diff --git a/.nojekyll b/.nojekyll index 2e0e5a3..ccd3241 100644 --- a/.nojekyll +++ b/.nojekyll @@ -1 +1 @@ -0de228ad \ No newline at end of file +e27ba027 \ No newline at end of file diff --git a/index.html b/index.html index ac9e7fd..cc1d617 100644 --- a/index.html +++ b/index.html @@ -143,7 +143,7 @@
-
+

@@ -166,7 +166,7 @@
-
+

@@ -189,7 +189,7 @@
-
+
 
@@ -212,7 +212,7 @@
-
+

diff --git a/posts/TDC2023.html b/posts/TDC2023.html index b0d1df4..2ecc5be 100644 --- a/posts/TDC2023.html +++ b/posts/TDC2023.html @@ -352,7 +352,7 @@

\(u_i\) is now a scalar for each x, given a collection of such x’s we can construct a z-score for our dataset as \((u_i - mean(u_i))/std(u_i)\), and rank them.

-

+

The Z-scores of activation vector similarity for the provided sample instances
@@ -722,7 +722,7 @@

Trojan recovery:

});
- + diff --git a/posts/catalog.html b/posts/catalog.html index 35bd0cb..d22f863 100644 --- a/posts/catalog.html +++ b/posts/catalog.html @@ -823,7 +823,7 @@

GitHub

});
- + diff --git a/posts/catalog.out.ipynb b/posts/catalog.out.ipynb index 0823287..b847b2b 100644 --- a/posts/catalog.out.ipynb +++ b/posts/catalog.out.ipynb @@ -297,7 +297,7 @@ "Pythia-12B is miscalibrated on 20% of the bigrams and 45% of the\n", "trigrams when we ask for prediction of $p \\geq 0.45$." ], - "id": "06c9d831-1c49-4989-a26e-4a8e9af18173" + "id": "5dd96a6b-f902-491d-897a-761873a63d92" }, { "cell_type": "code", @@ -313,7 +313,7 @@ } ], "source": [], - "id": "21157758-0f9e-4e85-b726-edfffbffd00b" + "id": "2a4e93cb-ea85-4eaa-840b-8fcc4afc2fee" }, { "cell_type": "markdown", @@ -377,7 +377,7 @@ "The dataset is available on Huggingface:\n", "[pile_scan_4](https://huggingface.co/datasets/Confirm-Labs/pile_scan_4)" ], - "id": "8a452dff-613c-4c99-9047-d6a42dbf494a" + "id": "1ab74550-3521-4dfa-ab3c-032021e2f440" }, { "cell_type": "code", @@ -391,7 +391,7 @@ } ], "source": [], - "id": "554bfd30-557d-45d0-81ff-1011fd3506b1" + "id": "330079d0-f60c-49e5-82e0-766e0ba7a91f" }, { "cell_type": "markdown", @@ -423,7 +423,7 @@ "Computational Linguistics, May 2022, pp. 95–136. doi:\n", "[10.18653/v1/2022.bigscience-1.9](https://doi.org/10.18653/v1/2022.bigscience-1.9)." ], - "id": "c8f430e1-ffd8-4dfe-8a28-b842a4f61b98" + "id": "875e30e7-98aa-49bd-a1c3-b23e310cc97d" } ], "nbformat": 4, diff --git a/posts/dreamy.html b/posts/dreamy.html index 73c5333..8413ab6 100644 --- a/posts/dreamy.html +++ b/posts/dreamy.html @@ -247,6 +247,7 @@

Fluent dreaming for language models

This is a companion page for our paper, “Fluent dreaming for language models.” (arXiv link).

+

There is an interactive demo of this page on Colab.

@@ -5211,7 +5212,7 @@

Causal token attr });

- + diff --git a/posts/dreamy.out.ipynb b/posts/dreamy.out.ipynb index 87fd7dc..15319dc 100644 --- a/posts/dreamy.out.ipynb +++ b/posts/dreamy.out.ipynb @@ -14,6 +14,9 @@ "This is a companion page for our paper, [“Fluent dreaming for language\n", "models.” (arXiv link)](http:///arxiv).\n", "\n", + "There is an [interactive demo of this page on\n", + "Colab.](https://colab.research.google.com/drive/1B0dM7du91BUkT7tSICXjKL7lrBAEdSa-?usp=sharing)\n", + "\n", "> **Dreaming Phi-2**\n", ">\n", ">
" ], - "id": "cb0cb1d8-1eee-4107-a764-b134ff396c70" + "id": "cf6c3a1e-a472-4426-9060-e22e51e9e9dd" } ], "nbformat": 4, diff --git a/search.json b/search.json index 917a400..55cfc61 100644 --- a/search.json +++ b/search.json @@ -39,7 +39,7 @@ "href": "posts/dreamy.html", "title": "Fluent dreaming for language models", "section": "", - "text": "This is a companion page for our paper, “Fluent dreaming for language models.” (arXiv link).\nDreaming is the process of maximizing some internal or output feature of a neural network by iteratively tweaking the input to the network. The most well-known example is DeepDream [1]. Besides making pretty images, dreaming is useful for interpreting the purpose of the internal components of a neural network [2]–[4]. To our knowledge, Dreaming has previously only been applied to vision models because the input space to a vision model is approximately continuous and algorithms like gradient descent work well. For language models, the input space is discrete and very different algorithms are needed. Extending work in the adversarial attacks literature [5], in the paper, we introduce the Evolutionary Prompt Optimization (EPO) algorithm for dreaming with language models.\nOn this page, we demonstrate running the EPO algorithm for a neuron in Phi-2. There is also a Colab notebook version of this page available." + "text": "This is a companion page for our paper, “Fluent dreaming for language models.” (arXiv link).\nThere is an interactive demo of this page on Colab.\nDreaming is the process of maximizing some internal or output feature of a neural network by iteratively tweaking the input to the network. The most well-known example is DeepDream [1]. Besides making pretty images, dreaming is useful for interpreting the purpose of the internal components of a neural network [2]–[4]. To our knowledge, Dreaming has previously only been applied to vision models because the input space to a vision model is approximately continuous and algorithms like gradient descent work well. For language models, the input space is discrete and very different algorithms are needed. Extending work in the adversarial attacks literature [5], in the paper, we introduce the Evolutionary Prompt Optimization (EPO) algorithm for dreaming with language models.\nOn this page, we demonstrate running the EPO algorithm for a neuron in Phi-2. There is also a Colab notebook version of this page available." }, { "objectID": "posts/dreamy.html#installation-and-setup", diff --git a/sitemap.xml b/sitemap.xml index fb41c8b..1982143 100644 --- a/sitemap.xml +++ b/sitemap.xml @@ -2,22 +2,22 @@ https://confirmlabs.org/posts/catalog.html - 2024-01-24T16:53:58.704Z + 2024-01-24T18:26:08.485Z https://confirmlabs.org/posts/dreamy.html - 2024-01-24T16:53:55.428Z + 2024-01-24T18:26:05.281Z https://confirmlabs.org/index.html - 2024-01-24T16:53:52.504Z + 2024-01-24T18:26:02.353Z https://confirmlabs.org/posts/TDC2023.html - 2024-01-24T16:53:53.968Z + 2024-01-24T18:26:03.825Z https://confirmlabs.org/posts/fight_the_illusion.html - 2024-01-24T16:53:56.140Z + 2024-01-24T18:26:05.997Z