Spaces:
Sleeping
Sleeping
Delete app.py
Browse files
app.py
DELETED
@@ -1,191 +0,0 @@
|
|
1 |
-
import gradio as gr
|
2 |
-
import spaces
|
3 |
-
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
|
4 |
-
import torch
|
5 |
-
from huggingface_hub import InferenceClient
|
6 |
-
import os
|
7 |
-
|
8 |
-
# Initialize Cerebras client for Llama 4
|
9 |
-
cerebras_client = InferenceClient(
|
10 |
-
"meta-llama/Llama-4-Scout-17B-16E-Instruct",
|
11 |
-
provider="cerebras",
|
12 |
-
token=os.getenv("HF_TOKEN"),
|
13 |
-
)
|
14 |
-
|
15 |
-
# Global variables for models and tokenizers
|
16 |
-
en_es_tokenizer = None
|
17 |
-
en_es_model = None
|
18 |
-
es_en_tokenizer = None
|
19 |
-
es_en_model = None
|
20 |
-
|
21 |
-
@spaces.GPU(duration=60)
|
22 |
-
def translate_en_to_es(text):
|
23 |
-
global en_es_tokenizer, en_es_model
|
24 |
-
|
25 |
-
# Initialize EN->ES model if needed
|
26 |
-
if en_es_tokenizer is None or en_es_model is None:
|
27 |
-
en_es_tokenizer = AutoTokenizer.from_pretrained("facebook/nllb-200-distilled-600M", src_lang="eng_Latn", tgt_lang="spa_Latn")
|
28 |
-
en_es_model = AutoModelForSeq2SeqLM.from_pretrained(
|
29 |
-
"facebook/nllb-200-distilled-600M",
|
30 |
-
torch_dtype=torch.float16
|
31 |
-
).cuda()
|
32 |
-
|
33 |
-
# Translate
|
34 |
-
inputs = en_es_tokenizer(text, return_tensors="pt", max_length=512, truncation=True).to("cuda")
|
35 |
-
with torch.no_grad():
|
36 |
-
outputs = en_es_model.generate(
|
37 |
-
**inputs,
|
38 |
-
forced_bos_token_id=en_es_tokenizer.convert_tokens_to_ids("spa_Latn"),
|
39 |
-
max_length=512,
|
40 |
-
num_beams=5,
|
41 |
-
early_stopping=True
|
42 |
-
)
|
43 |
-
|
44 |
-
translation = en_es_tokenizer.decode(outputs[0], skip_special_tokens=True)
|
45 |
-
return translation
|
46 |
-
|
47 |
-
@spaces.GPU(duration=60)
|
48 |
-
def translate_es_to_en(text):
|
49 |
-
global es_en_tokenizer, es_en_model
|
50 |
-
|
51 |
-
# Initialize ES->EN model if needed
|
52 |
-
if es_en_tokenizer is None or es_en_model is None:
|
53 |
-
es_en_tokenizer = AutoTokenizer.from_pretrained("facebook/nllb-200-distilled-600M", src_lang="spa_Latn", tgt_lang="eng_Latn")
|
54 |
-
es_en_model = AutoModelForSeq2SeqLM.from_pretrained(
|
55 |
-
"facebook/nllb-200-distilled-600M",
|
56 |
-
torch_dtype=torch.float16
|
57 |
-
).cuda()
|
58 |
-
|
59 |
-
# Translate
|
60 |
-
inputs = es_en_tokenizer(text, return_tensors="pt", max_length=512, truncation=True).to("cuda")
|
61 |
-
with torch.no_grad():
|
62 |
-
outputs = es_en_model.generate(
|
63 |
-
**inputs,
|
64 |
-
forced_bos_token_id=es_en_tokenizer.convert_tokens_to_ids("eng_Latn"),
|
65 |
-
max_length=512,
|
66 |
-
num_beams=5,
|
67 |
-
early_stopping=True
|
68 |
-
)
|
69 |
-
|
70 |
-
translation = es_en_tokenizer.decode(outputs[0], skip_special_tokens=True)
|
71 |
-
return translation
|
72 |
-
|
73 |
-
def refine_with_llama(original_text, translation, direction, region="general", formality="neutral"):
|
74 |
-
if direction == "en_to_es":
|
75 |
-
refine_prompt = f"""You are an expert Spanish translator. Refine the following translation to address these common issues:
|
76 |
-
|
77 |
-
1. Context and ambiguity resolution
|
78 |
-
2. Cultural nuances and regional variations for {region}
|
79 |
-
3. Tone and formality ({formality})
|
80 |
-
4. Grammatical correctness
|
81 |
-
5. Idiomatic expressions
|
82 |
-
|
83 |
-
Original English: {original_text}
|
84 |
-
Initial Spanish translation: {translation}
|
85 |
-
Region preference: {region}
|
86 |
-
|
87 |
-
Provide only the refined Spanish translation, nothing else."""
