Commit
•
704ce7b
1
Parent(s):
58e928c
functional w dags
Browse files- .gitignore +1 -0
- app.py +372 -0
- non_well_spec.png +0 -0
- well_spec.png +0 -0
.gitignore
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
venv*
|
app.py
ADDED
@@ -0,0 +1,372 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# !pip install gradio -q
|
2 |
+
# !pip install transformers -q
|
3 |
+
|
4 |
+
# %%
|
5 |
+
import gradio as gr
|
6 |
+
import matplotlib.pyplot as plt
|
7 |
+
import numpy as np
|
8 |
+
import pandas as pd
|
9 |
+
import random
|
10 |
+
from matplotlib.ticker import MaxNLocator
|
11 |
+
from transformers import pipeline
|
12 |
+
|
13 |
+
# %%
|
14 |
+
MODEL_NAMES = [
|
15 |
+
"bert-base-uncased",
|
16 |
+
"roberta-base",
|
17 |
+
"bert-large-uncased",
|
18 |
+
"roberta-large",
|
19 |
+
]
|
20 |
+
OWN_MODEL_NAME = "add-a-model"
|
21 |
+
|
22 |
+
DECIMAL_PLACES = 1
|
23 |
+
EPS = 1e-5 # to avoid /0 errors
|
24 |
+
# %%
|
25 |
+
|
26 |
+
# Fire up the models
|
27 |
+
models = dict()
|
28 |
+
|
29 |
+
for bert_like in MODEL_NAMES:
|
30 |
+
models[bert_like] = pipeline("fill-mask", model=bert_like)
|
31 |
+
|
32 |
+
# %%
|
33 |
+
|
34 |
+
def clean_tokens(tokens):
|
35 |
+
return [token.strip() for token in tokens]
|
36 |
+
|
37 |
+
|
38 |
+
def prepare_text_for_masking(input_text, mask_token, gendered_tokens, split_key):
|
39 |
+
text_w_masks_list = [
|
40 |
+
mask_token if word.lower() in gendered_tokens else word
|
41 |
+
for word in input_text.split()
|
42 |
+
]
|
43 |
+
num_masks = len([m for m in text_w_masks_list if m == mask_token])
|
44 |
+
|
45 |
+
text_portions = " ".join(text_w_masks_list).split(split_key)
|
46 |
+
return text_portions, num_masks
|
47 |
+
|
48 |
+
|
49 |
+
def get_avg_prob_from_pipeline_outputs(mask_filled_text, gendered_token, num_preds):
|
50 |
+
pronoun_preds = [
|
51 |
+
sum(
|
52 |
+
[
|
53 |
+
pronoun["score"]
|
54 |
+
if pronoun["token_str"].strip().lower() in gendered_token
|
55 |
+
else 0.0
|
56 |
+
for pronoun in top_preds
|
57 |
+
]
|
58 |
+
)
|
59 |
+
for top_preds in mask_filled_text
|
60 |
+
]
|
61 |
+
return round(sum(pronoun_preds) / (EPS + num_preds) * 100, DECIMAL_PLACES)
|
62 |
+
|
63 |
+
|
64 |
+
|
65 |
+
|
66 |
+
def get_figure(df, gender, n_fit=1):
|
67 |
+
df = df.set_index("x-axis")
|
68 |
+
cols = df.columns
|
69 |
+
xs = list(range(len(df)))
|
70 |
+
ys = df[cols[0]]
|
71 |
+
fig, ax = plt.subplots()
|
72 |
+
# Trying small fig due to rendering issues on HF, not on VS Code
|
73 |
+
fig.set_figheight(3)
|
74 |
+
fig.set_figwidth(9)
|
75 |
+
|
76 |
+
# find stackoverflow reference
|
77 |
+
p, C_p = np.polyfit(xs, ys, n_fit, cov=1)
|
78 |
+
t = np.linspace(min(xs)-1, max(xs)+1, 10*len(xs))
|
79 |
+
TT = np.vstack([t**(n_fit-i) for i in range(n_fit+1)]).T
|
80 |
+
|
81 |
+
# matrix multiplication calculates the polynomial values
|
82 |
+
yi = np.dot(TT, p)
|
83 |
+
C_yi = np.dot(TT, np.dot(C_p, TT.T)) # C_y = TT*C_z*TT.T
|
84 |
+
sig_yi = np.sqrt(np.diag(C_yi)) # Standard deviations are sqrt of diagonal
|
85 |
+
|
86 |
+
ax.fill_between(t, yi+sig_yi, yi-sig_yi, alpha=.25)
|
87 |
+
ax.plot(t, yi, '-')
|
88 |
+
ax.plot(df, "ro")
|
89 |
+
ax.legend(list(df.columns))
|
90 |
+
|
91 |
+
ax.axis("tight")
|
92 |
+
ax.set_xlabel("Value injected into input text")
|
93 |
+
ax.set_title(f"Probability of predicting {gender} tokens.")
