Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -3,22 +3,30 @@ import gradio as gr
|
|
3 |
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
|
4 |
import torch
|
5 |
import os
|
|
|
6 |
|
7 |
-
# Load the model
|
8 |
model_name = "google/flan-t5-base"
|
9 |
-
hf_token = os.environ.get("HF_TOKEN") #
|
10 |
|
11 |
tokenizer = AutoTokenizer.from_pretrained(model_name, use_auth_token=hf_token)
|
12 |
model = AutoModelForSeq2SeqLM.from_pretrained(model_name, use_auth_token=hf_token)
|
13 |
|
14 |
# Move the model to CPU (or GPU if available)
|
15 |
-
|
|
|
16 |
|
17 |
-
# Function to generate
|
18 |
def generate_prompt(original, translation):
|
19 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
|
21 |
-
#
|
22 |
def predict_scores(file):
|
23 |
df = pd.read_csv(file.name, sep="\t")
|
24 |
scores = []
|
@@ -26,32 +34,35 @@ def predict_scores(file):
|
|
26 |
for _, row in df.iterrows():
|
27 |
prompt = generate_prompt(row["original"], row["translation"])
|
28 |
|
29 |
-
# Tokenize and
|
30 |
-
inputs = tokenizer(prompt, return_tensors="pt").to(
|
31 |
outputs = model.generate(**inputs, max_new_tokens=10)
|
32 |
-
|
33 |
-
# Decode and extract the score from the response
|
34 |
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
35 |
|
36 |
-
#
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
|
|
|
|
|
|
|
|
42 |
|
43 |
scores.append(score_val)
|
44 |
|
45 |
df["predicted_score"] = scores
|
46 |
return df
|
47 |
|
48 |
-
#
|
49 |
iface = gr.Interface(
|
50 |
fn=predict_scores,
|
51 |
inputs=gr.File(label="Upload dev.tsv"),
|
52 |
outputs=gr.Dataframe(label="QE Output with Predicted Score"),
|
53 |
-
title="MT QE with
|
|
|
54 |
)
|
55 |
|
56 |
-
# Launch
|
57 |
iface.launch()
|
|
|
3 |
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
|
4 |
import torch
|
5 |
import os
|
6 |
+
import re
|
7 |
|
8 |
+
# Load the model and tokenizer
|
9 |
model_name = "google/flan-t5-base"
|
10 |
+
hf_token = os.environ.get("HF_TOKEN") # Set as a secret in Hugging Face Space settings
|
11 |
|
12 |
tokenizer = AutoTokenizer.from_pretrained(model_name, use_auth_token=hf_token)
|
13 |
model = AutoModelForSeq2SeqLM.from_pretrained(model_name, use_auth_token=hf_token)
|
14 |
|
15 |
# Move the model to CPU (or GPU if available)
|
16 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
17 |
+
model.to(device)
|
18 |
|
19 |
+
# Function to generate a clean prompt
|
20 |
def generate_prompt(original, translation):
|
21 |
+
return (
|
22 |
+
f"Rate the quality of this translation from 0 (poor) to 1 (excellent). "
|
23 |
+
f"Only respond with a number.\n\n"
|
24 |
+
f"Source: {original}\n"
|
25 |
+
f"Translation: {translation}\n"
|
26 |
+
f"Score:"
|
27 |
+
)
|
28 |
|
29 |
+
# Main prediction function
|
30 |
def predict_scores(file):
|
31 |
df = pd.read_csv(file.name, sep="\t")
|
32 |
scores = []
|
|
|
34 |
for _, row in df.iterrows():
|
35 |
prompt = generate_prompt(row["original"], row["translation"])
|
36 |
|
37 |
+
# Tokenize and send to model
|
38 |
+
inputs = tokenizer(prompt, return_tensors="pt").to(device)
|
39 |
outputs = model.generate(**inputs, max_new_tokens=10)
|
|
|
|
|
40 |
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
41 |
|
42 |
+
# Debug print (optional)
|
43 |
+
print("Response:", response)
|
44 |
+
|
45 |
+
# Extract numeric score using regex
|
46 |
+
match = re.search(r"\b([01](?:\.\d+)?)\b", response)
|
47 |
+
if match:
|
48 |
+
score_val = float(match.group(1))
|
49 |
+
score_val = max(0, min(score_val, 1)) # Clamp between 0 and 1
|
50 |
+
else:
|
51 |
+
score_val = -1 # fallback if model output is invalid
|
52 |
|
53 |
scores.append(score_val)
|
54 |
|
55 |
df["predicted_score"] = scores
|
56 |
return df
|
57 |
|
58 |
+
# Gradio UI
|
59 |
iface = gr.Interface(
|
60 |
fn=predict_scores,
|
61 |
inputs=gr.File(label="Upload dev.tsv"),
|
62 |
outputs=gr.Dataframe(label="QE Output with Predicted Score"),
|
63 |
+
title="MT QE with FLAN-T5-Base",
|
64 |
+
description="Upload a dev.tsv file with columns: 'original' and 'translation'."
|
65 |
)
|
66 |
|
67 |
+
# Launch app
|
68 |
iface.launch()
|