Spaces:
Running
on
Zero
Running
on
Zero
import gradio as gr | |
import spaces | |
from transformers import pipeline | |
import torch | |
DESCRIPTION=""" | |
### a Turkish encoder-decoder language model | |
Welcome to our Huggingface space, where you can explore the capabilities of TURNA. | |
**Key Features of TURNA:** | |
- **Powerful Architecture:** TURNA contains 1.1B parameters, and was pre-trained with an encoder-decoder architecture following the UL2 framework on 43B tokens from various domains. | |
- **Diverse Training Data:** Our model is trained on a varied dataset of 43 billion tokens, covering a wide array of domains. | |
- **Broad Applications:** TURNA is fine-tuned for a variety of generation and understanding tasks, including: | |
- Summarization | |
- Paraphrasing | |
- News title generation | |
- Sentiment classification | |
- Text categorization | |
- Named entity recognition | |
- Part-of-speech tagging | |
- Semantic textual similarity | |
- Natural language inference | |
**Note:** First inference might take time as the models are downloaded on-the-go. | |
*TURNA can generate toxic content or provide erroneous information. Double-check before usage.* | |
""" | |
CITATION = """ | |
Refer to our [paper](https://arxiv.org/abs/2401.14373) for more details. | |
### Citation | |
```bibtex | |
@misc{uludogan2024turna, | |
title={TURNA: A Turkish Encoder-Decoder Language Model for Enhanced Understanding and Generation}, | |
author={Gökçe Uludoğan and Zeynep Yirmibeşoğlu Balal and Furkan Akkurt and Melikşah Türker and Onur Güngör and Susan Üsküdarlı}, | |
year={2024}, | |
eprint={2401.14373}, | |
archivePrefix={arXiv}, | |
primaryClass={cs.CL} | |
} | |
``` | |
""" | |
sentiment_example = [["Bu üründen çok memnun kaldım."]] | |
long_text = [["Eyfel Kulesi (Fransızca: La tour Eiffel [la tuʀ ɛˈfɛl]), Paris'teki demir kule. Kule, aynı zamanda tüm dünyada Fransa'nın sembolü halini almıştır. İsmini, inşa ettiren Fransız inşaat mühendisi Gustave Eiffel'den alır.[1] En büyük turizm cazibelerinden biri olan Eyfel Kulesi, yılda 6 milyon turist çeker. 2002 yılında toplam ziyaretçi sayısı 200 milyona ulaşmıştır."], ["Kalp krizi geçirenlerin yaklaşık üçte birinin kısa bir süre önce grip atlattığı düşünülüyor. Peki grip virüsü ne yapıyor da kalp krizine yol açıyor? Karpuz şöyle açıkladı: Grip virüsü kanın yapışkanlığını veya pıhtılaşmasını artırıyor."]] | |
ner_example = [["Benim adım Turna."]] | |
t2t_example = [["Paraphrase: Bu üründen çok memnun kaldım."]] | |
nli_example = [["Bunu çok beğendim.", "Bunu çok sevdim."]] | |
text_category_example = [[" anadolu_efes e 18 lik star ! beko_basketbol_ligi nde iddialı bir kadroyla sezona giren anadolu_efes transfer harekatına devam ediyor"]] | |
def nli(first_input, second_input, model_choice="turna_nli_nli_tr"): | |
if model_choice=="turna_nli_nli_tr": | |
input = f"hipotez: {first_input} önerme: {second_input}" | |
nli_model = pipeline(model="boun-tabi-LMG/turna_nli_nli_tr", device=0) | |
return nli_model(input)[0]["generated_text"] | |
else: | |
input = f"ilk cümle: {first_input} ikinci cümle: {second_input}" | |
stsb_model = pipeline(model="boun-tabi-LMG/turna_semantic_similarity_stsb_tr", device=0) | |
return stsb_model(input)[0]["generated_text"] | |
def sentiment_analysis(input, model_choice="turna_classification_17bintweet_sentiment"): | |
sentiment_model = pipeline(model=f"boun-tabi-LMG/{model_choice}", device=0) | |
