Prompt Order Experiment
Prompt Order Experiment
This experiment aims to explore various scenarios for prompt fine-tuning using structured generation. We'll test how the order of elements in a prompt affects model performance. The elements we consider are:
- (Q): Question
- (AC): Answer Choices
- (R): Reasoning
- (FA): Final Answer
Scenarios
We will evaluate the following prompt orders:
Scenario 1: Q - AC - R - FA (Mistral and GPT3.5)
This is the most natural order. The model generates reasoning before the final answer, providing the most information prior to making a selection. This order leverages decoding mechanics effectively.
This is our user message, we can see the question and answer choices
Click to show prompt!
<seq>[INST] Answer the Question and include your Reasoning and the Final Answer in a json like: {"Reasoning: "...", "Final Answer": "x"} where x is a letter that corresponds to the answer choice which is a letter between a and h. Question: What is satellite technology used for predicting? Answer Choices: (a) Seconds and minutes (b) The strength and magnitude of an earthquake (c) What it's like outside each day (d) 70-75 degrees fahrenheit (e) Rapid changes occur (f) Dead-ends and false starts. (g) Snow, ice, and rock (h) Around 5 to 27 degrees celsius[/INST]
This is our assistant message, you can see that we are forcing a JSON (note I added spacing for visual purposes), and we are putting the reasoning first. Using a JSON in fine-tuning will improve our structured generation results as the model will get used to responding in that "space".
{ "Reasoning": "a) Seconds and minutes: This option is incorrect because satellite technology is not used for predicting time intervals. Satellite technology is used for various purposes such as communication, navigation, and weather forecasting, but it is not used for predicting time intervals. b) The strength and magnitude of an earthquake: This option is incorrect because satellite technology is not used for predicting earthquakes. Earthquake prediction is a complex process that involves seismology and other scientific methods, but satellite technology is not one of them. d) 70-75 degrees Fahrenheit: This option is incorrect because satellite technology is not used for predicting specific temperature ranges. While satellite technology can provide temperature data, it is not used for predicting specific temperature ranges. e) Rapid changes occur: This option is too vague and does not provide enough information to determine whether it is correct or not. Satellite technology can be used to monitor changes in various environmental factors, but it is not used specifically for predicting rapid changes. f) Dead-ends and false starts: This option is incorrect because it is not related to satellite technology or any type of prediction. g) Snow, ice, and rock: This option is incorrect because it is too specific and does not cover the broad range of predictions that satellite technology can be used for. While satellite technology can be used to monitor snow, ice, and rock formations, it is not used exclusively for this purpose. h) Around 5 to 27 degrees Celsius: This option is incorrect because it is too specific and does not cover the broad range of temperature predictions that satellite technology can be used for. While satellite technology can provide temperature data, it is not used exclusively for predicting temperatures within a specific range. Therefore, the correct answer is c) what it's like outside each day, as satellite technology is commonly used for weather forecasting and predicting daily weather conditions.", "Final Answer": "c" }</seq>
Scenario 2: Q - AC - FA - R (Mistral and GPT3.5)
An awkward order, placing reasoning after the final answer. While it is faster, it assumes the model can "know" reasoning internally before generating it. This approach saves tokens but is a skeptical case worth testing.
Click to show prompt!
```text{ 'Final Answer': 'c', 'Reasoning': "a) Seconds and minutes: This option is incorrect because satellite technology is not used for predicting time intervals. Satellite technology is used for various purposes such as communication, navigation, and weather forecasting, but it is not used for predicting time intervals. b) The strength and magnitude of an earthquake: This option is incorrect because satellite technology is not used for predicting earthquakes. Earthquake prediction is a complex process that involves seismology and other scientific methods, but satellite technology is not one of them. d) 70-75 degrees Fahrenheit: This option is incorrect because satellite technology is not used for predicting specific temperature ranges. While satellite technology can provide temperature data, it is not used for predicting specific temperature ranges. e) Rapid changes occur: This option is too vague and does not provide enough information to determine whether it is correct or not. Satellite technology can be used to monitor changes in various environmental factors, but it is not used specifically for predicting rapid changes. f) Dead-ends and false starts: This option is incorrect because it is not related to satellite technology or any type of prediction. g) Snow, ice, and rock: This option is incorrect because it is too specific and does not cover the broad range of predictions that satellite technology can be used for. While satellite technology can be used to monitor snow, ice, and rock formations, it is not used exclusively for this purpose. h) Around 5 to 27 degrees Celsius: This option is incorrect because it is too specific and does not cover the broad range of temperature predictions that satellite technology can be used for. While satellite technology can provide temperature data, it is not used exclusively for predicting temperatures within a specific range. Therefore, the correct answer is c) what it's like outside each day, as satellite technology is commonly used for weather forecasting and predicting daily weather conditions." }</seq>
Scenario 3: Q - AC - FA
This serves as a fine-tuning control. No reasoning is provided in the output.
Scenario 4: Base
An un-fine-tuned control for comparison purposes.
Structured Generation
Structured generation ensures consistent response formats, which is crucial for reliable fine-tuning. Initial experiments faced difficulties with response consistency and structured generation can solve this.