Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
|
@@ -4,20 +4,19 @@ import argparse
|
|
| 4 |
import pandas as pd
|
| 5 |
from functools import partial
|
| 6 |
import subprocess
|
| 7 |
-
"""
|
| 8 |
-
Eventually add filters based on headers
|
| 9 |
-
"""
|
| 10 |
-
HEADERS = ["__index_level_0__", "problem", "username", "entrypoint", "tests_passed", "total_tests"]
|
| 11 |
-
DATATYPES = ["number", "str", "str", "str", "number", "number"]
|
| 12 |
|
| 13 |
-
|
| 14 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
|
| 16 |
def capture_output(prompt, completion, prints):
|
| 17 |
code = "\n".join([prompt, " "+" \n".join(completion.split("\n")), prints])
|
| 18 |
outputs = subprocess.run(["python", "-c", code], stdout=subprocess.PIPE, stderr=subprocess.PIPE)
|
| 19 |
stderr = gr.Textbox(outputs.stderr.decode("utf-8").strip(), label="Code Errors", type="text")
|
| 20 |
-
stdout = gr.Code(outputs.stdout.decode("utf-8").strip(), label="Code
|
| 21 |
return stderr, stdout
|
| 22 |
|
| 23 |
def update_components(
|
|
@@ -27,11 +26,7 @@ def update_components(
|
|
| 27 |
success_data,
|
| 28 |
prompt,
|
| 29 |
submitted_text,
|
| 30 |
-
completion,
|
| 31 |
assertions,
|
| 32 |
-
prints,
|
| 33 |
-
code_err,
|
| 34 |
-
code_output
|
| 35 |
):
|
| 36 |
if isinstance(ds, gr.State):
|
| 37 |
ds = ds.value
|
|
@@ -57,31 +52,20 @@ def update_components(
|
|
| 57 |
row = row.iloc[0]
|
| 58 |
prompt = gr.Code(row["prompt"], language="python", label="Prompt")
|
| 59 |
submitted_text = gr.Textbox(row["submitted_text"], type="text", label="Submitted Text")
|
| 60 |
-
completion = gr.Code(row["completion"], language="python", label="Completion")
|
| 61 |
assertions = gr.Code(row["assertions"], language="python", label="Assertions")
|
| 62 |
-
prints = gr.Code(row["prints"], language="python", label="Prints")
|
| 63 |
-
code_err = gr.Textbox("__stderr__", label="Code Errors", type="text")
|
| 64 |
-
code_output = gr.Code("__stdout__", language="python", label="Code Outputs")
|
| 65 |
slider = gr.Slider(0, len(ds) - 1, step=1, label="Problem ID (click and arrow keys to navigate):", value=slider)
|
| 66 |
return [slider, header_data, success_data, prompt, submitted_text,
|
| 67 |
-
|
| 68 |
|
| 69 |
def filter_by(
|
| 70 |
dataset_name,
|
| 71 |
dataset_split,
|
| 72 |
-
fs_box, ls_box, ff_box, lf_box, f_box, l_box, is_success_box,
|
| 73 |
problem_box,
|
| 74 |
student_box,
|
| 75 |
slider,
|
| 76 |
*components_to_update):
|
| 77 |
ds = load_dataset(dataset_name, split=dataset_split)
|
| 78 |
-
success_boxes = [fs_box, ls_box, ff_box, lf_box, f_box, l_box, is_success_box]
|
| 79 |
ds = ds.to_pandas()
|
| 80 |
-
labels = ["is_first_success","is_last_success","is_first_failure","is_last_failure",
|
| 81 |
-
"first_attempt","last_attempt","is_success"]
|
| 82 |
-
for label, box in zip(labels, success_boxes):
|
| 83 |
-
if box:
|
| 84 |
-
ds = ds[ds[label] == box]
|
| 85 |
|
| 86 |
if problem_box != None:
|
| 87 |
ds = ds[ds["problem"] == problem_box]
|
|
@@ -89,20 +73,16 @@ def filter_by(
|
|
| 89 |
if student_box != None:
|
| 90 |
ds = ds[ds["username"] == student_box]
|
| 91 |
|
| 92 |
-
|
| 93 |
-
return [dataset, *update_components(ds, 0, *components_to_update)]
|
| 94 |
|
| 95 |
def next_example(ds, *components):
|
| 96 |
slider_value = components[0]
|
| 97 |
-
print(slider_value)
|
| 98 |
-
# Adjust slider value within bounds (min 0, max len(ds)-1)
|
| 99 |
new_slider_value = int(slider_value)+1 if slider_value < len(ds)-1 else len(ds)-1
|
| 100 |
lesscomponents = components[1:]
|
| 101 |
return update_components(ds, new_slider_value, *lesscomponents)
|
| 102 |
|
| 103 |
def prev_example(ds, *components):
|
| 104 |
slider_value = components[0]
|
| 105 |
-
# Adjust slider value within bounds (min 0, max len(ds)-1)
|
| 106 |
new_slider_value = int(slider_value)-1 if slider_value > 0 else 0
|
| 107 |
lesscomponents = components[1:]
|
| 108 |
return update_components(ds, new_slider_value, *lesscomponents)
