Update Readme.md

README.md

@@ -1,10 +1,8 @@
 # Title
-Analyzing Bitcointalk.org with Large Language Models
-
-- Researchers: Cyrille Grumbach, Didier Sornette
-- Research Assistant: Timothé Laborie
-
+"Bitcoin Forum analysis underscores enduring and substantial carbon footprint of Bitcoin"
 
+- Cyrille Grumbach, ETH Zurich, Switzerland ([email protected])
+- Didier Sornette, Southern University of Science and Technology, China ([email protected])
 
 # Folders
 
@@ -12,34 +10,31 @@ Analyzing Bitcointalk.org with Large Language Models
 
 File main.ipynb is used to scrape the forum.
 
-
 ## hardwarelist
 
 Includes the following:
 
-- pmaxv1 folder: contains the maximum hardware efficiency for each date alongside some manually added updates, originally made by Cyrille.
-- get_hwd_asicminervalue.js and get_hwd_bitcoinwiki.js: Can be pasted into the browser console in the URLs listed within the files, used to extract the hardware efficiency table
+- pmaxv1 folder: contains the maximum hardware efficiency for each date alongside some manually added updates originally made by Cyrille.
+- get_hwd_asicminervalue.js and get_hwd_bitcoinwiki.js: These can be pasted into the browser console in the URLs listed within the files, used to extract the hardware efficiency table
 - hardware_asicminervalue.txt and hardware_bitcoinwiki.txt: The raw output from the above scripts
 - 1_cleanup_hardware_table.ipynb: Used to clean up the raw output, to create hardware_asicminervalue.csv and hardware_bitcoinwiki.csv
 - 2_merge_tables.ipynb: Merges the two tables into hardware_merged.csv
 - 3_paper_list.ipynb: Creates 4 things. 1: The hardware table in the appendix. 2: The pmaxv2.csv file, which uses the hardware_merged.csv file to create an improved table with the maximum hardware efficiency for each date. 3: The pmax evolution table for the paper. 4: The paper_list.csv file, which is used to create an excel sheet later
 - 4_create_pmaxv3.ipynb: Creates the pmaxv3.csv file, which is the max between the pmaxv1.csv and pmaxv2.csv files
 
-
 ## bitcoinforum
 
 ### 1_forum_dataset
 
-Contains the raw HTML from the forum, and code to parse it and combine it into dataframes.
+It contains the raw HTML from the forum and code to parse it and combine it into data frames.
 
 ### 2_train_set_creation
 
-Combines the forum sections into one, truncates long threads, passes a random sample to GPT4 to get the training set for Mistral 7B, also creates the inputs that will be given to Mistral 7B after training.
+Combines the forum sections into one, truncates long threads, passes a random sample to GPT4 to get the training set for Mistral 7B, and also creates the inputs that will be given to Mistral 7B after training.
 
 ### 3_training
 
-Trains Mistral 7B using LoRA, on the dataset generated earlier, and saves the merged model
-
+Trains Mistral 7B using LoRA on the dataset generated earlier and saves the merged model.
 
 ### 4_inference
 
@@ -59,45 +54,41 @@ Includes the following files:
 
 Averages the forum efficiency on a monthly basis, then merges it alongside the Bitcoin price, hashrate, coins per block, and maximum hardware efficiency to create monthly_stuff.csv
 
-monthly_stuff.csv contains columns: date,price,hashrate,coins_per_block,efficiency,max_efficiency
-
+monthly_stuff.csv contains columns: date, price, hashrate, coins_per_block, efficiency, max_efficiency
 
 ## plots
 
 Includes the following:
 
-- carboncomparison folder: Contains the 17 sources used to create the carbon comparison table
+- carbon-comparison folder: Contains the 17 sources used to create the carbon comparison table
 - carbonintensity.html: Cambridge's table for the yearly gCO2e/kWh values, found at https://ccaf.io/cbnsi/cbeci/ghg/methodology
 - appendix2.ipynb: Creates all plots from appendix2
 
-
-
-
 # System requirements
 
-Running the training or inference of Mistral 7B requires an NVIDIA GPU with at least 24GB of VRAM (can also be a Runpod instance).
+Running Mistral 7 B's training or inference requires an NVIDIA GPU with at least 24GB of VRAM (it can also be a Runpod instance).
 