return sentiment_model(input, max_new_tokens = 4)[0]["generated_text"] | |
def pos(input, model_choice, max_new_tokens, length_penalty, no_repeat_ngram_size): | |
if model_choice=="turna_pos_imst": | |
pos_imst = pipeline(model="boun-tabi-LMG/turna_pos_imst", device=0) | |
return pos_imst(input, max_new_tokens = max_new_tokens, length_penalty=length_penalty, no_repeat_ngram_size=no_repeat_ngram_size)[0]["generated_text"] | |
else: | |
pos_boun = pipeline(model="boun-tabi-LMG/turna_pos_boun", device=0) | |
return pos_boun(input, max_new_tokens = max_new_tokens, length_penalty=length_penalty, no_repeat_ngram_size=no_repeat_ngram_size)[0]["generated_text"] | |
def ner(input, model_choice, max_new_tokens, length_penalty, no_repeat_ngram_size): | |
if model_choice=="turna_ner_wikiann": | |
ner_wikiann = pipeline(model="boun-tabi-LMG/turna_ner_wikiann", device=0) | |
return ner_wikiann(input, max_new_tokens = max_new_tokens, length_penalty=length_penalty, no_repeat_ngram_size=no_repeat_ngram_size)[0]["generated_text"] | |
else: | |
ner_model = pipeline(model="boun-tabi-LMG/turna_ner_milliyet", device=0) | |
return ner_model(input, max_new_tokens = max_new_tokens, length_penalty=length_penalty, no_repeat_ngram_size=no_repeat_ngram_size)[0]["generated_text"] | |
def paraphrase(input, model_choice, max_new_tokens): | |
if model_choice=="turna_paraphrasing_tatoeba": | |
paraphrasing = pipeline(model="boun-tabi-LMG/turna_paraphrasing_tatoeba", device=0) | |
return paraphrasing(input, max_new_tokens = max_new_tokens)[0]["generated_text"] | |
else: | |
paraphrasing_sub = pipeline(model="boun-tabi-LMG/turna_paraphrasing_opensubtitles", device=0) | |
return paraphrasing_sub(input, max_new_tokens = max_new_tokens)[0]["generated_text"] | |
def summarize(input, model_choice, max_new_tokens, length_penalty, no_repeat_ngram_size): | |
model_mapping = {"turna_summarization_tr_news": "boun-tabi-LMG/turna_summarization_tr_news", | |
"turna_summarization_mlsum": "boun-tabi-LMG/turna_summarization_mlsum"} | |
summarization_model = pipeline(model=model_mapping[model_choice], device=0) | |
return summarization_model(input, max_new_tokens = max_new_tokens, length_penalty=length_penalty, no_repeat_ngram_size=no_repeat_ngram_size)[0]["generated_text"] | |
def generate_title(input, model_choice, max_new_tokens, length_penalty, no_repeat_ngram_size): | |
model_mapping = {"turna_title_generation_tr_news": "boun-tabi-LMG/turna_title_generation_tr_news", | |
"turna_title_generation_mlsum": "boun-tabi-LMG/turna_title_generation_mlsum"} | |
summarization_model = pipeline(model=model_mapping[model_choice], device=0) | |
return summarization_model(input, max_new_tokens = max_new_tokens, length_penalty=length_penalty, no_repeat_ngram_size=no_repeat_ngram_size)[0]["generated_text"] | |
def categorize(input): | |
ttc = pipeline(model="boun-tabi-LMG/turna_classification_ttc4900", device=0) | |
return ttc(input, max_new_tokens = 8)[0]["generated_text"] | |
def turna(input, max_new_tokens, length_penalty, | |
top_k, top_p, temp, num_beams, | |
do_sample, no_repeat_ngram_size, repetition_penalty): | |
turna = pipeline(model="boun-tabi-LMG/TURNA", device=0) | |
input = f"[S2S] {input}<EOS>" | |
return turna(input, max_new_tokens = max_new_tokens, length_penalty=length_penalty, | |
top_k=top_k, top_p=top_p, temperature=temp, num_beams=num_beams, | |
