Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -1,50 +1,20 @@
|
|
1 |
import os
|
2 |
import torch
|
3 |
import spaces
|
4 |
-
from collections.abc import Iterator
|
5 |
-
from threading import Thread
|
6 |
import gradio as gr
|
|
|
|
|
7 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
8 |
|
9 |
MAX_MAX_NEW_TOKENS = 4096
|
10 |
-
DEFAULT_MAX_NEW_TOKENS = 2048
|
11 |
MAX_INPUT_TOKEN_LENGTH = 4096
|
12 |
-
|
13 |
HF_TOKEN = os.environ['HF_TOKEN']
|
14 |
|
15 |
-
DESCRIPTION = """\
|
16 |
-
## π IndicTrans3-beta π: Multilingual Translation for 22 Indic Languages
|
17 |
-
|
18 |
-
IndicTrans3 is the latest state-of-the-art (SOTA) translation model from AI4Bharat, designed to handle translations across **22 Indic languages** with high accuracy. It supports **document-level machine translation (MT)** and is built to match the performance of other leading SOTA models.
|
19 |
-
|
20 |
-
π’ **Training data will be released soon!**
|
21 |
-
|
22 |
-
### πΉ Features
|
23 |
-
β
Supports **22 Indic languages**
|
24 |
-
β
Enables **document-level translation**
|
25 |
-
β
Achieves **SOTA performance** in Indic MT
|
26 |
-
β
Optimized for **real-world applications**
|
27 |
-
|
28 |
-
### π Try It Out!
|
29 |
-
1οΈβ£ Enter text in any supported language
|
30 |
-
2οΈβ£ Select the target language
|
31 |
-
3οΈβ£ Click **Translate** and get high-quality results!
|
32 |
-
|
33 |
-
Built for **linguistic diversity and accessibility**, IndicTrans3 is a major step forward in **Indic language AI**.
|
34 |
-
|
35 |
-
π‘ **Source:** AI4Bharat | Powered by Hugging Face
|
36 |
-
"""
|
37 |
-
|
38 |
-
# if not torch.cuda.is_available():
|
39 |
-
# DESCRIPTION += "\n<p>Running on CPU π₯Ά This demo does not work on CPU.</p>"
|
40 |
-
|
41 |
-
|
42 |
-
# if torch.cuda.is_available():
|
43 |
model_id = "ai4bharat/IndicTrans3-beta"
|
44 |
-
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto",
|
45 |
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-3B-Instruct")
|
46 |
|
47 |
-
|
48 |
LANGUAGES = {
|
49 |
"Hindi": "hin_Deva",
|
50 |
"Bengali": "ben_Beng",
|
@@ -69,59 +39,33 @@ LANGUAGES = {
|
|
69 |
"Bodo": "brx_Deva"
|
70 |
}
|
71 |
|
72 |
-
|
73 |
-
|
74 |
-
tgt_lang: str,
|
75 |
-
message: str,
|
76 |
-
max_new_tokens: int = 1024,
|
77 |
-
temperature: float = 0.6,
|
78 |
-
top_p: float = 0.9,
|
79 |
-
top_k: int = 50,
|
80 |
-
repetition_penalty: float = 1.2,
|
81 |
-
) -> str:
|
82 |
-
conversation = []
|
83 |
-
conversation.append({"role": "user", "content": f"Translate the following text to {tgt_lang}: {message}"})
|
84 |
-
|
85 |
-
input_ids = tokenizer.apply_chat_template(conversation, return_tensors="pt", add_generation_prompt=True)
|
86 |
-
if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
|
87 |
-
input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
|
88 |
-
input_ids = input_ids.to(model.device)
|
89 |
-
|
90 |
-
outputs = model.generate(
|
91 |
-
input_ids=input_ids,
|
92 |
-
max_new_tokens=max_new_tokens,
|
93 |
-
do_sample=True,
|
94 |
-
top_p=top_p,
|
95 |
-
top_k=top_k,
|
96 |
-
temperature=temperature,
|
97 |
-
num_beams=1,
|
98 |
-
repetition_penalty=repetition_penalty,
|
99 |
-
)
|
100 |
-
|
101 |
-
return tokenizer.decode(outputs[0][input_ids.shape[1]:], skip_special_tokens=True)
|
102 |
-
|
103 |
|
104 |
@spaces.GPU
|
105 |
-
def
|
106 |
-
tgt_lang: str,
|
107 |
message: str,
|
|
|
|
|
108 |
max_new_tokens: int = 1024,
|
109 |
temperature: float = 0.6,
|
110 |
top_p: float = 0.9,
|
111 |
top_k: int = 50,
|
112 |
repetition_penalty: float = 1.2,
|
113 |
) -> Iterator[str]:
|
114 |
-
|
115 |
conversation = []
|
116 |
-
|
|
|
|
|
|
|
117 |
|
118 |
input_ids = tokenizer.apply_chat_template(conversation, return_tensors="pt", add_generation_prompt=True)
|
119 |
if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
|
120 |
input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
|
121 |
-
gr.Warning(f"Trimmed input
|
122 |
input_ids = input_ids.to(model.device)
|
123 |
|
124 |
-
streamer = TextIteratorStreamer(tokenizer, timeout=
|
125 |
generate_kwargs = dict(
|
126 |
{"input_ids": input_ids},
|
127 |
streamer=streamer,
|
@@ -154,84 +98,84 @@ def store_feedback(rating, feedback_text):
|
|
154 |
return "Thank you for your feedback!"
