Update README.md
Browse files
README.md
CHANGED
@@ -3,10 +3,12 @@ license: mit
|
|
3 |
task_categories:
|
4 |
- text-generation
|
5 |
tags:
|
6 |
-
- human
|
7 |
- rlhf
|
8 |
- preferences
|
9 |
- reddit
|
|
|
|
|
10 |
size_categories:
|
11 |
- 100K<n<1M
|
12 |
language:
|
@@ -20,7 +22,7 @@ SHP is a dataset of **385K aggregate human preferences** over responses to quest
|
|
20 |
It is primarily intended to be used for training reward models for RLHF and automatic evaluation models for NLG.
|
21 |
|
22 |
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 (in aggregate).
|
23 |
-
SHP exploits the fact that if comment A was written *after* comment B but has a higher score nonetheless, then A is
|
24 |
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.
|
25 |
|
26 |
How is SHP different from [Anthropic's HH-RLHF dataset](https://huggingface.co/datasets/Anthropic/hh-rlhf)?
|
@@ -32,7 +34,8 @@ Most notably, all the data in SHP is human-written, whereas the responses in HH-
|
|
32 |
| HH-RLHF | 91K | Dialogue with LLM | Individual Human Preference Label | unclear (not labelled) | Multi-turn Dialogue | up to 1.5K T5 tokens |
|
33 |
|
34 |
How is SHP different from other datasets that have scraped reddit, like [ELI5](https://huggingface.co/datasets/eli5#source-data)?
|
35 |
-
Most notably, SHP exploits the timestamp hack to infer preferences, rather than only providing the comments
|
|
|
36 |
|
37 |
| Dataset | Size | Comments + Scores | Preferences | Number of Domains |
|
38 |
| -------------------- | ---- | ------------------ | -------------| ------------------ |
|
|
|
3 |
task_categories:
|
4 |
- text-generation
|
5 |
tags:
|
6 |
+
- human feedback
|
7 |
- rlhf
|
8 |
- preferences
|
9 |
- reddit
|
10 |
+
- preference model
|
11 |
+
- RL
|
12 |
size_categories:
|
13 |
- 100K<n<1M
|
14 |
language:
|
|
|
22 |
It is primarily intended to be used for training reward models for RLHF and automatic evaluation models for NLG.
|
23 |
|
24 |
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 (in aggregate).
|
25 |
+
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.
|
26 |
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.
|
27 |
|
28 |
How is SHP different from [Anthropic's HH-RLHF dataset](https://huggingface.co/datasets/Anthropic/hh-rlhf)?
|
|
|
34 |
| HH-RLHF | 91K | Dialogue with LLM | Individual Human Preference Label | unclear (not labelled) | Multi-turn Dialogue | up to 1.5K T5 tokens |
|
35 |
|
36 |
How is SHP different from other datasets that have scraped reddit, like [ELI5](https://huggingface.co/datasets/eli5#source-data)?
|
37 |
+
Most notably, SHP exploits the timestamp hack to infer preferences, rather than only providing the comments.
|
38 |
+
It also contains data from far more domains:
|
39 |
|
40 |
| Dataset | Size | Comments + Scores | Preferences | Number of Domains |
|
41 |
| -------------------- | ---- | ------------------ | -------------| ------------------ |
|