do_sample = do_sample, no_repeat_ngram_size=no_repeat_ngram_size, repetition_penalty=repetition_penalty)[0]["generated_text"] | |
with gr.Blocks(theme="abidlabs/Lime") as demo: | |
gr.Markdown("# TURNA") | |
gr.Image("images/turna-logo.png", width=100, show_label=False, show_download_button=False, show_share_button=False) | |
with gr.Tab("TURNA"): | |
gr.Markdown(DESCRIPTION) | |
with gr.Tab("Sentiment Analysis"): | |
gr.Markdown("TURNA fine-tuned on sentiment analysis. Enter text to analyse sentiment and pick the model (tweets or product reviews).") | |
with gr.Column(): | |
with gr.Row(): | |
with gr.Column(): | |
sentiment_choice = gr.Radio(choices = ["turna_classification_17bintweet_sentiment", "turna_classification_tr_product_reviews"], label ="Model", value="turna_classification_17bintweet_sentiment") | |
sentiment_input = gr.Textbox(label="Sentiment Analysis Input") | |
sentiment_submit = gr.Button() | |
sentiment_output = gr.Textbox(label="Sentiment Analysis Output") | |
sentiment_submit.click(sentiment_analysis, inputs=[sentiment_input, sentiment_choice], outputs=sentiment_output) | |
sentiment_examples = gr.Examples(examples = sentiment_example, inputs = [sentiment_input, sentiment_choice], outputs=sentiment_output, fn=sentiment_analysis) | |
with gr.Tab("Text Categorization"): | |
gr.Markdown("TURNA fine-tuned on text categorization. Enter text to categorize text or try the example.") | |
with gr.Column(): | |
with gr.Row(): | |
with gr.Column(): | |
text_input = gr.Textbox(label="Text Categorization Input") | |
text_submit = gr.Button() | |
text_output = gr.Textbox(label="Text Categorization Output") | |
text_submit.click(categorize, inputs=[text_input], outputs=text_output) | |
text_examples = gr.Examples(examples = text_category_example,inputs=[text_input], outputs=text_output, fn=categorize) | |
with gr.Tab("NLI & STS"): | |
gr.Markdown("TURNA fine-tuned on natural language inference or semantic textual similarity. Enter text to infer entailment or measure semantic similarity. ") | |
with gr.Column(): | |
with gr.Row(): | |
with gr.Column(): | |
nli_choice = gr.Radio(choices = ["turna_nli_nli_tr", "turna_semantic_similarity_stsb_tr"], label ="Model", value="turna_nli_nli_tr") | |
nli_first_input = gr.Textbox(label="First Sentence") | |
nli_second_input = gr.Textbox(label="Second Sentence") | |
nli_submit = gr.Button() | |
nli_output = gr.Textbox(label="NLI Output") | |
nli_submit.click(nli, inputs=[nli_first_input, nli_second_input, nli_choice], outputs=nli_output) | |
nli_examples = gr.Examples(examples = nli_example, inputs = [nli_first_input, nli_second_input, nli_choice], outputs=nli_output, fn=nli) | |
with gr.Tab("POS"): | |
gr.Markdown("TURNA fine-tuned on part-of-speech-tagging. Enter text to parse parts of speech and pick the model.") | |
with gr.Column(): | |
with gr.Row(): | |
with gr.Column(): | |
pos_choice = gr.Radio(choices = ["turna_pos_imst", "turna_pos_boun"], label ="Model", value="turna_pos_imst") | |
with gr.Accordion("Advanced Generation Parameters"): | |
max_new_tokens = gr.Slider(label = "Maximum length", | |
minimum = 0, | |
maximum = 64, | |
value = 64) | |
length_penalty = gr.Slider(label = "Length penalty", | |
minimum = -10, | |
maximum = 10, | |
value=2.0) | |
no_repeat_ngram_size =gr.Slider(label="No Repeat N-Gram Size", minimum=0,value=3,) | |
with gr.Column(): | |
