Update README.md

README.md

@@ -33,7 +33,20 @@ How is SHP different from [Anthropic's HH-RLHF dataset](https://huggingface.co/d
 
 ## Data Structure
 
-Here's an example from the `askculinary` training data:
+There are 18 directories, one for each subreddit, and each directory contains a JSONL file for the training, validation, and test data. 
+Here's how to get the data using Huggingface's `datasets` library:
+
+```python
+from datasets import load_dataset
+
+# Load all the data (share the same schema)
+dataset = load_dataset("stanfordnlp/shp")
+
+# Load one of the harmless subsets
+dataset = load_dataset("stanfordnlp/shp", data_dir="askculinary")
+```
+
+Here's an example from `askculinary`/train.json:
 ```
 {
     `post_id`:"qt3nxl",
@@ -68,7 +81,7 @@ where the fields are:
 - ```human_ref_A```: text of comment A (string)
 - ```human_ref_B```: text of comment B (string)
 - ```labels```: the preference label -- it is 1 if A is preferred to B; 0 if B is preferred to A. This was randomized such that the label distribution is roughly 50/50. (integer)
-- ```seconds_difference```: how many seconds after the less preferred comment the more preferred one was created (will always be positive) (integer)
+- ```seconds_difference```: how many seconds after the less preferred comment the more preferred one was created (will always be >= 0) (integer)
 - ```score_ratio```: the ratio score_A:score B (will be >= 2) (float)
 
 
@@ -111,36 +124,48 @@ Since different posts have different numbers of comments, the number of preferen
 | legaladvice        | 21170 | 1106 | 1011 | 23287 | 
 | ALL | 348718 | 18436 | 18409 | 385563 |
 
+### Post and Comment Selection
 
-### 
-
-of 
-The input in SHP contains more [FLANT5-usable information](https://icml.cc/virtual/2022/oral/16634) about the preference label than in 
-
-Specifically, given a post P and two comments (A,B) we only included the preference A > B in the dataset if 
-1. A was written *no later than* B.
-2. Despite being written later, A has a score that is at least 2 times as high as B's.
-3. Both comments have a score >= 2 and the post has a score >= 10.
-4. The post is a self-post (i.e., a body of text and not a link to another page) made before 2023, was not edited, and is not NSFW (over 18).
-5. Neither comment was made by a deleted user, a moderator, or the post creator. The post was not made by a deleted user or moderator.
-
-Since comments made earlier get more visibility, the first condition is needed to ensure that A's higher score is not the result of a first-mover advantage.
-Since the comment score is also a noisy estimate of the comment's utility, the second and third conditions were enforced to ensure that the preference is genuine.
-
+Given a post P and two comments (A,B) we only included the preference A > B in the dataset if 
+1. A was written *no later than* B and A has a higher score than B.
+2. The post is a self-post (i.e., a body of text and not a link to another page) made before 2023, was not edited, and is not NSFW (over 18).
+3. Neither comment was made by a deleted user, a moderator, or the post creator. The post was not made by a deleted user or moderator.
+4. The post P has a score >= 10 and each comment has a score >= 2 (upvoted at least once).
 
+Reddit makes it very difficult to get anything beyond the top 1000 posts for subreddit.
+We started with the top-scoring 1000 posts (of all time) and searched for the 25 most similar posts to each one using the Reddit search function.
+By doing this recursively, we scraped up to 7500 post IDs for each subreddit and then used the AsyncPRAW API to scrape the top 50 comments from each post.
+We limited the scraping to 50 comments per post because the number of comments per post is Pareto-distributed, and we did not want a relatively small number of posts dominating the data.
 
-## Files
+### Preprocessing
 
+We tried to keep preprocessing to a minimum. Subreddit-specific abbreviations were expanded ("CMV" to "Change my view that"). 
+In hyperlinks, only the referring text was kept and the URL was removed (if the URL was written out, then it was kept).
 
 
+## Building a Preference Model
 
+### Finetuning
 
+If you want to finetune a model to predict human preferences (e.g., for NLG evaluation or an RLHF reward model), here are some helpful tips:
 
+1. **Use a sufficiently large model.** With FLAN-T5-xl, you can get 65-85% percent accuracies depending on the subreddit.