|
155 |
|
156 |
css = """
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
162 |
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
label="Translate To",
|
184 |
-
elem_id="translate-to"
|
185 |
-
)
|
186 |
-
|
187 |
-
text_output = gr.Textbox(
|
188 |
-
label="",
|
189 |
-
lines=10,
|
190 |
-
max_lines=100,
|
191 |
-
elem_id="output-text"
|
192 |
-
)
|
193 |
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
text_input,
|
200 |
-
gr.Number(value=4096, visible=False),
|
201 |
-
gr.Number(value=0.1, visible=False),
|
202 |
-
gr.Number(value=0.9, visible=False),
|
203 |
-
gr.Number(value=50, visible=False),
|
204 |
-
gr.Number(value=1.0, visible=False)
|
205 |
-
],
|
206 |
-
outputs=text_output
|
207 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
208 |
|
209 |
gr.Examples(
|
210 |
examples=[
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
|
216 |
],
|
217 |
-
inputs=
|
218 |
-
tgt_lang,
|
219 |
-
text_input,
|
220 |
-
gr.Number(value=4096, visible=False),
|
221 |
-
gr.Number(value=0.1, visible=False),
|
222 |
-
gr.Number(value=0.9, visible=False),
|
223 |
-
gr.Number(value=50, visible=False),
|
224 |
-
gr.Number(value=1.0, visible=False)
|
225 |
-
],
|
226 |
-
outputs=text_output,
|
227 |
-
fn=generate_for_examples,
|
228 |
-
cache_examples=True,
|
229 |
-
examples_per_page=5
|
230 |
)
|
231 |
|
232 |
-
|
|
|
233 |
gr.Markdown("## Rate Translation & Provide Feedback π")
|
234 |
-
gr.Markdown("Help us improve the translation quality by providing your feedback
|
235 |
with gr.Row():
|
236 |
rating = gr.Radio(
|
237 |
["1", "2", "3", "4", "5"],
|
@@ -246,11 +190,93 @@ with gr.Blocks(theme=gr.themes.Default(), css=css) as demo:
|
|
246 |
|
247 |
feedback_submit = gr.Button("Submit Feedback")
|
248 |
feedback_result = gr.Textbox(label="", visible=False)
|
249 |
-
|
250 |
-
|
251 |
-
|
252 |
-
|
253 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
254 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
255 |
|
256 |
-
|
|
|
|
1 |
import os
|
2 |
import torch
|
3 |
import spaces
|
|
|
|
|
4 |
import gradio as gr
|
5 |
+
from threading import Thread
|
6 |
+
from collections.abc import Iterator
|
7 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
8 |
|
9 |
MAX_MAX_NEW_TOKENS = 4096
|
|
|
10 |
MAX_INPUT_TOKEN_LENGTH = 4096
|
11 |
+
DEFAULT_MAX_NEW_TOKENS = 2048
|
12 |
HF_TOKEN = os.environ['HF_TOKEN']
|
13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
model_id = "ai4bharat/IndicTrans3-beta"
|
15 |
+
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto", token=HF_TOKEN)
|
16 |
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-3B-Instruct")
|
17 |
|
|
|
18 |
LANGUAGES = {
|
19 |
"Hindi": "hin_Deva",
|
20 |
"Bengali": "ben_Beng",
|
|
|
39 |
"Bodo": "brx_Deva"
|
40 |
}
|
41 |
|
42 |
+
def format_message_for_translation(message, target_lang):
|
43 |
+
return f"Translate the following text to {target_lang}: {message}"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
44 |
|
45 |
@spaces.GPU
|
46 |
+
def translate_message(
|
|
|
47 |
message: str,
|
48 |
+
chat_history: list[dict],
|
49 |
+
target_language: str = "Hindi",
|
50 |
max_new_tokens: int = 1024,
|
51 |
temperature: float = 0.6,
|
52 |
top_p: float = 0.9,
|
53 |
top_k: int = 50,
|
54 |
repetition_penalty: float = 1.2,
|
55 |
) -> Iterator[str]:
|
|
|
56 |
conversation = []
|
57 |
+
|
58 |
+
translation_request = format_message_for_translation(message, target_language)
|
59 |
+
print(f"Translation request: {translation_request}")
|
60 |
+
conversation.append({"role": "user", "content": translation_request})
|
61 |
|
62 |
input_ids = tokenizer.apply_chat_template(conversation, return_tensors="pt", add_generation_prompt=True)
|
63 |
if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
|
64 |
input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
|
65 |
+
gr.Warning(f"Trimmed input as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")
|
66 |
input_ids = input_ids.to(model.device)
|
67 |
|
68 |
+
streamer = TextIteratorStreamer(tokenizer, timeout=240.0, skip_prompt=True, skip_special_tokens=True)
|
69 |
generate_kwargs = dict(
|
70 |
{"input_ids": input_ids},
|
71 |
streamer=streamer,
|
|
|
98 |
return "Thank you for your feedback!"