pos_input = gr.Textbox(label="POS Input") | |
pos_submit = gr.Button() | |
pos_output = gr.Textbox(label="POS Output") | |
pos_submit.click(pos, inputs=[pos_input, pos_choice, max_new_tokens, length_penalty, no_repeat_ngram_size], outputs=pos_output) | |
pos_examples = gr.Examples(examples = ner_example, inputs = [pos_input, pos_choice, max_new_tokens, length_penalty, no_repeat_ngram_size], outputs=pos_output, fn=pos) | |
with gr.Tab("NER"): | |
gr.Markdown("TURNA fine-tuned on named entity recognition. Enter text to parse named entities and pick the model.") | |
with gr.Column(): | |
with gr.Row(): | |
with gr.Column(): | |
ner_choice = gr.Radio(choices = ["turna_ner_wikiann", "turna_ner_milliyet"], label ="Model", value="turna_ner_wikiann") | |
with gr.Accordion("Advanced Generation Parameters"): | |
max_new_tokens = gr.Slider(label = "Maximum length", | |
minimum = 0, | |
maximum = 64, | |
value = 64) | |
length_penalty = gr.Slider(label = "Length penalty", | |
minimum = -10, | |
maximum = 10, | |
value=2.0) | |
no_repeat_ngram_size =gr.Slider(label="No Repeat N-Gram Size", minimum=0,value=3,) | |
with gr.Column(): | |
ner_input = gr.Textbox(label="NER Input") | |
ner_submit = gr.Button() | |
ner_output = gr.Textbox(label="NER Output") | |
ner_submit.click(ner, inputs=[ner_input, ner_choice, max_new_tokens, length_penalty, no_repeat_ngram_size], outputs=ner_output) | |
ner_examples = gr.Examples(examples = ner_example, inputs = [ner_input, ner_choice, max_new_tokens, length_penalty, no_repeat_ngram_size], outputs=ner_output, fn=ner) | |
with gr.Tab("Paraphrase"): | |
gr.Markdown("TURNA fine-tuned on paraphrasing. Enter text to paraphrase and pick the model.") | |
with gr.Column(): | |
with gr.Row(): | |
with gr.Column(): | |
paraphrasing_choice = gr.Radio(choices = ["turna_paraphrasing_tatoeba", "turna_paraphrasing_opensubtitles"], label ="Model", value="turna_paraphrasing_tatoeba") | |
with gr.Accordion("Advanced Generation Parameters"): | |
max_new_tokens = gr.Slider(label = "Maximum length", | |
minimum = 0, | |
maximum = 20, | |
value = 20) | |
with gr.Column(): | |
paraphrasing_input = gr.Textbox(label = "Paraphrasing Input") | |
paraphrasing_submit = gr.Button() | |
paraphrasing_output = gr.Text(label="Paraphrasing Output") | |
paraphrasing_submit.click(paraphrase, inputs=[paraphrasing_input, paraphrasing_choice, max_new_tokens], outputs=paraphrasing_output) | |
paraphrase_examples = gr.Examples(examples = long_text, inputs = [paraphrasing_input, paraphrasing_choice, max_new_tokens], outputs=paraphrasing_output, fn=paraphrase) | |
with gr.Tab("Summarization"): | |
gr.Markdown("TURNA fine-tuned on summarization. Enter text to summarize and pick the model.") | |
with gr.Column(): | |
with gr.Row(): | |
with gr.Column(): | |
sum_choice = gr.Radio(choices = ["turna_summarization_mlsum", "turna_summarization_tr_news"], label ="Model", value="turna_summarization_mlsum") | |
with gr.Accordion("Advanced Generation Parameters"): | |
max_new_tokens = gr.Slider(label = "Maximum length", | |
minimum = 0, | |
maximum = 512, | |
value = 128) | |
length_penalty = gr.Slider(label = "Length penalty", | |
minimum = -10, | |
maximum = 10, | |
value=2.0) | |
no_repeat_ngram_size =gr.Slider(label="No Repeat N-Gram Size", minimum=0,value=3,) | |
with gr.Column(): | |
sum_input = gr.Textbox(label = "Summarization Input") | |
sum_submit = gr.Button() | |
