emilylearning commited on
Commit
058f4ca
·
1 Parent(s): 7234ad2

Update display text for paper link. Add TLDR instructions

Browse files
Files changed (1) hide show
  1. app.py +5 -2
app.py CHANGED
@@ -209,7 +209,7 @@ with demo:
209
  gr.Markdown(
210
  "#### LLMs are pretty good at reporting their uncertainty. We just need to ask the right way.")
211
  gr.Markdown("Using our uncertainty metric informed by applying causal inference techniques in \
212
- [Exploiting Selection Bias on Underspecified Tasks in Large Language Models](https://arxiv.org/abs/2210.00131 ), \
213
  we are able to identify likely spurious correlations and exploit them in \
214
  the scenario of gender underspecified tasks. (Note that introspecting softmax probabilities alone is insufficient, as in the sentences \
215
  below, LLMs may report a softmax prob of ~0.9 despite the task being underspecified.)")
@@ -218,7 +218,10 @@ with demo:
218
  eight syntactically similar sentences. However semantically, \
219
  only two of the sentences are gender-specified while the rest remain gender-underspecified")
220
  gr.Markdown("If a model can reliably tell us when it is uncertain about its predictions, one can replace only those uncertain predictions with\
221
- an appropriate heuristic.")
 
 
 
222
 
223
  with gr.Row():
224
  model_name = gr.Radio(
 
209
  gr.Markdown(
210
  "#### LLMs are pretty good at reporting their uncertainty. We just need to ask the right way.")
211
  gr.Markdown("Using our uncertainty metric informed by applying causal inference techniques in \
212
+ ['Selection Induced Collider Bias: A Gender Pronoun Uncertainty Case Study'](https://arxiv.org/abs/2210.00131 ), \
213
  we are able to identify likely spurious correlations and exploit them in \
214
  the scenario of gender underspecified tasks. (Note that introspecting softmax probabilities alone is insufficient, as in the sentences \
215
  below, LLMs may report a softmax prob of ~0.9 despite the task being underspecified.)")
 
218
  eight syntactically similar sentences. However semantically, \
219
  only two of the sentences are gender-specified while the rest remain gender-underspecified")
220
  gr.Markdown("If a model can reliably tell us when it is uncertain about its predictions, one can replace only those uncertain predictions with\
221
+ an appropriate heuristic or information retrieval process.")
222
+
223
+ gr.Markdown("#### TL;DR")
224
+ gr.Markdown("Follow steps below to test out one of the pre-loaded options. Once you get the hang of it, you can load a new model and/or provide your own input texts.")
225
 
226
  with gr.Row():
227
  model_name = gr.Radio(