-Everything else can be run on a normal desktop/laptop computer with python 3.10 installed.
+Everything else can be run on a normal desktop/laptop computer with Python 3.10 installed.
 
 # Operating system
 
-Code which is not related to training or inference of Mistral 7B has been tested on Windows 10.
+Code unrelated to training or inference of Mistral 7B has been tested on Windows 10.
 
 Code for Mistral 7B training and inference has been tested on Runpod instances.
 
 # Installation guide for software dependencies
 
-For the code which is not related to training or inference of Mistral 7B, use the packages listed in requirements.txt
+For the code unrelated to training or inference of Mistral 7B, use the packages listed in requirements.txt
 
 ## Installation guide for Mistral 7B training and inference
 
-Setup a Runpod instance with the axolotl docker image, then install unsloth using the instructions at https://github.com/unslothai/unsloth
+Setup a Runpod instance with the axolotl docker image, then install Unsloth using the instructions at https://github.com/unslothai/unsloth
 
-Also install SGLang for inference.
+Also, install SGLang for inference.
 
 ## Typical install time on a "normal" desktop computer
 
-For the code which is not related to training or inference of Mistral 7B, the install time is around 5 minutes.
+For the code unrelated to training or inference of Mistral 7B, the install time is around 5 minutes.
 
 For Mistral 7B training and inference, the install time is around 1 hour.
 
@@ -107,17 +98,17 @@ For Mistral 7B training and inference, the install time is around 1 hour.
 
 Run the code in the order listed in the folders section above.
 
-Note: There are 3 files that normally take a long time to run. I have included a const "DEMO_MODE" at the top of each file. When turned on, the files will run on a tiny subset of the data. The original runtimes are as follows:
+Note: 3 files normally take a long time to run. I have included a const "DEMO_MODE" at the top of each file. When turned on, the files will run on a tiny subset of the data. The original runtimes are as follows:
 
 - The scraper takes over 12 hours to run.
-- The process of creating the training set for Mistral 7B takes around 3 hours and costs about 10$ of OpenAI credits.
-- The process of mapping the hardware names to those of the efficiency table takes around 3 hour and also costs about 10$ of OpenAI credits.
+- The process of creating the training set for Mistral 7B takes around 3 hours and costs about 10$ OpenAI credits.
+- The process of mapping the hardware names to those of the efficiency table takes around 3 hours and also costs about 10$ of OpenAI credits.
 
 All other files can be run in a few minutes.
 
 ## Expected output
 
-You should re-obtain the csv files that are already in the folders, and the plots used in the paper.
+You should re-obtain the CSV files already in the folders and the plots used in the paper.
 
 ## Expected run time for demo on a "normal" desktop computer
 
@@ -125,4 +116,12 @@ The expected run time to run every notebook on a "normal" desktop computer is ar
 
 ## Instructions for use on custom data
 
-The code is designed only to analyse the mining section of bitcointalk.org.
+The code is designed only to analyze the mining section of bitcointalk.org.
+
+**Acknowledgments**
+This work was partially supported by the National Natural Science Foundation of China (Grant No.
+T2350710802 and No. U2039202), Shenzhen Science and Technology Innovation Commission
+Project (Grants No. GJHZ20210705141805017 and No. K23405006), and the Center for
+Computational Science and Engineering at Southern University of Science and Technology. The
+authors acknowledge T. Laborie for excellent research assistance and Y. Cui and M. von Krosigk
+for helpful comments. Any errors are our own.