waterfall 0.1.0__py3-none-any.whl → 0.1.2__py3-none-any.whl

{waterfall-0.1.0.dist-info → waterfall-0.1.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: waterfall
-Version: 0.1.0
+Version: 0.1.2
 Summary: Scalable Framework for Robust Text Watermarking and Provenance for LLMs
 Author-email: Xinyuan Niu <aperture@outlook.sg>
 License-Expression: Apache-2.0
@@ -26,7 +26,7 @@ Gregory Kang Ruey Lau*, Xinyuan Niu*, Hieu Dao, Jiangwei Chen, Chuan-Sheng Foo,
 
 ## TL;DR: Training-free framework for text watermarking that is scalable, robust to LLM attacks, and applicable to original text of multiple types 
 
-![Alt text](Images/Problem_formulation.jpg "")
+![Alt text](https://raw.githubusercontent.com/aoi3142/Waterfall/main/Images/Problem_formulation.jpg "")
 
 1. Watermark original text $T_o$ with watermark key $\mu$ → watermarked text $T_w$ with same semantic content
 
@@ -41,7 +41,7 @@ Protecting intellectual property (IP) of text such as articles and code is incre
 
 # Watermark process
 
-![Alt text](Images/Watermarking_process.png "")
+![Alt text](https://raw.githubusercontent.com/aoi3142/Waterfall/main/Images/Watermarking_process.png "")
 
 1. Original text $T_o$ is fed into LLM paraphraser to produce initial logits $L$.
 
@@ -57,7 +57,7 @@ Protecting intellectual property (IP) of text such as articles and code is incre
 
 # Verification of un/watermarked text
 
-![Alt text](Images/Illustration.gif "Text watermarked with a sine-watermark shows the watermark signal when verified with the correct key")
+![Alt text](https://raw.githubusercontent.com/aoi3142/Waterfall/main/Images/Illustration.gif "Text watermarked with a sine-watermark shows the watermark signal when verified with the correct key")
 
 1. For each token $\hat{w}$ in the watermarked text $T_w$ (original text is not required), use the unique ID $\mu$ and preceding $n-1$ tokens to permute the token index of $\hat{w}$ from $V_o$ space into $V_w$ space.
 
@@ -77,6 +77,13 @@ Protecting intellectual property (IP) of text such as articles and code is incre
 
 # Using our code
 
+Install our package from pypi
+```sh
+pip install waterfall
+```
+
+## Alterntive installation from source
+
 [Optional]
 If using `conda` (or other pkg managers), it is highly advisable to create a new environment
 
@@ -85,7 +92,7 @@ conda create -n waterfall python=3.11 --yes   `# Compatible with python version
 conda activate waterfall
 ```
 
-Clone and install our package
+Clone and install our package from source
 ```sh
 git clone https://github.com/aoi3142/Waterfall.git
 pip install -e Waterfall        `# Install in 'editable' mode with '-e', can be omitted`
@@ -98,6 +105,8 @@ Use the command `waterfall_demo` to watermark a piece of text, and then verify t
 waterfall_demo
 ```
 
+\* Ensure that your device (cuda/cpu/mps) has enough memory to load the model (~16GB for default Llama 3.1 8B model)
+
 Additional arguments
 ```sh
 waterfall_demo \
@@ -107,9 +116,11 @@ waterfall_demo \
   --kappa        2                                `# Watermark strength`          \
   --model        meta-llama/Llama-3.1-8B-Instruct `# Paraphrasing LLM`            \
   --watermark_fn fourier                          `# fourier/square watermark`    \
-  --device       cuda                             `# Use cuda/cpu`
+  --device       cuda                             `# Use cuda/cpu/mps`
 ```
 
+\* By default, `--device` automatically selects among `cuda`/`cpu`/`mps` if not set
+
 ## Using our code to watermark and verify
 
 To watermark texts
@@ -134,7 +145,7 @@ test_texts = ["...", "..."]   # Suspected texts to verify
 watermark_strength = verify_texts(test_texts, id)[0]   # np array of floats
 ```
 
-# Code structure
+## Code structure
 
 - `watermark.py`              : Sample watermarking script used by with `watermark_demo` command, includes beam search and other optimizations
 - `WatermarkerBase.py`        : Underlying generation and verification code provided by `Watermarker` class

{waterfall-0.1.0.dist-info → waterfall-0.1.2.dist-info}/RECORD

@@ -5,9 +5,9 @@ waterfall/WatermarkingFnSquare.py,sha256=2PAO05DdKT02npo7GDf_82D520nP7kGAWK6H4E4
 waterfall/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 waterfall/permute.py,sha256=RwxOHFhx_VSOhhFwy5s79YgwTUBkfW2-LCCXYR3VT2o,2582
 waterfall/watermark.py,sha256=whiNhPwWNNIZwXMH6r7QzEE3A7Niq2Ro9elA1iSRoxI,11952
-waterfall-0.1.0.dist-info/licenses/LICENSE,sha256=zAtaO-k41Q-Q4Etl4bzuh7pgNJsPH-dYfzvznRa0OvM,11341
-waterfall-0.1.0.dist-info/METADATA,sha256=ONpoos0Pyx43h4ntb9Fs4l78F3GZNxq__Ox05GMFD8A,8221
-waterfall-0.1.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-waterfall-0.1.0.dist-info/entry_points.txt,sha256=XXnUzuWXu2nc9j4WAll9tq6HyodN_8WJLjeG0O4Y2Gw,60
-waterfall-0.1.0.dist-info/top_level.txt,sha256=5rTgijeT9V5GRCwIDZmhjeZ4khgH1lmfhS9ZmdUUCKQ,10
-waterfall-0.1.0.dist-info/RECORD,,
+waterfall-0.1.2.dist-info/licenses/LICENSE,sha256=zAtaO-k41Q-Q4Etl4bzuh7pgNJsPH-dYfzvznRa0OvM,11341
+waterfall-0.1.2.dist-info/METADATA,sha256=zDciBMgLwmhhf8voqFzLeCOruNQBdqtohMUYKDMzmdQ,8715
+waterfall-0.1.2.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+waterfall-0.1.2.dist-info/entry_points.txt,sha256=XXnUzuWXu2nc9j4WAll9tq6HyodN_8WJLjeG0O4Y2Gw,60
+waterfall-0.1.2.dist-info/top_level.txt,sha256=5rTgijeT9V5GRCwIDZmhjeZ4khgH1lmfhS9ZmdUUCKQ,10
+waterfall-0.1.2.dist-info/RECORD,,