Mooler0410 · murali1999-tech · Apr 30, 2023 · Apr 30, 2023 · Apr 30, 2023 · Apr 30, 2023
diff --git a/README.md b/README.md
diff --git a/awesome_examples/tableQA.md b/awesome_examples/tableQA.md
@@ -4,6 +4,8 @@ In this lesson, we ask the model to answer questions based on a table. The table
 
 Comparing the following two examples, ChatGPT is vulnerable to table row order perturbation, while GPT4 is robust to table row order perturbation. Such robustness could probably be due to two reasons. The first reason is larger model size and more pretraining data of GPT4. Secondly, better truthfulness stemming from better RLHF alignment could help GPT4 follow different formats of the same instructions better. 
 
+Note that smaller finetuned models heavily suffer from such non-robustness issue, according to the paper: [TableFormer: Robust Transformer Modeling for Table-Text Encoding](https://arxiv.org/pdf/2203.00274.pdf)
+
 # Example 1 (2022/04/29)
 
 ## ChatGPT

diff --git a/imgs/qr_version.jpg b/imgs/qr_version.jpg
diff --git a/imgs/tree.jpg b/imgs/tree.jpg
diff --git a/imgs/tree.png b/imgs/tree.png
diff --git a/source/README.md b/source/README.md
@@ -1,4 +1,5 @@
 ### change log
 
 - V1 (04/07/2023): First version of the figure. 
-- V2 (04/29/2023): Second version of the figure. (The gif version is not updated) 
+- V2 (04/29/2023): Second version of the figure. (The gif version is not updated)
+- V3 (08/06/2023): added Claude 2 and LLama-2-Chat
diff --git a/source/figure_gif.pptx b/source/figure_gif.pptx
diff --git a/source/figure_still.pptx b/source/figure_still.pptx