|
88 |
-
else:
|
89 |
-
refine_prompt = f"""You are an expert English translator. Refine the following translation to address these common issues:
|
90 |
-
|
91 |
-
1. Context and ambiguity resolution
|
92 |
-
2. Cultural nuances and natural English expressions
|
93 |
-
3. Tone and formality ({formality})
|
94 |
-
4. Grammatical correctness
|
95 |
-
5. Idiomatic expressions
|
96 |
-
|
97 |
-
Original Spanish: {original_text}
|
98 |
-
Initial English translation: {translation}
|
99 |
-
Formality: {formality}
|
100 |
-
|
101 |
-
Provide only the refined English translation, nothing else."""
|
102 |
-
|
103 |
-
try:
|
104 |
-
response = cerebras_client.chat_completion(
|
105 |
-
messages=[{"role": "user", "content": refine_prompt}],
|
106 |
-
max_tokens=512,
|
107 |
-
temperature=0.3
|
108 |
-
)
|
109 |
-
return response.choices[0].message.content.strip()
|
110 |
-
except Exception as e:
|
111 |
-
return f"Refinement error: {str(e)}"
|
112 |
-
|
113 |
-
def complete_translation(text, direction, region, formality):
|
114 |
-
if not text.strip():
|
115 |
-
return "", ""
|
116 |
-
|
117 |
-
try:
|
118 |
-
# Step 1: Initial translation
|
119 |
-
if direction == "English to Spanish":
|
120 |
-
initial_translation = translate_en_to_es(text)
|
121 |
-
refined_translation = refine_with_llama(text, initial_translation, "en_to_es", region, formality)
|
122 |
-
else: # Spanish to English
|
123 |
-
initial_translation = translate_es_to_en(text)
|
124 |
-
refined_translation = refine_with_llama(text, initial_translation, "es_to_en", region, formality)
|
125 |
-
|
126 |
-
return initial_translation, refined_translation
|
127 |
-
except Exception as e:
|
128 |
-
return f"Error: {str(e)}", ""
|
129 |
-
|
130 |
-
# Create Gradio interface
|
131 |
-
with gr.Blocks(title="Bidirectional English-Spanish Translator") as demo:
|
132 |
-
gr.Markdown("# Bidirectional English-Spanish Translator")
|
133 |
-
gr.Markdown("Powered by NLLB-200 + Llama 4 via Cerebras for context-aware, culturally nuanced translations")
|
134 |
-
|
135 |
-
with gr.Row():
|
136 |
-
with gr.Column(scale=2):
|
137 |
-
input_text = gr.Textbox(
|
138 |
-
label="Text to Translate",
|
139 |
-
placeholder="Enter text in English or Spanish...",
|
140 |
-
lines=6
|
141 |
-
)
|
142 |
-
|
143 |
-
with gr.Row():
|
144 |
-
direction = gr.Dropdown(
|
145 |
-
choices=["English to Spanish", "Spanish to English"],
|
146 |
-
value="English to Spanish",
|
147 |
-
label="Translation Direction"
|
148 |
-
)
|
149 |
-
|
150 |
-
with gr.Row():
|
151 |
-
region = gr.Dropdown(
|
152 |
-
choices=["general", "Mexico", "Spain", "Argentina", "Colombia", "Peru"],
|
153 |
-
value="general",
|
154 |
-
label="Spanish Variant (for ES translations)"
|
155 |
-
)
|
156 |
-
formality = gr.Dropdown(
|
157 |
-
choices=["neutral", "formal", "informal"],
|
158 |
-
value="neutral",
|
159 |
-
label="Formality Level"
|
160 |
-
)
|
161 |
-
|
162 |
-
translate_btn = gr.Button("Translate", variant="primary", size="lg")
|
163 |
-
|
164 |
-
with gr.Column(scale=2):
|
165 |
-
with gr.Row():
|
166 |
-
initial_output = gr.Textbox(
|
167 |
-
label="Initial Translation (NLLB-200)",
|
168 |
-
lines=3,
|
169 |
-
interactive=False
|
170 |
-
)
|
171 |
-
refined_output = gr.Textbox(
|
172 |
-
label="Refined Translation (Llama 4)",
|
173 |
-
lines=3,
|
174 |
-
interactive=False
|
175 |
-
)
|
176 |
-
|
177 |
-
# Connect function
|
178 |
-
translate_btn.click(
|
179 |
-
fn=complete_translation,
|
180 |
-
inputs=[input_text, direction, region, formality],
|
181 |
-
outputs=[initial_output, refined_output]
|
182 |
-
)
|
183 |
-
|
184 |
-
input_text.submit(
|
185 |
-
fn=complete_translation,
|
186 |
-
inputs=[input_text, direction, region, formality],
|
187 |
-
outputs=[initial_output, refined_output]
|
188 |
-
)
|
189 |
-
|
190 |
-
if __name__ == "__main__":
|
191 |
-
demo.launch()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|