+2. **Do in-domain prediction.** Out-of-domain performance will be poor if the subreddits are unrelated (e.g., if you fine-tune on `askculinary` preferences and test on `askcarguys` preferences).
+3. **Preprocess the data**. The total input length should fit under the model's token limit (usually 512 tokens).
+   Although models like FLAN-T5 use positional embeddings, we found that the loss would not converge if we finetuned it on the entire input.
+   To avoid this, truncate the post text (in the `history` field) as much as possible, such that the whole input is under 512 tokens (do not truncate the comment(s) however).
+   If this is still over 512 tokens, simply skip the example.
+5. **Train for 1 epoch only**, as the [InstructGPT paper](https://arxiv.org/abs/2203.02155) paper suggests.
+   Since the same comment appears in multiple preferences, it is easy to overfit to the data.
+6. **Train on less data.**
+   Preferences with a large score ratio (e.g., comment A having 2x the score of comment B) will provide a stronger signal for finetuning the model, so you may only want to consider preferences above a certain `score_ratio`.
+   The number of preferences per post is Pareto-distributed, so to prevent the model from over-fitting to certain posts, you may want to limit the number of preferences from a particular post.
+   
 
 
 ## Disclaimer
 
-Although we filtered out posts with NSFW (over 18) content, some of the data may contain discriminatory or harmful language.
+Although we filtered out posts with NSFW (over 18) content and chose an innocuous set of subreddits, some of the data may contain discriminatory or harmful language.
 The data does not reflect the views of the dataset creators.
 Please only engage with the data in accordance with your own personal risk tolerance. 
 
@@ -148,34 +173,7 @@ Reddit users on these subreddits are also not necessarily representative of the
 As always, remember to evaluate!
 
 
-## FAQs
-
-**Q**: *I'm trying to train a FLAN-T5/T5 model on these preferences, but the loss won't converge. Help!*
-
-**A**: The most likely problem is that you're feeding the post text AND one or both comments as input, which is a lot larger than the 512 tokens these models can support.
-Even though they use relative position embeddings, in our experience, this is not helpful when training a preference/reward model on this data.
-To avoid this, truncate the post text as much as possible, such that the whole input is under 512 tokens (do not truncate the comment(s) however). If this is still over 512 tokens, simply skip the example.
-This should allow you to still train on most of the examples and get a preference model that is still ~75% accurate at predicting human preferencess.
-We are currently training a preference model on this data and will make it available shortly.
-
-**Q**: *Why did you use threshold the score ratio rather than the score difference when filtering preferences?*
-
-**A**: Some Reddit posts get far less traffic than others, which means their comments have lower absolute scores. 
-An absolute difference threshold would disproportionately exclude comments from these posts, a kind of bias that we didn't want to introduce.
-
-**Q**: *Did you scrape every post on those 18 subreddits?*
-
-**A**: No. Reddit makes it very difficult to get anything beyond the top 1000 posts. 
-We started with the top-scoring 1000 posts (of all time) and searched for the 25 most similar posts to each one using the Reddit search function.
-By doing this recursively, we scraped up to 7500 post IDs for each subreddit and then used the AsyncPRAW API to scrape the top 50 comments from each post.
-We limited the scraping to 50 comments per post because the number of comments per post is Pareto-distributed, and we did not want a relatively small number of posts dominating the data.
-
-**Q**: *How did you preprocess the text?*
-
-**A**: We tried to keep preprocessing to a minimum. Subreddit-specific abbreviations were expanded ("CMV" to "Change my view that"). 
-In hyperlinks, only the referring text was kept and the URL was removed (if the URL was written out, then it was kept).
-
-
 ## Contact
 
-Please contact kawin@stanford.edu if you have any questions about the data.
+Please contact kawin@stanford.edu if you have any questions about the data.
+This project is being maintained by Kawin Ethayarajh, Heidi (Chenyu) Zhang, and Yizhong Wang.