Emily McMilin
commited on
Commit
·
eef17b3
1
Parent(s):
c8a64af
updated for parity with ICML version
Browse files
app.py
CHANGED
@@ -7,9 +7,8 @@ import random
|
|
7 |
from matplotlib.ticker import MaxNLocator
|
8 |
from transformers import pipeline
|
9 |
|
10 |
-
MODEL_NAMES = ["bert-base-uncased",
|
11 |
-
|
12 |
-
OWN_MODEL_NAME = 'add-your-own'
|
13 |
|
14 |
DECIMAL_PLACES = 1
|
15 |
EPS = 1e-5 # to avoid /0 errors
|
@@ -290,45 +289,47 @@ description = """
|
|
290 |
|
291 |
"""
|
292 |
|
293 |
-
|
294 |
-
|
|
|
295 |
'',
|
296 |
-
', '.join(
|
297 |
-
'
|
298 |
"False",
|
299 |
1,
|
300 |
-
'She was
|
301 |
]
|
302 |
|
303 |
-
|
|
|
304 |
MODEL_NAMES[0],
|
305 |
'',
|
306 |
-
', '.join(
|
307 |
-
'
|
308 |
"False",
|
309 |
-
|
310 |
-
'She
|
311 |
]
|
312 |
|
313 |
|
314 |
subreddit_example = [
|
315 |
-
MODEL_NAMES[
|
316 |
'',
|
317 |
', '.join(SUBREDDITS),
|
318 |
'SUBREDDIT',
|
319 |
"False",
|
320 |
1,
|
321 |
-
'
|
322 |
]
|
323 |
|
324 |
own_model_example = [
|
325 |
OWN_MODEL_NAME,
|
326 |
-
'
|
327 |
', '.join(DATES),
|
328 |
'DATE',
|
329 |
"False",
|
330 |
-
|
331 |
-
'
|
332 |
]
|
333 |
|
334 |
|
@@ -351,39 +352,42 @@ def your_fn():
|
|
351 |
# %%
|
352 |
demo = gr.Blocks()
|
353 |
with demo:
|
354 |
-
gr.Markdown("
|
355 |
-
gr.Markdown("
|
356 |
-
|
357 |
-
gr.Markdown("
|
358 |
-
|
|
|
|
|
|
|
359 |
gr.Markdown("2) Check out the pre-populated fields as you scroll down to the ['Hit Submit...'] button!")
|
360 |
gr.Markdown("3) Repeat steps (1) and (2) with more pre-populated inputs or with your own values in the input fields!")
|
361 |
|
362 |
-
gr.Markdown("
|
363 |
gr.Markdown("Click a button below to pre-populate input fields with example values. Then scroll down to Hit Submit to generate predictions.")
|
364 |
with gr.Row():
|
365 |
-
gr.Markdown("X-axis sorted by older to more recent dates:")
|
366 |
date_gen = gr.Button('Click for date example inputs')
|
|
|
367 |
|
368 |
-
gr.Markdown(
|
369 |
-
"X-axis sorted by bottom 10 and top 10 [Global Gender Gap](https://www3.weforum.org/docs/WEF_GGGR_2021.pdf) ranked countries:")
|
370 |
place_gen = gr.Button('Click for country example inputs')
|
371 |
-
|
372 |
gr.Markdown(
|
373 |
-
"
|
|
|
374 |
subreddit_gen = gr.Button('Click for Subreddit example inputs')
|
|
|
|
|
375 |
|
376 |
-
gr.
|
377 |
-
|
378 |
|
379 |
-
gr.Markdown("
|
380 |
gr.Markdown(
|
381 |
-
f"A) Pick a spectrum of comma separated values for text injection and x-axis
|
382 |
|
383 |
with gr.Row():
|
384 |
x_axis = gr.Textbox(
|
385 |
-
lines=
|
386 |
-
label="A)
|
387 |
)
|
388 |
|
389 |
|
@@ -394,15 +398,15 @@ with demo:
|
|
394 |
model_name = gr.Radio(
|
395 |
MODEL_NAMES + [OWN_MODEL_NAME],
|
396 |
type="value",
|
397 |
-
label="B)
|
398 |
)
|
399 |
own_model_name = gr.Textbox(
|
400 |
-
label="C) If you selected an 'add-
|
401 |
)
|
402 |
|
403 |
gr.Markdown("D) Pick if you want to the predictions normalied to these gendered terms only.")
