dataset_info:
features:
- name: prompt
dtype: string
- name: prompt_id
dtype: string
- name: chosen
list:
- name: content
dtype: string
- name: role
dtype: string
- name: rejected
list:
- name: content
dtype: string
- name: role
dtype: string
- name: messages
list:
- name: content
dtype: string
- name: role
dtype: string
- name: score_chosen
dtype: float64
- name: score_rejected
dtype: float64
- name: other_info
struct:
- name: domain
dtype: string
- name: post_id
dtype: string
- name: raw_score_chosen
dtype: int64
- name: raw_score_ratio
dtype: float64
- name: raw_score_rejected
dtype: int64
- name: seconds_difference
dtype: float64
- name: source
dtype: string
- name: upvote_ratio
dtype: float64
splits:
- name: train
num_bytes: 1815446429
num_examples: 348718
- name: validation
num_bytes: 93098840
num_examples: 18436
- name: test
num_bytes: 95879141
num_examples: 18409
download_size: 262070837
dataset_size: 2004424410
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
Dataset Card for Dataset Name
Reformatted from stanfordnlp/SHP dataset. To make it consistent with other preference dsets, we:
- convert upvotes to scores in a [1, 10] scale. This is achieved by 1) convert the better response's upvotes to score of [5.0, 10.0] by:
def shp_map_score(score, threshold=78): # 78 is chosen because about the best 10% data has score > 78 if score > threshold: return 10.0 # linearly map the rest # start with 5.0 because we assume that any human written reponses that can receive any upvote should already reflect decent quality return 5.0 + (score / 78) * 5.0score_ratioin the original dataset, we use it to model score difference between the chosen and the rejected score. Therefore, the rejected score is calculated by:remaped_chosen_score = # from above ratio_diff = data_row['score_ratio'] - 1.0 rejected_score = max(remaped_chosen_score - ratio_diff, 0.0) - all other information is kept intact: since the original data is already paired, we simply reformat to use the better response as
chosen, and the other asrejected.
convert all scores to a [1, 10] scale by np.mean([helpfulness+1, correctness+1, coherence+1, complexity+1, 4-verbosity])*2.0 the original dset considers 4 responses per prompt. We construct preference pairs by 1) take the best scoring response as chosen, and 2) randomly sample responses with score lower than best response as rejected. We skip prompts/data rows where all responses have the same score.
Dataset Details
Dataset Description
- Curated by: [More Information Needed]
- Funded by [optional]: [More Information Needed]
- Shared by [optional]: [More Information Needed]
- Language(s) (NLP): [More Information Needed]
- License: [More Information Needed]
Dataset Sources [optional]
- Repository: [More Information Needed]
- Paper [optional]: [More Information Needed]
- Demo [optional]: [More Information Needed]
Uses
Direct Use
[More Information Needed]
Out-of-Scope Use
[More Information Needed]
Dataset Structure
[More Information Needed]
Dataset Creation
Curation Rationale
[More Information Needed]
Source Data
Data Collection and Processing
[More Information Needed]
Who are the source data producers?
[More Information Needed]
Annotations [optional]
Annotation process
[More Information Needed]
Who are the annotators?
[More Information Needed]
Personal and Sensitive Information
[More Information Needed]
Bias, Risks, and Limitations
[More Information Needed]
Recommendations
Users should be made aware of the risks, biases and limitations of the dataset. More information needed for further recommendations.
Citation [optional]
BibTeX:
[More Information Needed]
APA:
[More Information Needed]
Glossary [optional]
[More Information Needed]
More Information [optional]
[More Information Needed]
Dataset Card Authors [optional]
[More Information Needed]
Dataset Card Contact
[More Information Needed]