|
|
--- |
|
|
license: mit |
|
|
task_categories: |
|
|
- text-classification |
|
|
language: |
|
|
- en |
|
|
tags: |
|
|
- llm training |
|
|
- chemistry |
|
|
- biology |
|
|
- code |
|
|
- finance |
|
|
- math |
|
|
pretty_name: 'Add Human Feedback to Your AI ' |
|
|
size_categories: |
|
|
- 10K<n<100K |
|
|
--- |
|
|
General Overview |
|
|
|
|
|
This dataset is to be used for LLM Trainings, This is a sample, visit https://www.cntxt.tech/ to learn more |
|
|
|
|
|
The dataset consists of 50 rows (excluding headers) and 8 columns. The columns capture various aspects of ranked responses to prompts, including: |
|
|
Numeric_ID (Unique Identifier - Integer) |
|
|
Prompt (The Question or Task - Text) |
|
|
Answer_A / Answer_B (Response Options - Text) |
|
|
Category (Type of Task - Categorical) |
|
|
Best Answer (Preferred Response - Categorical) |
|
|
LikeRT Score (Ranking Score - Numeric) |
|
|
Comments (Reviewer Comments) |
|
|
Statistical Insights |
|
|
Category Distribution |
|
|
|
|
|
|
|
|
General Knowledge (30%) |
|
|
Instruction Following (26%) |
|
|
Safety (24%) |
|
|
Reasoning (20%) |
|
|
|
|
|
Best Answer Trends |
|
|
|
|
|
|
|
|
"A" was chosen as the best answer 25 times (50%). |
|
|
"B" was chosen 24 times (48%). |
|
|
One case had both "A/B" as the correct response. |
|
|
|
|
|
LikeRT Score Analysis |
|
|
|
|
|
|
|
|
‘Mode’ Score: 2 |
|
|
Minimum Score: 2 |
|
|
Maximum Score: 6 |
|
|
The majority of scores fall within the 2-6 range, indicating a fair rating distribution. |
|
|
|
|
|
|
|
|
|
|
|
Data Trends & Patterns |
|
|
Distribution of LikeRT Scores |
|
|
|
|
|
Most scores hover around 2, 3, or 6, forming three distinct peaks. |
|
|
The dataset does not exhibit extreme outliers beyond the expected range. |
|
|
|
|
|
Category-Based Performance Trends |
|
|
|
|
|
"General Knowledge" and "Safety" questions tend to have higher LikeRT scores, indicating clearer or more objective evaluation criteria. |
|
|
"Instruction-Following" has more variability, due to subjective interpretations. |
|
|
|
|
|
Best Answer vs. LikeRT Score Correlation |
|
|
|
|
|
No significant bias between "A" or "B" in terms of overall LikeRT scores. |
|
|
The choice of "Best Answer" is distributed fairly evenly between the two responses. |
|
|
|
|
|
|
|
|
Approach to Data Collection & Review |
|
|
The dataset was compiled through a structured annotation process, likely involving the following steps: |
|
|
Prompt Selection: Diverse questions were designed across multiple categories. |
|
|
Response Generation: Two alternative answers (A & B) were provided for each prompt. |
|
|
Validation Process: Annotators selected the preferred response and assigned a LikeRT Score. |
|
|
Quality Assurance: Reviewers then reviewed the data to ensure it was error-free. |
|
|
Final Summary & Conclusion |
|
|
The dataset is well-structured and provides a clear ranking system. |
|
|
There is balance in the selection of Best Answers, indicating an unbiased reviewing process. |
|
|
LikeRT Scores are evenly distributed, without major outliers or anomalies. |