|
404 |
gr.Markdown("E) 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.")
|
405 |
-
gr.Markdown("And F) the degree of polynomial fit used for high-lighting
|
406 |
|
407 |
|
408 |
with gr.Row():
|
@@ -412,11 +416,11 @@ with demo:
|
|
412 |
type="index",
|
413 |
)
|
414 |
place_holder = gr.Textbox(
|
415 |
-
label="E) Special token place-holder
|
416 |
)
|
417 |
n_fit = gr.Dropdown(
|
418 |
list(range(1, 5)),
|
419 |
-
label="F) Degree of polynomial fit
|
420 |
type="value",
|
421 |
)
|
422 |
|
@@ -425,11 +429,11 @@ with demo:
|
|
425 |
|
426 |
with gr.Row():
|
427 |
input_text = gr.Textbox(
|
428 |
-
lines=
|
429 |
-
label="G) Input text
|
430 |
)
|
431 |
|
432 |
-
gr.Markdown("
|
433 |
#gr.Markdown("Scroll down and 'Hit Submit'!")
|
434 |
with gr.Row():
|
435 |
btn = gr.Button("Hit submit to generate predictions!")
|
@@ -465,50 +469,6 @@ with demo:
|
|
465 |
outputs=[sample_text, female_fig, male_fig, df])
|
466 |
|
467 |
|
468 |
-
gr.Markdown("### How does this work?")
|
469 |
-
gr.Markdown("We are able to test the pre-trained LLMs without any modification to the models, as the gender-pronoun prediction task is simply a special case of the masked language modeling (MLM) task, with which all these models were pre-trained. Rather than random masking, the gender-pronoun prediction task masks only non-gender-neutral terms (listed in prior [Space](https://huggingface.co/spaces/emilylearning/causing_gender_pronouns_two)).")
|
470 |
-
gr.Markdown("For the pre-trained LLMs the final prediction is a softmax over the entire tokenizer's vocabulary, from which we sum up the portion of the probability mass from the top five prediction words that are gendered terms (and normalize or not, based on selected preference above.")
|
471 |
-
|
472 |
-
|
473 |
-
|
474 |
-
gr.Markdown("### What is Causing these Spurious Correlations?")
|
475 |
-
|
476 |
-
gr.Markdown("Spurious correlations are often considered undesirable, as they do not match our intuition about the real-world domain from which we derive samples for inference-time prediction.")
|
477 |
-
gr.Markdown("Selection of samples into datasets can be a zero-sum-game, with even our high quality datasets forced to trade off one for another, thus inducing selection bias into the learned associations of the model.")
|
478 |
-
|
479 |
-
gr.Markdown("### Dose-response Relationship")
|
480 |
-
gr.Markdown("One intuitive way to see the impact that changing one variable may have upon another is to look for a dose-response relationship, in which a larger intervention in the treatment (the value in text form injected in the otherwise unchanged text sample) produces a larger response in the output (the softmax probability of a gendered pronoun).")
|
481 |
-
gr.Markdown("This dose-response plot requires a range of values along which we may see a spectrum of gender representation (or misrepresentation) in our datasets.")
|
482 |
-
|
483 |
-
|
484 |
-
gr.Markdown("### Data Generating Process")
|
485 |
-
gr.Markdown("To pick values below that are most likely to cause spurious correlations, it helps to make some assumptions about the training dataset's likely data generating process, and where selection bias may come in.")
|
486 |
-
|
487 |
-
gr.Markdown("A plausible data generating processes for Wiki-Bio and Reddit datasets is shown as a DAG below. The variables `W` : birth place, birth date or subreddit interest, and `G`: gender, are both independent variables that have no ancestral variables. However, `W` and `G` may have a role in causing one's access, `Z`. In the case of Wiki-Bio a functional form of `Z` may capture the general trend that access has become less gender-dependent over time, but not in every place. In the case of Reddit TLDR, `Z` may capture that despite some subreddits having gender-neutral topics, the specific style of moderation and community in the subreddit may reduce access to some genders.")
|
488 |
-
|
489 |
-
gr.Markdown("This DAG structure is prone to collider bias between `W` and `G` when conditioning on access, `Z`. In other words, although in real life *place*, *date*, and (subreddit) *interest* vs *gender* are unconditionally independent, when we condition on their common effect, *access*, they become unconditionally dependent.")
