Add toxicity evaluation on prompted generation

neulab · Nov 13, 2024 · 6fb8f9e · 6fb8f9e
1 parent ad11f80
commit 6fb8f9e
Show file tree

Hide file tree

Showing 12 changed files with 1,174 additions and 99 deletions.
diff --git a/examples/detoxification_bias/Detoxify_LM_demo.ipynb b/examples/detoxification_bias/Detoxify_LM_demo.ipynb
diff --git a/examples/detoxification_bias/Llama2-7b_perplexity_scores_plot.png b/examples/detoxification_bias/Llama2-7b_perplexity_scores_plot.png
diff --git a/examples/detoxification_bias/gpt2_perplexity_scores_plot.png b/examples/detoxification_bias/gpt2_perplexity_scores_plot.png
diff --git a/examples/detoxification_bias/image/perplexity_gpt2.png b/examples/detoxification_bias/image/perplexity_gpt2.png
diff --git a/examples/detoxification_bias/image/perplexity_llama2.png b/examples/detoxification_bias/image/perplexity_llama2.png
diff --git a/examples/detoxification_bias/image/toxicity_gpt2.png b/examples/detoxification_bias/image/toxicity_gpt2.png
diff --git a/examples/detoxification_bias/image/toxicity_llama2.png b/examples/detoxification_bias/image/toxicity_llama2.png
diff --git a/examples/detoxification_bias/perplexity/gpt2_DAPT.txt b/examples/detoxification_bias/perplexity/gpt2_DAPT.txt
@@ -0,0 +1,4 @@
+nontoxic_aae.txt: 548.1239176862463
+nontoxic_wae.txt: 147.63300804273632
+toxic_aae.txt: 1029.3297598141162
+toxic_wae.txt: 219.54834864455597
diff --git a/examples/detoxification_bias/perplexity/gpt2_vanilla.txt b/examples/detoxification_bias/perplexity/gpt2_vanilla.txt
@@ -0,0 +1,4 @@
+nontoxic_aae.txt: 246.03108213983714
+nontoxic_wae.txt: 82.38113356642384
+toxic_aae.txt: 331.3004681927659
+toxic_wae.txt: 90.21565652496795
diff --git a/examples/detoxification_bias/perplexity/llama2_DAPT.txt b/examples/detoxification_bias/perplexity/llama2_DAPT.txt
@@ -0,0 +1,4 @@
+nontoxic_aae.txt: 130.09595642687992
+nontoxic_wae.txt: 64.82364552393581
+toxic_aae.txt: 150.9991039804579
+toxic_wae.txt: 72.54242631595599
diff --git a/examples/detoxification_bias/perplexity/llama2_vanilla.txt b/examples/detoxification_bias/perplexity/llama2_vanilla.txt
@@ -0,0 +1,4 @@
+nontoxic_aae.txt: 108.29028081355527
+nontoxic_wae.txt: 54.45030579864216
+toxic_aae.txt: 119.11823342766786
+toxic_wae.txt: 55.3702726134379
diff --git a/examples/detoxification_bias/requirements.txt b/examples/detoxification_bias/requirements.txt
@@ -12,4 +12,4 @@ tokenizers
 torch
 torchtext
 transformers
-tweet-preprocessor
+tqdm