Update README.md

README.md

@@ -9,7 +9,7 @@ size_categories:
 ---
 # Dataset Summary
 
-The "humor-chains" dataset is a machine-filtered collection of the most upvoted Reddit submissions and their replies of humor-related subreddits. Generally, a humor chain is when a short post triggers a chain of one or more replies *that Redditors find entertaining*. In other words, some entries might be NSFW, topical, or internal jokes.
+The "humor-chains" dataset is a machine-filtered collection of the most upvoted Reddit submissions and their replies on humor-related subreddits. Generally, a humor chain is when a short post triggers a chain of one or more replies *that Redditors find entertaining*. In other words, some entries might be NSFW, topical, or internal jokes.
 
 For example ([original thread](https://www.reddit.com/r/funny/comments/cmb3z/im_dating_a_midget/?sort=top)):
 ![Screenshot-reddit-chain.png](https://cdn-uploads.huggingface.co/production/uploads/6564a7606c9793d92e715f33/YKtjZD6V1auH1DVaofQo2.png)
@@ -25,11 +25,13 @@ Reddit's license states in section "2.4 User Content" that such content is [owne
 For privacy reasons, all usernames were removed and replaced by "human" (original submission) and "gpt" (all replies). 
 That format is convenient for fine-tuning GPT models, e.g., in [Axolotl](https://github.com/OpenAccess-AI-Collective/axolotl) that is ["sharegpt" format type](https://openaccess-ai-collective.github.io/axolotl/docs/dataset-formats/conversation.html#sharegpt).
 
+UPDATE 2024-7-9: Axolotl's finetunimg formats excpect single GPT reply, so dataset was reduced to include submission and a single top reply.
+
 ## Dataset Creation
 
 The dataset was created using the following procedure:
 1. Reddit's API was not used directly, as API license switched to a more restrictive license (see [2023 Reddit API controversy](https://en.wikipedia.org/wiki/2023_Reddit_API_controversy)). Instead, I used older Reddit dump [Subreddit comments/submissions 2005-06 to 2023-12](https://academictorrents.com/details/56aa49f9653ba545f48df2e33679f014d2829c10). Inside the big torrent file are zstandard compressed ndjson files, each representing a subreddit.
-2. I downloaded a portion of subreddits from [List of funny subreddits](https://www.reddit.com/r/redditlists/comments/128ayc/list_of_funny_subreddits/).
+2. Downloaded a portion of subreddits from [List of funny subreddits](https://www.reddit.com/r/redditlists/comments/128ayc/list_of_funny_subreddits/).
 3. Python script filtered .jsonl files inside .zst to include only the most upvoted thread for each submission.
 4. Filtered submissions satisfy ALL of the following:
    * score >= min (10 upvotes),