# Education Dialogue Dataset

Education Dialogue (ED) contains conversations, generated by prompting Gemini Ultra.
These are conversations between a teacher and a student, where the teacher is prompted with a specific topic to teach the student, and the student is prompted with their learning preferences.

For more details on the design and content of the dataset, please see the paper [Multi-turn Reinforcement Learning from Preference Human Feedback](https://arxiv.org/abs/2405.14655).

## Data Description

ED contains 40,000 training examples and 7,234 examples for testing.
Each example is a full conversation between a teacher and a student, including metadata on the topic and the teacher/student preferences.
In the paper, we perform experiments in multi-turn reinforcement learning on the dataset.

## Data Format

The data is composed of 6 JSON files: 5 for the training data and 1 for the test data. Each entry represents a conversation with the following fields:
* background_info - the contextual information of the conversation, containing the following fields:
  * topic -  the topic that the teacher needs to teach.
  * student_prefrences - the ways in which the student prefers to learn, e.g., lecture-based learning or hands-on activities.
  * teacher_prefrences - the ways in which the teacher prefers to teach, e.g., lecture-based learning or hands-on activities.
  * student_reactions - the reaction of the student if they don’t learn in their preferred way, e.g., gets disengaged or might adapt to other methods.
  * teacher_reactions - the reaction of the teacher if the student doesn't learn in their preferred teaching way, e.g., gets frustrated or might adapt to the student.
* conversation - a list of turns that represent the conversation between the teacher and the student. Each turn has a field (called role) that identifies the speaker and a field (called text) of the content.

The data was generated by prompting Gemini Ultra with the following prompt:

*Simulate a conversation between a teacher in school and a student.
There is a small chance that the teacher is successful in teaching the student so he understands the topic.
The conversation lasts roughly 10-15 turns but ends when either side says \[end of conversation\].
The teacher wants to teach the student about {topic}.  The student likes {student_pref}.  The teacher does not know that beforehand. The student prefers to learn this way, {student_reaction}. The teacher likes {teacher_pref}.  He prefers to teach thisway, {teacher_reaction}.  Output the conversation and the probability that the student understood the material, in the following format.\
\#\
Conversation:\
\[\
   Teacher:  "...",\
   Student:  "...",\
   Teacher:  "...",\
   Student:  "...",\
\]\
Probability:  "...",\
\#*