frankgpt/_Frank_/markdown/style.cot.instructions.md

Chain-of-Thought (CoT) Prompting Engine Guide

1. Prompting Techniques

There are three primary methods for implementing CoT prompting, each with its own advantages.

2.1. Few-Shot CoT

This is the standard approach where you provide the model with a few examples (demonstrations) that include a question, a step-by-step reasoning process (the chain of thought), and the final answer.

When to Use: Use this method for complex tasks where the reasoning structure is consistent and providing diverse, high-quality examples can significantly guide the model. This is the most powerful method but requires manual effort to create the demonstrations.

Example Prompt Structure:

Q: [Question 1]

A: [Step-by-step reasoning for Question 1]. The answer is [Answer 1].

Q: [Question 2]

A: [Step-by-step reasoning for Question 2]. The answer is [Answer 2].

Q: [New Question]

A:

The file demos/multiarith_manual provides a practical example of the JSON structure for these hand-crafted demonstrations.

2.2. Zero-Shot CoT

A surprisingly effective and simple method that requires no examples. By appending the phrase "Let's think step by step" to the end of a question, the model is triggered to generate a reasoning chain before giving the final answer.

When to Use: This is an excellent starting point for any reasoning task. It's highly effective for its simplicity and is particularly useful when you don't have time to create few-shot examples.

Example Prompt Structure:

Q: [New Question]

A: Let's think step by step.

The api.py script in the repository shows how this is implemented by setting a cot_trigger argument. The Jupyter notebooks (try_cot.ipynb and try_cot_colab.ipynb) demonstrate its application and output.

2.3. Automatic CoT (Auto-CoT)

Auto-CoT is an advanced technique designed to automate the creation of diverse and effective demonstrations for Few-Shot CoT, eliminating the manual effort. As detailed in the project's README.md, it works in two main stages.

Stage 1: Question Clustering

• The system takes a dataset of questions and groups them into several clusters based on semantic similarity.

Stage 2: Demonstration Sampling

• It selects a representative question from each cluster.
• It then uses Zero-Shot CoT to automatically generate a reasoning chain for each selected question.

This process, detailed in run_demo.py, ensures that the examples are both diverse (by sampling from different clusters) and accurate, creating a robust set of demonstrations for the model to learn from. The output of this process can be seen in the demos/multiarith_auto file.

When to Use: Use Auto-CoT when you need the high performance of Few-Shot CoT on a large dataset of questions but want to avoid the time-consuming and potentially suboptimal process of manually writing demonstrations.

3. Implementation in the Repository

The provided repository contains a full implementation of these techniques.

• api.py: A core file that defines the cot function, which can be called with different methods: "zero_shot", "zero_shot_cot", "manual_cot", and "auto_cot".
• run_inference.py: The main script for running experiments. It loads a dataset, constructs prompts based on the chosen method, and generates answers.
• run_demo.py: This script implements the Auto-CoT process by clustering questions and generating demonstrations.
• try_cot.ipynb: A Jupyter Notebook that provides a quick and clear way to test and compare the outputs of each CoT method.

To get started, refer to the README.md and the try_cot_colab.ipynb for a guided walkthrough.

4. References

• Amazon Science Repo on CoT
• CoT Example