Spaces:
Running
on
Zero
Running
on
Zero
da03
commited on
Commit
·
d87049c
1
Parent(s):
5c05a33
app.py
CHANGED
@@ -3,156 +3,42 @@ import torch
|
|
3 |
import gradio as gr
|
4 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
5 |
|
6 |
-
# Load
|
7 |
-
implicit_cot_model_name = 'yuntian-deng/
|
8 |
-
implicit_cot_model = AutoModelForCausalLM.from_pretrained(implicit_cot_model_name)
|
9 |
tokenizer = AutoTokenizer.from_pretrained(implicit_cot_model_name)
|
10 |
|
11 |
-
no_cot_model_name = 'yuntian-deng/gpt2-no-cot-multiplication'
|
12 |
-
no_cot_model = AutoModelForCausalLM.from_pretrained(no_cot_model_name)
|
13 |
-
|
14 |
-
explicit_cot_model_name = 'yuntian-deng/gpt2-explicit-cot-multiplication'
|
15 |
-
explicit_cot_model = AutoModelForCausalLM.from_pretrained(explicit_cot_model_name)
|
16 |
-
|
17 |
-
models = {'implicit': implicit_cot_model, 'no': no_cot_model, 'explicit': explicit_cot_model}
|
18 |
-
|
19 |
# Constants
|
20 |
-
|
21 |
-
|
22 |
-
def preprocess(num):
|
23 |
-
num = str(num).strip().replace(' ', '')
|
24 |
-
reversed_num = ' '.join(num[::-1])
|
25 |
-
return reversed_num
|
26 |
-
|
27 |
-
def postprocess(raw_output):
|
28 |
-
prediction = raw_output.replace(' ', '')[::-1]
|
29 |
-
return prediction
|
30 |
|
31 |
@spaces.GPU
|
32 |
-
def
|
33 |
-
input_text =
|
34 |
inputs = tokenizer(input_text, return_tensors='pt').to('cuda' if torch.cuda.is_available() else 'cpu')
|
35 |
-
|
36 |
|
37 |
input_ids = inputs['input_ids']
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
try:
|
44 |
-
num1_int = int(num1)
|
45 |
-
num2_int = int(num2)
|
46 |
-
ground_truth_product = str(num1_int * num2_int)
|
47 |
-
ground_truth_digits_reversed = list(ground_truth_product)[::-1]
|
48 |
-
except ValueError:
|
49 |
-
valid_input = False
|
50 |
-
|
51 |
-
generated_ids_per_model = {model_name: inputs['input_ids'].data.clone() for model_name in models}
|
52 |
-
finished_per_model = {model_name: False for model_name in models}
|
53 |
-
past_key_values_per_model = {model_name: None for model_name in models}
|
54 |
-
predicted_annotations_per_model = {}
|
55 |
-
for step in range(max(MAX_PRODUCT_DIGITS_PER_MODEL.values())): # Set a maximum limit to prevent infinite loops
|
56 |
-
# Ground Truth
|
57 |
-
if not valid_input:
|
58 |
-
ground_truth_annotations = [('Invalid Input!', None)]
|
59 |
-
else:
|
60 |
-
ground_truth_annotations = [(ground_truth_digit, None) for ground_truth_digit in ground_truth_digits_reversed[:step+1]]
|
61 |
-
ground_truth_annotations = ground_truth_annotations[::-1]
|
62 |
-
# Predicted
|
63 |
-
for model_name in models:
|
64 |
-
model = models[model_name]
|
65 |
-
if finished_per_model[model_name]:
|
66 |
-
continue
|
67 |
-
if step >= MAX_PRODUCT_DIGITS_PER_MODEL[model_name]:
|
68 |
-
continue
|
69 |
-
generation_kwargs = {
|
70 |
-
'input_ids': generated_ids_per_model[model_name],
|
71 |
-
'max_new_tokens': 1,
|
72 |
-
'do_sample': False,
|
73 |
-
'past_key_values': past_key_values_per_model[model_name],
|
74 |
-
'return_dict_in_generate': True,
|
75 |
-
'use_cache': True
|
76 |
-
}
|
77 |
-
if step == 0:
|
78 |
-
del generation_kwargs['past_key_values']
|
79 |
-
outputs = model.generate(**generation_kwargs)
|
80 |
-
generated_ids = outputs.sequences
|
81 |
-
next_token_id = generated_ids[0, -1]
|
82 |
-
#print (next_token_id)
|
83 |