|
490 |
-
|
491 |
-
gr.Markdown("The obvious solution to not condition on access is unavailable to us, as we are required to in order to represent the process of selection into the dataset. Thus, a statistical relationship between `W` and `G` can be induced by the dataset formation, leading to possible spurious correlations, as shown here.")
|
492 |
-
|
493 |
-
|
494 |
-
|
495 |
-
gr.Markdown("""
|
496 |
-
<style>
|
497 |
-
img {
|
498 |
-
width: 30%;
|
499 |
-
max-width: 600px;
|
500 |
-
}
|
501 |
-
</style>
|
502 |
-
<center>
|
503 |
-
<img src="https://www.dropbox.com/s/4f07djirinl2qvy/show_g_crop.png?raw=1"
|
504 |
-
alt="DAG of possible data generating process for datasets used in training some of our LLMs.">
|
505 |
-
</center>
|
506 |
-
""")
|
507 |
-
|
508 |
-
gr.Markdown("### I Don't Buy It")
|
509 |
-
gr.Markdown("See something wrong above? Do you think we cherry picked our examples? Try your own, including your own x-axis. Think we cherry picked LLMs? Try the `add-your-own` model option. This demo _should_ work with any Hugging Face model that supports the [fill-mask](https://huggingface.co/models?pipeline_tag=fill-mask) task.")
|
510 |
-
gr.Markdown("Think our data generating process is wrong, or found an interesting spurious correlation you'd like to set as a default example? Use the community tab to discuss or pull request your fix.")
|
511 |
-
|
512 |
demo.launch(debug=True)
|
513 |
|
514 |
|
|
|
7 |
from matplotlib.ticker import MaxNLocator
|
8 |
from transformers import pipeline
|
9 |
|
10 |
+
MODEL_NAMES = ["bert-base-uncased", "roberta-base", "bert-large-uncased", "roberta-large"]
|
11 |
+
OWN_MODEL_NAME = 'add-a-model'
|
|
|
12 |
|
13 |
DECIMAL_PLACES = 1
|
14 |
EPS = 1e-5 # to avoid /0 errors
|
|
|
289 |
|
290 |
"""
|
291 |
|
292 |
+
|
293 |
+
date_example = [
|
294 |
+
MODEL_NAMES[1],
|
295 |
'',
|
296 |
+
', '.join(DATES),
|
297 |
+
'DATE',
|
298 |
"False",
|
299 |
1,
|
300 |
+
'She was a teenager in DATE.'
|
301 |
]
|
302 |
|
303 |
+
|
304 |
+
place_example = [
|
305 |
MODEL_NAMES[0],
|
306 |
'',
|
307 |
+
', '.join(PLACES),
|
308 |
+
'PLACE',
|
309 |
"False",
|
310 |
+
1,
|
311 |
+
'She became an adult in PLACE.'
|
312 |
]
|
313 |
|
314 |
|
315 |
subreddit_example = [
|
316 |
+
MODEL_NAMES[3],
|
317 |
'',
|
318 |
', '.join(SUBREDDITS),
|
319 |
'SUBREDDIT',
|
320 |
"False",
|
321 |
1,
|
322 |
+
'She was a kid. SUBREDDIT.'
|
323 |
]
|
324 |
|
325 |
own_model_example = [
|
326 |
OWN_MODEL_NAME,
|
327 |
+
'emilyalsentzer/Bio_ClinicalBERT',
|
328 |
', '.join(DATES),
|
329 |
'DATE',
|
330 |
"False",
|
331 |
+
1,
|
332 |
+
'She was exposed to the virus in DATE.'
|
333 |
]
|
334 |
|
335 |
|
|
|
352 |
# %%
|
353 |
demo = gr.Blocks()
|
354 |
with demo:
|
355 |
+
gr.Markdown("# Spurious Correlation Evaluation for Pre-trained LLMs")
|
356 |
+
gr.Markdown("Find spurious correlations between seemingly independent variables (for example between `gender` and `time`) in almost any BERT-like LLM on Hugging Face, below.")
|
357 |
+
|
358 |
+
gr.Markdown("See why this happens how in our paper, [Selection Bias Induced Spurious Correlations in Large Language Models](https://arxiv.org/pdf/2207.08982.pdf), presented at [ICML 2022 Workshop on Spurious Correlations, Invariance, and Stability](https://sites.google.com/view/scis-workshop/home).")