sum_output = gr.Textbox(label = "Summarization Output") | |
sum_submit.click(summarize, inputs=[sum_input, sum_choice, max_new_tokens, length_penalty, no_repeat_ngram_size], outputs=sum_output) | |
sum_examples = gr.Examples(examples = long_text, inputs = [sum_input, sum_choice, max_new_tokens, length_penalty, no_repeat_ngram_size], outputs=sum_output, fn=summarize) | |
with gr.Tab("Title Generation"): | |
gr.Markdown("TURNA fine-tuned on news title generation. Enter news text to generate a title.") | |
with gr.Column(): | |
with gr.Row(): | |
with gr.Column(): | |
title_choice = gr.Radio(choices = ["turna_title_generation_tr_news", "turna_title_generation_mlsum"], label ="Model", value="turna_title_generation_tr_news") | |
with gr.Accordion("Advanced Generation Parameters"): | |
max_new_tokens = gr.Slider(label = "Maximum length", | |
minimum = 0, | |
maximum = 64, | |
value = 64) | |
length_penalty = gr.Slider(label = "Length penalty", | |
minimum = -10, | |
maximum = 10, | |
value=2.0) | |
no_repeat_ngram_size =gr.Slider(label="No Repeat N-Gram Size", minimum=0,value=3,) | |
with gr.Column(): | |
title_input = gr.Textbox(label = "News Title Generation Input") | |
title_submit = gr.Button() | |
title_output = gr.Textbox(label = "News Title Generation Output") | |
title_submit.click(generate_title, inputs=[title_input, title_choice, max_new_tokens, length_penalty, no_repeat_ngram_size], outputs=title_output) | |
title_examples = gr.Examples(examples = long_text, inputs = [title_input, title_choice, max_new_tokens, length_penalty, no_repeat_ngram_size], outputs=title_output, fn=generate_title) | |
with gr.Tab("Text Generation"): | |
gr.Markdown("Pre-trained TURNA. Enter text to start generating.") | |
with gr.Column(): | |
with gr.Row(): | |
with gr.Column(): | |
with gr.Accordion("Advanced Generation Parameters"): | |
max_new_tokens = gr.Slider(label = "Maximum length", | |
minimum = 0, | |
maximum = 512, | |
value = 128) | |
length_penalty = gr.Slider(label = "Length penalty", | |
value=1.0) | |
top_k = gr.Slider(label = "Top-k", value=10) | |
top_p = gr.Slider(label = "Top-p", value=0.95) | |
temp = gr.Slider(label = "Temperature", value=1.0, minimum=0.1, maximum=100.0) | |
no_repeat_ngram_size =gr.Slider(label="No Repeat N-Gram Size", minimum=0,value=3,) | |
repetition_penalty = gr.Slider(label = "Repetition Penalty", minimum=0.0, value=3.1, step=0.1) | |
num_beams = gr.Slider(label = "Number of beams", minimum=1, | |
maximum=10, value=3) | |
do_sample = gr.Radio(choices = [True, False], value = True, label = "Sampling") | |
with gr.Column(): | |
text_gen_input = gr.Textbox(label="Text Generation Input") | |
text_gen_submit = gr.Button() | |
text_gen_output = gr.Textbox(label="Text Generation Output") | |
text_gen_submit.click(turna, inputs=[text_gen_input, max_new_tokens, length_penalty, | |
top_k, top_p, temp, num_beams, | |
do_sample, no_repeat_ngram_size, repetition_penalty], outputs=text_gen_output) | |
text_gen_example = [["Bir varmış, bir yokmuş, evvel zaman içinde, kalbur saman içinde, uzak diyarların birinde bir turna"]] | |
text_gen_examples = gr.Examples(examples = text_gen_example, inputs = [text_gen_input, max_new_tokens, length_penalty, | |
top_k, top_p, temp, num_beams, do_sample, no_repeat_ngram_size, repetition_penalty], outputs=text_gen_output, fn=turna) | |
gr.Markdown(CITATION) | |
demo.launch() |