-
|
84 |
-
if next_token_id.item() == tokenizer.eos_token_id:
|
85 |
-
finished_per_model[model_name] = True
|
86 |
-
if valid_input:
|
87 |
-
if len([item for item in predicted_annotations_per_model[model_name] if item[1] is not None]) < len(ground_truth_digits_reversed):
|
88 |
-
predicted_annotations_per_model[model_name].insert(0, ('⠀', 'wrong'))
|
89 |
-
continue
|
90 |
-
|
91 |
-
generated_ids_per_model[model_name] = generated_ids
|
92 |
-
past_key_values_per_model[model_name] = outputs.past_key_values
|
93 |
-
|
94 |
-
output_text = tokenizer.decode(generated_ids[0, input_len:], skip_special_tokens=True)
|
95 |
-
predicted_digits_reversed = output_text.strip().split(' ')
|
96 |
-
|
97 |
-
predicted_annotations = []
|
98 |
-
is_correct_sofar = True
|
99 |
-
if model_name == 'explicit':
|
100 |
-
if '=' not in predicted_digits_reversed:
|
101 |
-
predicted_annotations = [(predicted_digit, None) for predicted_digit in predicted_digits_reversed]
|
102 |
-
predicted_digits_reversed = []
|
103 |
-
else:
|
104 |
-
equal_sign_position = predicted_digits_reversed.index('=')
|
105 |
-
predicted_annotations = [(predicted_digit, None) for predicted_digit in predicted_digits_reversed[:equal_sign_position+1]]
|
106 |
-
predicted_digits_reversed = predicted_digits_reversed[equal_sign_position+1:]
|
107 |
|
108 |
-
|
109 |
-
predicted_digit = predicted_digits_reversed[i]
|
110 |
-
if not valid_input:
|
111 |
-
is_correct_digit = None
|
112 |
-
elif i >= len(ground_truth_digits_reversed):
|
113 |
-
if predicted_digit == '0' and is_correct_sofar:
|
114 |
-
is_correct_digit = True
|
115 |
-
else:
|
116 |
-
is_correct_digit = False
|
117 |
-
else:
|
118 |
-
ground_truth_digit = ground_truth_digits_reversed[i]
|
119 |
-
if predicted_digit == ground_truth_digit:
|
120 |
-
is_correct_digit = True
|
121 |
-
else:
|
122 |
-
is_correct_digit = False
|
123 |
-
if not is_correct_digit:
|
124 |
-
is_correct_sofar = False
|
125 |
-
if is_correct_digit is None:
|
126 |
-
predicted_annotations.append((predicted_digit, None))
|
127 |
-
elif is_correct_digit:
|
128 |
-
predicted_annotations.append((predicted_digit, "correct"))
|
129 |
-
else:
|
130 |
-
predicted_annotations.append((predicted_digit, "wrong"))
|
131 |
-
predicted_annotations = predicted_annotations[::-1]
|
132 |
-
predicted_annotations_per_model[model_name] = predicted_annotations
|
133 |
|
134 |
-
|
135 |
-
predicted_annotations_nocot = predicted_annotations_per_model['no']
|
136 |
-
predicted_annotations_explicit_cot = predicted_annotations_per_model['explicit']
|
137 |
-
|
138 |
-
yield ground_truth_annotations, predicted_annotations_implicit_cot, predicted_annotations_nocot, predicted_annotations_explicit_cot
|
139 |
|
140 |
color_map = {"correct": "green", "wrong": "red"}
|
141 |
|
142 |
demo = gr.Interface(
|
143 |
-
fn=
|
144 |
inputs=[
|
145 |
-
gr.Textbox(label='
|
146 |
-
gr.Textbox(label='Second Number (up to 15 digits)', value='987654321'),
|
147 |
],
|
148 |
outputs=[
|
149 |
-
gr.HighlightedText(label='
|
150 |
-
gr.HighlightedText(label='Implicit CoT Prediction (Ours)', combine_adjacent=False, show_legend=False, color_map=color_map, show_inline_category=False),
|
151 |
-
gr.HighlightedText(label='No CoT Prediction', combine_adjacent=False, show_legend=False, color_map=color_map, show_inline_category=False),
|