|
94 |
+
ax.set_ylabel(f"Softmax prob")
|
95 |
+
ax.tick_params(axis="x", labelrotation=5)
|
96 |
+
ax.set_ylim(0, 100)
|
97 |
+
return fig
|
98 |
+
|
99 |
+
|
100 |
+
|
101 |
+
# %%
|
102 |
+
def predict_masked_tokens(
|
103 |
+
model_name,
|
104 |
+
own_model_name,
|
105 |
+
group_a_tokens,
|
106 |
+
group_b_tokens,
|
107 |
+
indie_vars,
|
108 |
+
split_key,
|
109 |
+
normalizing,
|
110 |
+
n_fit,
|
111 |
+
input_text,
|
112 |
+
):
|
113 |
+
"""Run inference on input_text for each model type, returning df and plots of percentage
|
114 |
+
of gender pronouns predicted as female and male in each target text.
|
115 |
+
"""
|
116 |
+
if model_name not in MODEL_NAMES:
|
117 |
+
model = pipeline("fill-mask", model=own_model_name)
|
118 |
+
else:
|
119 |
+
model = models[model_name]
|
120 |
+
|
121 |
+
mask_token = model.tokenizer.mask_token
|
122 |
+
|
123 |
+
indie_vars_list = indie_vars.split(",")
|
124 |
+
|
125 |
+
group_a_tokens = clean_tokens(group_a_tokens.split(","))
|
126 |
+
group_b_tokens = clean_tokens(group_b_tokens.split(","))
|
127 |
+
|
128 |
+
text_segments, num_preds = prepare_text_for_masking(
|
129 |
+
input_text, mask_token, group_b_tokens + group_a_tokens, split_key
|
130 |
+
)
|
131 |
+
|
132 |
+
male_pronoun_preds = []
|
133 |
+
female_pronoun_preds = []
|
134 |
+
for indie_var in indie_vars_list:
|
135 |
+
target_text = f"{indie_var}".join(text_segments)
|
136 |
+
mask_filled_text = model(target_text)
|
137 |
+
# Quick hack as realized return type based on how many MASKs in text.
|
138 |
+
if type(mask_filled_text[0]) is not list:
|
139 |
+
mask_filled_text = [mask_filled_text]
|
140 |
+
|
141 |
+
female_pronoun_preds.append(
|
142 |
+
get_avg_prob_from_pipeline_outputs(
|
143 |
+
mask_filled_text, group_a_tokens, num_preds
|
144 |
+
)
|
145 |
+
)
|
146 |
+
male_pronoun_preds.append(
|
147 |
+
get_avg_prob_from_pipeline_outputs(
|
148 |
+
mask_filled_text, group_b_tokens, num_preds
|
149 |
+
)
|
150 |
+
)
|
151 |
+
|
152 |
+
if normalizing:
|
153 |
+
total_gendered_probs = np.add(female_pronoun_preds, male_pronoun_preds)
|
154 |
+
female_pronoun_preds = np.around(
|
155 |
+
np.divide(female_pronoun_preds, total_gendered_probs + EPS) * 100,
|
156 |
+
decimals=DECIMAL_PLACES,
|
157 |
+
)
|
158 |
+
male_pronoun_preds = np.around(
|
159 |
+
np.divide(male_pronoun_preds, total_gendered_probs + EPS) * 100,
|
160 |
+
decimals=DECIMAL_PLACES,
|
161 |
+
)
|
162 |
+
|
163 |
+
results_df = pd.DataFrame({"x-axis": indie_vars_list})
|
164 |
+
results_df["group_a"] = female_pronoun_preds
|
165 |
+
results_df["group_b"] = male_pronoun_preds
|
166 |
+
female_fig = get_figure(
|
167 |
+
results_df.drop("group_b", axis=1),
|
168 |
+
"group_a",
|
169 |
+
n_fit,
|
170 |
+
)
|
171 |
+
male_fig = get_figure(
|
172 |
+
results_df.drop("group_a", axis=1),
|
173 |
+
"group_b",
|
174 |
+
n_fit,
|
175 |
+
)
|
176 |
+
display_text = f"{random.choice(indie_vars_list)}".join(text_segments)
|
177 |
+
|
178 |
+
return (
|
179 |
+
display_text,
|
180 |
+
female_fig,
|
181 |
+
male_fig,
|
182 |
+
results_df,
|
183 |
+
)
|
184 |
+
|
185 |
+
|
186 |
+
truck_fn_example = [
|
187 |
+
MODEL_NAMES[2],
|
188 |
+
'',
|
189 |
+
', '.join(['truck', 'pickup']),
|
190 |
+
', '.join(['car', 'sedan']),
|
191 |
+
', '.join(['city','neighborhood','farm']),
|
192 |
+
'PLACE',
|
193 |
+
"True",
|
194 |
+
1,
|
195 |
+
]
|
196 |
+
def truck_1_fn():
|
197 |
+
return truck_fn_example + [
|
198 |
+
'He loaded up his truck and drove to the PLACE.'
