuyen Truong for sentiment extraction. |
|
|
""" |
|
|
) |
|
|
with demo: |
|
|
gr.TabbedInterface([sentiment_extr], ["Sentiment text analysis"]) |
|
|
demo.launch() |
|
|
|