|
|
@@ -143,61 +123,31 @@ def main(args):
|
|
| 143 |
prev_btn = gr.Button("Previous")
|
| 144 |
next_btn = gr.Button("Next")
|
| 145 |
submitted_text = gr.Textbox("__submitted_text__", type="text", label="Submitted Text")
|
| 146 |
-
completion = gr.Code("__completion__", language="python", label="Completion")
|
| 147 |
-
|
| 148 |
|
|
|
|
| 149 |
with gr.Row():
|
| 150 |
assertions = gr.Code("__assertions__", language="python", label="Assertions")
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
runbtn = gr.Button("Run this code")
|
| 154 |
-
|
| 155 |
-
with gr.Column():
|
| 156 |
-
with gr.Row():
|
| 157 |
-
code_output = gr.Code("__stdout__", language="python", label="Code Outputs")
|
| 158 |
-
code_err = gr.Textbox("__stderr__", label="Code Errors", type="text")
|
| 159 |
-
|
| 160 |
-
gr.Markdown("**Logging**\n")
|
| 161 |
-
with gr.Column():
|
| 162 |
-
with gr.Row():
|
| 163 |
-
flagbtn = gr.Button("Flag this example to log file")
|
| 164 |
-
flagout = gr.Textbox(label="Num examples in logfile")
|
| 165 |
|
| 166 |
# updates
|
| 167 |
-
# run code
|
| 168 |
-
runbtn.click(fn=capture_output, inputs=[prompt, completion, prints], outputs=[code_err,code_output])
|
| 169 |
# change example on slider change
|
| 170 |
-
components = [slider, header_data, success_data, prompt, submitted_text,
|
| 171 |
slider.input(fn=update_components, inputs=[dataset, *components], outputs=components)
|
| 172 |
|
| 173 |
prev_btn.click(fn=prev_example, inputs=[dataset, *components], outputs=components)
|
| 174 |
next_btn.click(fn=next_example, inputs=[dataset, *components], outputs=components)
|
| 175 |
|
| 176 |
-
# log
|
| 177 |
-
callback.setup(components, "flagged_data_points")
|
| 178 |
-
flagbtn.click(lambda *args: callback.flag(list(args)), components, flagout,
|
| 179 |
-
preprocess=False, show_progress="full", trigger_mode="once")
|
| 180 |
-
|
| 181 |
# add filtering options
|
| 182 |
gr.Markdown("**Filtering (reload to clear all filters)**\n")
|
| 183 |
with gr.Row():
|
| 184 |
with gr.Column():
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
with gr.Column():
|
| 188 |
-
ff_box = gr.Checkbox(label="is_first_failure")
|
| 189 |
-
lf_box = gr.Checkbox(label="is_last_failure")
|
| 190 |
-
with gr.Column():
|
| 191 |
-
f_box = gr.Checkbox(label="first_attempt")
|
| 192 |
-
l_box = gr.Checkbox(label="last_attempt")
|
| 193 |
-
is_success_box = gr.Checkbox(label="is_success")
|
| 194 |
-
success_boxes = [fs_box, ls_box, ff_box, lf_box, f_box, l_box, is_success_box]
|
| 195 |
-
problem_box = gr.Dropdown(label="problem", choices = problem_names)
|
| 196 |
-
student_box = gr.Dropdown(label="username", choices = student_usernames)
|
| 197 |
filter_btn = gr.Button("Filter")
|
| 198 |
|
| 199 |
-
filter_btn.click(fn=partial(filter_by, args.dataset, args.split), inputs=[
|
| 200 |
outputs=[dataset, *components])
|
|
|
|
| 201 |
|
| 202 |
demo.launch(share=args.share)
|
| 203 |
|
|
@@ -207,5 +157,4 @@ if __name__ == "__main__":
|
|
| 207 |
parser.add_argument("--split", type=str, default="test")
|
| 208 |
parser.add_argument("--share", action="store_true")
|
| 209 |
args = parser.parse_args()
|
| 210 |
-
main(args)
|
| 211 |
-
|
|
|
|
| 4 |
import pandas as pd
|
| 5 |
from functools import partial
|
| 6 |
import subprocess
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
|
| 8 |
+
# Headers and datatypes for the remaining columns
|
| 9 |
+
HEADERS = ["__index_level_0__", "problem", "username", "entrypoint", "submitted_text", "prompt", "subset"]
|
| 10 |
+
DATATYPES = ["number", "str", "str", "str", "str", "str", "str"]
|
| 11 |
+
|
| 12 |
+
SUCCESS_HEADERS = ["subset"]
|
| 13 |
+
SUCCESS_DATATYPES = ["str"]
|
| 14 |
|
| 15 |
def capture_output(prompt, completion, prints):
|
| 16 |
code = "\n".join([prompt, " "+" \n".join(completion.split("\n")), prints])
|
| 17 |
outputs = subprocess.run(["python", "-c", code], stdout=subprocess.PIPE, stderr=subprocess.PIPE)
|
| 18 |