152 |
-
gr.HighlightedText(label='Explicit CoT Steps & Prediction', combine_adjacent=False, show_legend=False, color_map=color_map, show_inline_category=False),
|
153 |
],
|
154 |
-
title='
|
155 |
-
description='This demo showcases
|
156 |
article="""
|
157 |
- [Paper 1: Implicit Chain of Thought Reasoning via Knowledge Distillation](https://arxiv.org/pdf/2311.01460)
|
158 |
- [Paper 2: From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step](https://arxiv.org/pdf/2405.14838)
|
@@ -160,8 +46,8 @@ demo = gr.Interface(
|
|
160 |
- [Tweet Announcement](https://twitter.com/yuntiandeng/status/1795854740879774036)
|
161 |
""",
|
162 |
clear_btn=None,
|
163 |
-
submit_btn="
|
164 |
live=False,
|
165 |
concurrency_limit=1
|
166 |
)
|
167 |
-
demo.queue(max_size=
|
|
|
3 |
import gradio as gr
|
4 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
5 |
|
6 |
+
# Load the implicit CoT model
|
7 |
+
implicit_cot_model_name = 'yuntian-deng/implicit-cot-math-mistral7b'
|
8 |
+
implicit_cot_model = AutoModelForCausalLM.from_pretrained(implicit_cot_model_name, torch_dtype=torch.bfloat16)
|
9 |
tokenizer = AutoTokenizer.from_pretrained(implicit_cot_model_name)
|
10 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
# Constants
|
12 |
+
MAX_RESULT_TOKENS = 10
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
|
14 |
@spaces.GPU
|
15 |
+
def predict_answer(question):
|
16 |
+
input_text = ' '.join(question.split()).strip() + ' ' + tokenizer.eos_token
|
17 |
inputs = tokenizer(input_text, return_tensors='pt').to('cuda' if torch.cuda.is_available() else 'cpu')
|
18 |
+
implicit_cot_model.to('cuda' if torch.cuda.is_available() else 'cpu')
|
19 |
|
20 |
input_ids = inputs['input_ids']
|
21 |
+
outputs = implicit_cot_model.generate(input_ids=input_ids,
|
22 |
+
max_new_tokens=MAX_RESULT_TOKENS,
|
23 |
+
do_sample=False)
|
24 |
+
generated_ids = outputs.sequences[0]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
|
26 |
+
prediction = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
|
28 |
+
return prediction
|
|
|
|
|
|
|
|
|
29 |
|
30 |
color_map = {"correct": "green", "wrong": "red"}
|
31 |
|
32 |
demo = gr.Interface(
|
33 |
+
fn=predict_answer,
|
34 |
inputs=[
|
35 |
+
gr.Textbox(label='Question', value='A set of 7 spoons costs $21. If each spoon would be sold separately, how much would 5 spoons cost?'),
|
|
|
36 |
],
|
37 |
outputs=[
|
38 |
+
gr.HighlightedText(label='Implicit CoT Prediction', combine_adjacent=False, show_legend=False, color_map=color_map, show_inline_category=False),
|
|
|
|
|
|
|
39 |
],
|
40 |
+
title='Solving Grade School Math Problems with Implicit CoT',
|
41 |
+
description='This demo showcases Mistral-7B\'s ability to solve grade school math problems without producing intermediate steps, using our stepwise internalization method.',
|
42 |
article="""
|
43 |
- [Paper 1: Implicit Chain of Thought Reasoning via Knowledge Distillation](https://arxiv.org/pdf/2311.01460)
|
44 |
- [Paper 2: From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step](https://arxiv.org/pdf/2405.14838)
|
|
|
46 |
- [Tweet Announcement](https://twitter.com/yuntiandeng/status/1795854740879774036)
|
47 |
""",
|
48 |
clear_btn=None,
|
49 |
+
submit_btn="Get Answer!",
|
50 |
live=False,
|
51 |
concurrency_limit=1
|
52 |
)
|
53 |
+
demo.queue(max_size=5).launch()
|