|
359 |
+
|
360 |
+
|
361 |
+
gr.Markdown("## Instructions for this Demo")
|
362 |
+
gr.Markdown("1) Click on one of the examples below (where we sweep through a spectrum of `places`, `dates` and `subreddits`) to pre-populate the input fields.")
|
363 |
gr.Markdown("2) Check out the pre-populated fields as you scroll down to the ['Hit Submit...'] button!")
|
364 |
gr.Markdown("3) Repeat steps (1) and (2) with more pre-populated inputs or with your own values in the input fields!")
|
365 |
|
366 |
+
gr.Markdown("## Example inputs")
|
367 |
gr.Markdown("Click a button below to pre-populate input fields with example values. Then scroll down to Hit Submit to generate predictions.")
|
368 |
with gr.Row():
|
|
|
369 |
date_gen = gr.Button('Click for date example inputs')
|
370 |
+
gr.Markdown("<-- x-axis sorted by older to more recent dates:")
|
371 |
|
|
|
|
|
372 |
place_gen = gr.Button('Click for country example inputs')
|
|
|
373 |
gr.Markdown(
|
374 |
+
"<-- x-axis sorted by bottom 10 and top 10 [Global Gender Gap](https://www3.weforum.org/docs/WEF_GGGR_2021.pdf) ranked countries:")
|
375 |
+
|
376 |
subreddit_gen = gr.Button('Click for Subreddit example inputs')
|
377 |
+
gr.Markdown(
|
378 |
+
"<-- x-axis sorted in order of increasing self-identified female participation (see [bburky](http://bburky.com/subredditgenderratios/)): ")
|
379 |
|
380 |
+
your_gen = gr.Button('Add-a-model example inputs')
|
381 |
+
gr.Markdown("<-- x-axis dates, with your own model loaded! (If first time, try another example, it can take a while to load new model.)")
|
382 |
|
383 |
+
gr.Markdown("## Input fields")
|
384 |
gr.Markdown(
|
385 |
+
f"A) Pick a spectrum of comma separated values for text injection and x-axis.")
|
386 |
|
387 |
with gr.Row():
|
388 |
x_axis = gr.Textbox(
|
389 |
+
lines=3,
|
390 |
+
label="A) Comma separated values for text injection and x-axis",
|
391 |
)
|
392 |
|
393 |
|
|
|
398 |
model_name = gr.Radio(
|
399 |
MODEL_NAMES + [OWN_MODEL_NAME],
|
400 |
type="value",
|
401 |
+
label="B) BERT-like model.",
|
402 |
)
|
403 |
own_model_name = gr.Textbox(
|
404 |
+
label="C) If you selected an 'add-a-model' model, put any Hugging Face pipeline model name (that supports the fill-mask task) here.",
|
405 |
)
|
406 |
|
407 |
gr.Markdown("D) Pick if you want to the predictions normalied to these gendered terms only.")
|
408 |
gr.Markdown("E) 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.")
|
409 |
+
gr.Markdown("And F) the degree of polynomial fit used for high-lighting potential spurious association.")
|
410 |
|
411 |
|
412 |
with gr.Row():
|
|
|
416 |
type="index",
|
417 |
)
|
418 |
place_holder = gr.Textbox(
|
419 |
+
label="E) Special token place-holder",
|
420 |
)
|
421 |
n_fit = gr.Dropdown(
|
422 |
list(range(1, 5)),
|
423 |
+
label="F) Degree of polynomial fit",
|
424 |
type="value",
|
425 |
)
|
426 |
|
|
|
429 |
|
430 |
with gr.Row():
|
431 |
input_text = gr.Textbox(
|
432 |
+
lines=2,
|
433 |
+
label="G) Input text with pronouns and place-holder token",
|
434 |
)
|
435 |
|
436 |
+
gr.Markdown("## Outputs!")
|
437 |
#gr.Markdown("Scroll down and 'Hit Submit'!")
|
438 |
with gr.Row():
|
439 |
btn = gr.Button("Hit submit to generate predictions!")
|
|
|
469 |
outputs=[sample_text, female_fig, male_fig, df])
|
470 |
|
471 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
472 |
demo.launch(debug=True)
|
473 |
|
474 |
|