kawine commited on
Commit
2039e73
·
1 Parent(s): 1b93ac4

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +24 -17
README.md CHANGED
@@ -20,12 +20,13 @@ language:
20
 
21
  ## Summary
22
 
23
- SHP is a dataset of **385K collective human preferences** over responses to questions/instructions in 18 different subject areas, from cooking to legal advice (see the [Design section](https://huggingface.co/datasets/stanfordnlp/SHP#dataset-design) for a breakdown of the domains).
24
- It is primarily intended to be used for training reward models for RLHF and automatic evaluation models for NLG.
25
 
26
- Each example is a Reddit post and a pair of top-level comments for that post, where one comment is more preferred by Reddit users (collectively).
27
  SHP exploits the fact that if comment A was written *after* comment B but has a higher score nonetheless, then A is ostensibly more preferred to B.
28
  If A had been written before B, then we could not conclude this, since its higher score could have been the result of more visibility from being written first.
 
29
 
30
  How is SHP different from [Anthropic's HH-RLHF dataset](https://huggingface.co/datasets/Anthropic/hh-rlhf)?
31
  Most notably, all the data in SHP is naturally occurring and human-written, whereas the responses in HH-RLHF are machine-written, giving us two very different distributions that can complement each other.
@@ -36,7 +37,7 @@ Most notably, all the data in SHP is naturally occurring and human-written, wher
36
  | HH-RLHF | 91K | Dialogue with LLM | Individual Human Preference | not labelled | Live Chat (Multi-turn) | up to 1.5K T5 tokens |
37
 
38
  How is SHP different from other datasets that have scraped Reddit, like [ELI5](https://huggingface.co/datasets/eli5#source-data)?
39
- Most notably, SHP uses the timestamp information to infer preferences, while ELI5 only provides comments and scores -- the latter are not enough to infer preferences since comments made earlier tend to get higher scores from more visibility.
40
  It also contains data from more domains:
41
 
42
  | Dataset | Size | Comments + Scores | Preferences | Number of Domains |
@@ -101,19 +102,16 @@ where the fields are:
101
 
102
  ## Dataset Design
103
 
 
 
104
  The data is sourced from Reddit, which is a public forum organized into topic-specific fora called *subreddits*.
105
  For example, the `askculinary` subreddit is where users ask cooking-related questions and are answered by other users.
106
- The score of a post/comment is 1 plus the number of upvotes it gets from users, minus the number of downvotes it gets.
107
- The value of a score is relative; in subreddits(posts) with more traffic, there will be more higher-scoring posts(comments).
108
- Within a post, comments posted earlier will tend to have a higher score simply due to having more exposure.
109
-
110
-
111
- ### Subreddit Selection
112
 
113
  SHP contains a train, validation, and test split for comments scraped from 18 different subreddits. We chose subreddits based on:
114
  1. whether they were well-known (subscriber count >= 50K)
115
- 2. whether posts were expected to pose a question or instruction that the top-level comments were meant to answer
116
- 3. whether comments had to be rooted in some objectivity, instead of being entirely about personal experiences (e.g., `askscience` vs. `AskAmericans`)
 
117
 
118
  The train/validation/test splits were created by splitting the post IDs of a subreddit in 90%/5%/5% proportions respectively, so that no post would appear in multiple splits.
119
  Since different posts have different numbers of comments, the number of preferences in each split is not exactly 90%/5%/5%:
@@ -140,7 +138,11 @@ Since different posts have different numbers of comments, the number of preferen
140
  | legaladvice | 21170 | 1106 | 1011 | 23287 |
141
  | ALL | 348718 | 18436 | 18409 | 385563 |
142
 
143
- ### Post and Comment Selection
 
 
 
 
144
 
145
  Given a post P and two comments (A,B) we only included the preference A > B in the dataset if
146
  1. A was written *no later than* B and A has a higher score than B.
@@ -203,15 +205,20 @@ We encourage you to use SteamSHP for NLG evaluation, for building reward models
203
 
204
  ## Biases and Limitations
205
 
 
 
206
  Although we filtered out posts with NSFW (over 18) content, chose subreddits that were well-moderated and had policies against harassment and bigotry, some of the data may contain discriminatory or harmful language.
207
  The data does not reflect the views of the dataset creators.
208
- Reddit users on these subreddits are also not representative of the broader population. They are disproportionately from developed, Western, and English-speaking countries.
 
209
  One should keep that in mind before using any models trained on this data.
210
 
211
- It is also worth noting that the comment more preferred by Redditors is not necessarily the more correct one, and though some comments do provide citations to justify their response, most do not.
212
- There are exceptions to this, such as the `askhistorians` subreddit, which is heavily moderated and answers are expected to provide citations.
213
 
214
- As always, remember to evaluate!
 