stderr = gr.Textbox(outputs.stderr.decode("utf-8").strip(), label="Code Errors", type="text")
|
| 19 |
+
stdout = gr.Code(outputs.stdout.decode("utf-8").strip(), label="Code Outputs", language="python")
|
| 20 |
return stderr, stdout
|
| 21 |
|
| 22 |
def update_components(
|
|
|
|
| 26 |
success_data,
|
| 27 |
prompt,
|
| 28 |
submitted_text,
|
|
|
|
| 29 |
assertions,
|
|
|
|
|
|
|
|
|
|
| 30 |
):
|
| 31 |
if isinstance(ds, gr.State):
|
| 32 |
ds = ds.value
|
|
|
|
| 52 |
row = row.iloc[0]
|
| 53 |
prompt = gr.Code(row["prompt"], language="python", label="Prompt")
|
| 54 |
submitted_text = gr.Textbox(row["submitted_text"], type="text", label="Submitted Text")
|
|
|
|
| 55 |
assertions = gr.Code(row["assertions"], language="python", label="Assertions")
|
|
|
|
|
|
|
|
|
|
| 56 |
slider = gr.Slider(0, len(ds) - 1, step=1, label="Problem ID (click and arrow keys to navigate):", value=slider)
|
| 57 |
return [slider, header_data, success_data, prompt, submitted_text,
|
| 58 |
+
assertions]
|
| 59 |
|
| 60 |
def filter_by(
|
| 61 |
dataset_name,
|
| 62 |
dataset_split,
|
|
|
|
| 63 |
problem_box,
|
| 64 |
student_box,
|
| 65 |
slider,
|
| 66 |
*components_to_update):
|
| 67 |
ds = load_dataset(dataset_name, split=dataset_split)
|
|
|
|
| 68 |
ds = ds.to_pandas()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 69 |
|
| 70 |
if problem_box != None:
|
| 71 |
ds = ds[ds["problem"] == problem_box]
|
|
|
|
| 73 |
if student_box != None:
|
| 74 |
ds = ds[ds["username"] == student_box]
|
| 75 |
|
| 76 |
+
return [ds, *update_components(ds, 0, *components_to_update)]
|
|
|
|
| 77 |
|
| 78 |
def next_example(ds, *components):
|
| 79 |
slider_value = components[0]
|
|
|
|
|
|
|
| 80 |
new_slider_value = int(slider_value)+1 if slider_value < len(ds)-1 else len(ds)-1
|
| 81 |
lesscomponents = components[1:]
|
| 82 |
return update_components(ds, new_slider_value, *lesscomponents)
|
| 83 |
|
| 84 |
def prev_example(ds, *components):
|
| 85 |
slider_value = components[0]
|
|
|
|
| 86 |
new_slider_value = int(slider_value)-1 if slider_value > 0 else 0
|
| 87 |
lesscomponents = components[1:]
|
| 88 |
return update_components(ds, new_slider_value, *lesscomponents)
|
|
|
|
| 123 |
prev_btn = gr.Button("Previous")
|
| 124 |
next_btn = gr.Button("Next")
|
| 125 |
submitted_text = gr.Textbox("__submitted_text__", type="text", label="Submitted Text")
|
|
|
|
|
|
|
| 126 |
|
| 127 |
+
|
| 128 |
with gr.Row():
|
| 129 |
assertions = gr.Code("__assertions__", language="python", label="Assertions")
|
| 130 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 131 |
|
| 132 |
# updates
|
|
|
|
|
|
|
| 133 |
# change example on slider change
|
| 134 |
+
components = [slider, header_data, success_data, prompt, submitted_text, assertions]
|
| 135 |
slider.input(fn=update_components, inputs=[dataset, *components], outputs=components)
|
| 136 |
|
| 137 |
prev_btn.click(fn=prev_example, inputs=[dataset, *components], outputs=components)
|
| 138 |
next_btn.click(fn=next_example, inputs=[dataset, *components], outputs=components)
|
| 139 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 140 |
# add filtering options
|
| 141 |
gr.Markdown("**Filtering (reload to clear all filters)**\n")
|
| 142 |
with gr.Row():
|
| 143 |
with gr.Column():
|
| 144 |
+
problem_box = gr.Dropdown(label="problem", choices=problem_names)
|
| 145 |
+
student_box = gr.Dropdown(label="username", choices=student_usernames)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 146 |
filter_btn = gr.Button("Filter")
|
| 147 |
|
| 148 |
+
filter_btn.click(fn=partial(filter_by, args.dataset, args.split), inputs=[problem_box, student_box, *components],
|
| 149 |
outputs=[dataset, *components])
|
| 150 |
+
|
| 151 |
|
| 152 |
demo.launch(share=args.share)
|
| 153 |
|
|
|
|
| 157 |
parser.add_argument("--split", type=str, default="test")
|
| 158 |
parser.add_argument("--share", action="store_true")
|
| 159 |
args = parser.parse_args()
|
| 160 |
+
main(args)
|
|
|