|
99 |
|
100 |
css = """
|
101 |
+
body {
|
102 |
+
background-color: #f7f7f7;
|
103 |
+
}
|
104 |
+
.feedback-section {
|
105 |
+
margin-top: 30px;
|
106 |
+
border-top: 1px solid #ddd;
|
107 |
+
padding-top: 20px;
|
108 |
+
}
|
109 |
+
.container {
|
110 |
+
max-width: 90%;
|
111 |
+
margin: 0 auto;
|
112 |
+
}
|
113 |
+
.language-selector {
|
114 |
+
margin-bottom: 20px;
|
115 |
+
padding: 10px;
|
116 |
+
background-color: #ffffff;
|
117 |
+
border-radius: 8px;
|
118 |
+
box-shadow: 0 2px 5px rgba(0,0,0,0.1);
|
119 |
+
}
|
120 |
+
.advanced-options {
|
121 |
+
margin-top: 20px;
|
122 |
+
}
|
123 |
+
"""
|
124 |
|
125 |
+
DESCRIPTION = """\
|
126 |
+
IndicTrans3 is the latest state-of-the-art (SOTA) translation model from AI4Bharat, designed to handle translations across <b>22 Indic languages</b> with high accuracy. It supports <b>document-level machine translation (MT)</b> and is built to match the performance of other leading SOTA models. <br>
|
127 |
+
π’ <b>Training data will be released soon!</b>
|
128 |
+
<h3>πΉ Features</h3>
|
129 |
+
β
Supports <b>22 Indic languages</b>
|
130 |
+
β
Enables <b>document-level translation</b>
|
131 |
+
β
Achieves <b>SOTA performance</b> in Indic MT
|
132 |
+
β
Optimized for <b>real-world applications</b>
|
133 |
+
<h3>π Try It Out!</h3>
|
134 |
+
1οΈβ£ Enter text in any supported language
|
135 |
+
2οΈβ£ Select the target language
|
136 |
+
3οΈβ£ Click <b>Translate</b> and get high-quality results!
|
137 |
+
Built for <b>linguistic diversity and accessibility</b>, IndicTrans3 is a major step forward in <b>Indic language AI</b>.
|
138 |
+
π‘ <b>Source:</b> AI4Bharat | Powered by Hugging Face
|
139 |
+
"""
|
140 |
+
|
141 |
+
with gr.Blocks(css=css) as demo:
|
142 |
+
with gr.Column(elem_classes="container"):
|
143 |
+
gr.Markdown("# π IndicTrans3-beta π: Multilingual Translation for 22 Indic Languages </center>")
|
144 |
+
gr.Markdown(DESCRIPTION)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
145 |
|
146 |
+
target_language = gr.Dropdown(
|
147 |
+
list(LANGUAGES.keys()),
|
148 |
+
value="Hindi",
|
149 |
+
label="Which language would you like to translate to?",
|
150 |
+
elem_id="language-dropdown"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
151 |
)
|
152 |
+
|
153 |
+
chatbot = gr.Chatbot(height=400, elem_id="chatbot")
|
154 |
+
|
155 |
+
with gr.Row():
|
156 |
+
msg = gr.Textbox(
|
157 |
+
placeholder="Enter text to translate...",
|
158 |
+
show_label=False,
|
159 |
+
container=False,
|
160 |
+
scale=9
|
161 |
+
)
|
162 |
+
submit_btn = gr.Button("Translate", scale=1)
|
163 |
|
164 |
gr.Examples(
|
165 |
examples=[
|
166 |
+
"The Taj Mahal stands majestically along the banks of river Yamuna, a timeless symbol of eternal love.",
|
167 |
+
"Kumbh Mela is the world's largest gathering of people, where millions of pilgrims bathe in sacred rivers for spiritual purification.",
|
168 |
+
"India's classical dance forms like Bharatanatyam, Kathak, and Odissi beautifully blend rhythm, expression, and storytelling.",
|
169 |
+
"Ayurveda, the ancient Indian medical system, focuses on holistic wellness through natural herbs and balanced living.",
|
170 |
+
"During Diwali, homes across India are decorated with oil lamps, colorful rangoli patterns, and twinkling lights to celebrate the victory of light over darkness."