|
199 |
+
]
|
200 |
+
|
201 |
+
def truck_2_fn():
|
202 |
+
return truck_fn_example + [
|
203 |
+
'He loaded up the bed of his truck and drove to the PLACE.'
|
204 |
+
]
|
205 |
+
|
206 |
+
|
207 |
+
# # %%
|
208 |
+
|
209 |
+
|
210 |
+
|
211 |
+
demo = gr.Blocks()
|
212 |
+
with demo:
|
213 |
+
gr.Markdown("# Spurious Correlation Evaluation for Pre-trained LLMs")
|
214 |
+
|
215 |
+
|
216 |
+
gr.Markdown("## Instructions for this Demo")
|
217 |
+
gr.Markdown(
|
218 |
+
"1) Click on one of the examples below to pre-populate the input fields."
|
219 |
+
)
|
220 |
+
gr.Markdown(
|
221 |
+
"2) Check out the pre-populated fields as you scroll down to the ['Hit Submit...'] button!"
|
222 |
+
)
|
223 |
+
gr.Markdown(
|
224 |
+
"3) Repeat steps (1) and (2) with more pre-populated inputs or with your own values in the input fields!"
|
225 |
+
)
|
226 |
+
|
227 |
+
|
228 |
+
gr.Markdown("""The pre-populated inputs below are for a demo example of a location-vs-vehicle-type spurious correlation.
|
229 |
+
We can see this spurious correlation largely disappears in the well-specified example text.
|
230 |
+
|
231 |
+
<p align="center">
|
232 |
+
<img src="file/non_well_spec.png" alt="results" width="300"/>
|
233 |
+
</p>
|
234 |
+
|
235 |
+
|
236 |
+
<p align="center">
|
237 |
+
<img src="file/well_spec.png" alt="results" width="300"/>
|
238 |
+
</p>
|
239 |
+
""")
|
240 |
+
|
241 |
+
gr.Markdown("## Example inputs")
|
242 |
+
gr.Markdown(
|
243 |
+
"Click a button below to pre-populate input fields with example values. Then scroll down to Hit Submit to generate predictions."
|
244 |
+
)
|
245 |
+
with gr.Row():
|
246 |
+
truck_1_gen = gr.Button("Click for non-well-specified(?) vehicle-type example inputs")
|
247 |
+
gr.Markdown("<-- Multiple solutions with low training error. LLM sensitive to spurious(?) correlations.")
|
248 |
+
|
249 |
+
truck_2_gen = gr.Button("Click for well-specified vehicle-type example inputs")
|
250 |
+
gr.Markdown("<-- Fewer solutions with low training error. LLM less sensitive to spurious(?) correlations.")
|
251 |
+
|
252 |
+
gr.Markdown("## Input fields")
|
253 |
+
gr.Markdown(
|
254 |
+
f"A) Pick a spectrum of comma separated values for text injection and x-axis."
|
255 |
+
)
|
256 |
+
|
257 |
+
with gr.Row():
|
258 |
+
group_a_tokens = gr.Textbox(
|
259 |
+
type="text",
|
260 |
+
lines=3,
|
261 |
+
label="A) To-MASK tokens A: Comma separated words that account for accumulated group A softmax probs",
|
262 |
+
)
|
263 |
+
|
264 |
+
group_b_tokens = gr.Textbox(
|
265 |
+
type="text",
|
266 |
+
lines=3,
|
267 |
+
label="B) To-MASK tokens B: Comma separated words that account for accumulated group B softmax probs",
|
268 |
+
)
|
269 |
+
|
270 |
+
with gr.Row():
|
271 |
+
x_axis = gr.Textbox(
|
272 |
+
type="text",
|
273 |
+
lines=3,
|
274 |
+
label="C) Comma separated values for text injection and x-axis",
|
275 |
+
)
|
276 |
+
|
277 |
+
gr.Markdown("D) Pick a pre-loaded BERT-family model of interest on the right.")