 
 
215
 
216
 
217
  ## Contact
 
20
 
21
  ## Summary
22
 
23
+ SHP is a dataset of **385K collective human preferences** over responses to questions/instructions in 18 different subject areas, from cooking to legal advice.
24
+ The preferences are meant to reflect the helpfulness of one response over another, and are intended to be used for training RLHF reward models and NLG evaluation models (e.g., [SteamSHP](https://huggingface.co/stanfordnlp/SteamSHP-flan-t5-xl)).
25
 
26
+ Each example is a Reddit post with a question/instruction and a pair of top-level comments for that post, where one comment is more preferred by Reddit users (collectively).
27
  SHP exploits the fact that if comment A was written *after* comment B but has a higher score nonetheless, then A is ostensibly more preferred to B.
28
  If A had been written before B, then we could not conclude this, since its higher score could have been the result of more visibility from being written first.
29
+ We chose data where the preference label is intended to reflect which response is more *helpful* rather than which is less *harmful*, the latter being the focus of much past work.
30
 
31
  How is SHP different from [Anthropic's HH-RLHF dataset](https://huggingface.co/datasets/Anthropic/hh-rlhf)?
32
  Most notably, all the data in SHP is naturally occurring and human-written, whereas the responses in HH-RLHF are machine-written, giving us two very different distributions that can complement each other.
 
37
  | HH-RLHF | 91K | Dialogue with LLM | Individual Human Preference | not labelled | Live Chat (Multi-turn) | up to 1.5K T5 tokens |
38
 
39
  How is SHP different from other datasets that have scraped Reddit, like [ELI5](https://huggingface.co/datasets/eli5#source-data)?
40
+ SHP uses the timestamp information to infer preferences, while ELI5 only provides comments and scores -- the latter are not enough to infer preferences since comments made earlier tend to get higher scores from more visibility.
41
  It also contains data from more domains:
42
 
43
  | Dataset | Size | Comments + Scores | Preferences | Number of Domains |
 
102
 
103
  ## Dataset Design
104
 
105
+ ### Domain Selection
106
+
107
  The data is sourced from Reddit, which is a public forum organized into topic-specific fora called *subreddits*.
108
  For example, the `askculinary` subreddit is where users ask cooking-related questions and are answered by other users.
 
 
 
 
 
 
109
 
110
  SHP contains a train, validation, and test split for comments scraped from 18 different subreddits. We chose subreddits based on:
111
  1. whether they were well-known (subscriber count >= 50K)
112
+ 2. whether posts were expected to pose a question or instruction
113
+ 3. whether responses were valued based on how *helpful* they were
114
+ 4. whether comments had to be rooted in some objectivity, instead of being entirely about personal experiences (e.g., `askscience` vs. `AskAmericans`)
115
 
116
  The train/validation/test splits were created by splitting the post IDs of a subreddit in 90%/5%/5% proportions respectively, so that no post would appear in multiple splits.
117
  Since different posts have different numbers of comments, the number of preferences in each split is not exactly 90%/5%/5%:
 
138
  | legaladvice | 21170 | 1106 | 1011 | 23287 |
139
  | ALL | 348718 | 18436 | 18409 | 385563 |
140
 
141
+ ### Data Selection
142
+
143
+ The score of a post/comment is 1 plus the number of upvotes (approvals) it gets from users, minus the number of downvotes (disapprovals) it gets.
144
+ The value of a score is relative; in subreddits(posts) with more traffic, there will be more higher-scoring posts(comments).
145
+ Within a post, comments posted earlier will tend to have a higher score simply due to having more exposure, which is why using timestamp information is essential when inferring preferences.
146
 
147
  Given a post P and two comments (A,B) we only included the preference A > B in the dataset if
148
  1. A was written *no later than* B and A has a higher score than B.
 
205
 
206
  ## Biases and Limitations
207
 
208
+ ### Biases
209
+
210
  Although we filtered out posts with NSFW (over 18) content, chose subreddits that were well-moderated and had policies against harassment and bigotry, some of the data may contain discriminatory or harmful language.
211
  The data does not reflect the views of the dataset creators.
212
+ Reddit users on these subreddits are also not representative of the broader population.
213
+ Although subreddit-specific demographic information is not available, Reddit users overall are disproportionately male and from developed, Western, and English-speaking countries ([Pew Research](https://www.pewresearch.org/internet/2013/07/03/6-of-online-adults-are-reddit-users/)).
214
  One should keep that in mind before using any models trained on this data.
215
 
216
+ ### Limitations
 
217
 
218
+ The preference label in SHP is intended to reflect how *helpful* one response is relative to another, given an instruction/question.
219
+ However, the more preferred response is not necessarily the more factual one.
220
+ Though some comments do provide citations to justify their response, most do not.
221
+ There are exceptions to this, such as the `askhistorians` subreddit, which is heavily moderated and answers are expected to provide citations.
222
 
223
 
224
  ## Contact