|
171 |
],
|
172 |
+
inputs=msg
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
173 |
)
|
174 |
|
175 |
+
|
176 |
+
with gr.Accordion("Provide Feedback", open=True):
|
177 |
gr.Markdown("## Rate Translation & Provide Feedback π")
|
178 |
+
gr.Markdown("Help us improve the translation quality by providing your feedback.")
|
179 |
with gr.Row():
|
180 |
rating = gr.Radio(
|
181 |
["1", "2", "3", "4", "5"],
|
|
|
190 |
|
191 |
feedback_submit = gr.Button("Submit Feedback")
|
192 |
feedback_result = gr.Textbox(label="", visible=False)
|
193 |
+
|
194 |
+
with gr.Accordion("Advanced Options", open=False, elem_classes="advanced-options"):
|
195 |
+
max_new_tokens = gr.Slider(
|
196 |
+
label="Max new tokens",
|
197 |
+
minimum=1,
|
198 |
+
maximum=MAX_MAX_NEW_TOKENS,
|
199 |
+
step=1,
|
200 |
+
value=DEFAULT_MAX_NEW_TOKENS,
|
201 |
+
)
|
202 |
+
temperature = gr.Slider(
|
203 |
+
label="Temperature",
|
204 |
+
minimum=0.1,
|
205 |
+
maximum=1.0,
|
206 |
+
step=0.1,
|
207 |
+
value=0.1,
|
208 |
+
)
|
209 |
+
top_p = gr.Slider(
|
210 |
+
label="Top-p (nucleus sampling)",
|
211 |
+
minimum=0.05,
|
212 |
+
maximum=1.0,
|
213 |
+
step=0.05,
|
214 |
+
value=0.9,
|
215 |
+
)
|
216 |
+
top_k = gr.Slider(
|
217 |
+
label="Top-k",
|
218 |
+
minimum=1,
|
219 |
+
maximum=100,
|
220 |
+
step=1,
|
221 |
+
value=50,
|
222 |
+
)
|
223 |
+
repetition_penalty = gr.Slider(
|
224 |
+
label="Repetition penalty",
|
225 |
+
minimum=1.0,
|
226 |
+
maximum=2.0,
|
227 |
+
step=0.05,
|
228 |
+
value=1.0,
|
229 |
)
|
230 |
+
|
231 |
+
chat_state = gr.State([])
|
232 |
+
|
233 |
+
def user(user_message, history, target_lang):
|
234 |
+
return "", history + [[user_message, None]]
|
235 |
+
|
236 |
+
def bot(history, target_lang, max_tokens, temp, top_p_val, top_k_val, rep_penalty):
|
237 |
+
user_message = history[-1][0]
|
238 |
+
history[-1][1] = ""
|
239 |
+
|
240 |
+
for chunk in translate_message(
|
241 |
+
user_message,
|
242 |
+
history[:-1],
|
243 |
+
target_lang,
|
244 |
+
max_tokens,
|
245 |
+
temp,
|
246 |
+
top_p_val,
|
247 |
+
top_k_val,
|
248 |
+
rep_penalty
|
249 |
+
):
|
250 |
+
history[-1][1] = chunk
|
251 |
+
yield history
|
252 |
+
|
253 |
+
msg.submit(
|
254 |
+
user,
|
255 |
+
[msg, chatbot, target_language],
|
256 |
+
[msg, chatbot],
|
257 |
+
queue=False
|
258 |
+
).then(
|
259 |
+
bot,
|
260 |
+
[chatbot, target_language, max_new_tokens, temperature, top_p, top_k, repetition_penalty],
|
261 |
+
chatbot
|
262 |
+
)
|
263 |
+
|
264 |
+
submit_btn.click(
|
265 |
+
user,
|
266 |
+
[msg, chatbot, target_language],
|
267 |
+
[msg, chatbot],
|
268 |
+
queue=False
|
269 |
+
).then(
|
270 |
+
bot,
|
271 |
+
[chatbot, target_language, max_new_tokens, temperature, top_p, top_k, repetition_penalty],
|
272 |
+
chatbot
|
273 |
+
)
|
274 |
+
|
275 |
+
feedback_submit.click(
|
276 |
+
fn=store_feedback,
|
277 |
+
inputs=[rating, feedback_text],
|
278 |
+
outputs=feedback_result
|
279 |
+
)
|
280 |
|
281 |
+
if __name__ == "__main__":
|
282 |
+
demo.launch()
|