|
278 |
+
gr.Markdown(
|
279 |
+
f"Or E) select `{OWN_MODEL_NAME}`, then add the mame of any other Hugging Face model that supports the [fill-mask](https://huggingface.co/models?pipeline_tag=fill-mask) task on the right (note: this may take some time to load)."
|
280 |
+
)
|
281 |
+
|
282 |
+
with gr.Row():
|
283 |
+
model_name = gr.Radio(
|
284 |
+
MODEL_NAMES + [OWN_MODEL_NAME],
|
285 |
+
type="value",
|
286 |
+
label="D) BERT-like model.",
|
287 |
+
)
|
288 |
+
own_model_name = gr.Textbox(
|
289 |
+
label="E) If you selected an 'add-a-model' model, put any Hugging Face pipeline model name (that supports the fill-mask task) here.",
|
290 |
+
)
|
291 |
+
|
292 |
+
gr.Markdown(
|
293 |
+
"F) Pick if you want to the predictions normalied to only those from group A or B."
|
294 |
+
)
|
295 |
+
gr.Markdown(
|
296 |
+
"G) Also tell the demo what special token you will use in your input text, that you would like replaced with the spectrum of values you listed above."
|
297 |
+
)
|
298 |
+
gr.Markdown(
|
299 |
+
"And H) the degree of polynomial fit used for high-lighting potential spurious association."
|
300 |
+
)
|
301 |
+
|
302 |
+
with gr.Row():
|
303 |
+
to_normalize = gr.Dropdown(
|
304 |
+
["False", "True"],
|
305 |
+
label="D) Normalize model's predictions?",
|
306 |
+
type="index",
|
307 |
+
)
|
308 |
+
place_holder = gr.Textbox(
|
309 |
+
label="E) Special token place-holder",
|
310 |
+
)
|
311 |
+
n_fit = gr.Dropdown(
|
312 |
+
list(range(1, 5)),
|
313 |
+
label="F) Degree of polynomial fit",
|
314 |
+
type="value",
|
315 |
+
)
|
316 |
+
|
317 |
+
gr.Markdown(
|
318 |
+
"I) Finally, add input text that includes at least one of the '`To-MASK`' tokens from (A) or (B) and one place-holder token from (G)."
|
319 |
+
)
|
320 |
+
|
321 |
+
with gr.Row():
|
322 |
+
input_text = gr.Textbox(
|
323 |
+
lines=2,
|
324 |
+
label="I) Input text with a '`To-MASK`' and place-holder token",
|
325 |
+
)
|
326 |
+
|
327 |
+
gr.Markdown("## Outputs!")
|
328 |
+
with gr.Row():
|
329 |
+
btn = gr.Button("Hit submit to generate predictions!")
|
330 |
+
|
331 |
+
with gr.Row():
|
332 |
+
sample_text = gr.Textbox(
|
333 |
+
type="text", label="Output text: Sample of text fed to model"
|
334 |
+
)
|
335 |
+
with gr.Row():
|
336 |
+
female_fig = gr.Plot(type="auto")
|
337 |
+
male_fig = gr.Plot(type="auto")
|
338 |
+
with gr.Row():
|
339 |
+
df = gr.Dataframe(
|
340 |
+
show_label=True,
|
341 |
+
overflow_row_behaviour="show_ends",
|
342 |
+
label="Table of softmax probability for grouped predictions",
|
343 |
+
)
|
344 |
+
|
345 |
+
with gr.Row():
|
346 |
+
truck_1_gen.click(truck_1_fn, inputs=[], outputs=[model_name, own_model_name, group_a_tokens, group_b_tokens,
|
347 |
+
x_axis, place_holder, to_normalize, n_fit, input_text])
|
348 |
+
|
349 |
+
truck_2_gen.click(truck_2_fn, inputs=[], outputs=[model_name, own_model_name, group_a_tokens, group_b_tokens,
|
350 |
+
x_axis, place_holder, to_normalize, n_fit, input_text])
|
351 |
+
|
352 |
+
btn.click(
|
353 |
+
predict_masked_tokens,
|
354 |
+
inputs=[
|
355 |
+
model_name,
|
356 |
+
own_model_name,
|
357 |
+
group_a_tokens,
|
358 |
+
group_b_tokens,
|
359 |
+
x_axis,
|
360 |
+
place_holder,
|
361 |
+
to_normalize,
|
362 |
+
n_fit,
|
363 |
+
input_text,
|
364 |
+
],
|
365 |
+
outputs=[sample_text, female_fig, male_fig, df],
|
366 |
+
)
|
367 |
+
|
368 |
+
demo.launch(debug=True, share=True)
|
369 |
+
|
370 |
+
# %%
|
371 |
+
|
372 |
+
|
non_well_spec.png
ADDED
![]() |
well_spec.png
ADDED
![]() |