Code Generation with AlphaCodium: From Prompt Engineering to Flow Engineering

Official Implementation

Tal Ridnik, Dedy Kredo, Itamar Friedman <br/> CodiumAI

News 2024-17-05

Updated AlphaCodium leaderboard with scores of new GPT models, and Claude3 Opus. "GPT-4o" Is currently the leading model on AlphaCodium.

Abstract
Installation
How to run
Technical Q&A
Broader Applicability
Example Problem
Acknowledgments
Citation

Abstract

Code generation problems differ from common natural language problems - they require matching the exact syntax of the target language, identifying happy paths and edge cases, paying attention to numerous small details in the problem spec, and addressing other code-specific issues and requirements. Hence, many of the optimizations and tricks that have been successful in natural language generation may not be effective for code tasks.

In this work, we propose a new approach to code generation by LLMs, which we call AlphaCodium - a test-based, multi-stage, code-oriented iterative flow, that improves the performances of LLMs on code problems.

We tested AlphaCodium on a challenging code generation dataset called CodeContests, which includes competitive programming problems from platforms such as Codeforces. The proposed flow consistently and significantly improves results. On the validation set, for example, GPT-4 accuracy (pass@5) increased from 19% with a single well-designed direct prompt to 44% with the AlphaCodium flow.

Many of the principles and best practices we acquired in this work, we believe, are broadly applicable to general code generation tasks.

Installation

(1) setup a virtual environment：

python3 -m venv venv
source ./venv/bin/activate

and run: pip install -r requirements.txt.

(2) Duplicate the file alpha_codium/settings/.secrets_template.toml, rename it as alpha_codium/settings/.secrets.toml, and fill in your OpenAI API key:

[openai]
key = "..."

(3) Download the processed CodeContest validation and test dataset from hugging face, extract the zip file, and placed the extracted folder in the root of the project.

How to run

Configuration

The file: alpha_codium/settings/configuration.toml contains the configuration for the project. In the config section you can choose the model you want to use ("gpt-4", "gpt-3.5-turbo-16k", or others).

Solving a specific problem from CodeContest

To solve a specific problem with AlphaCodium, from the root folder run:

python -m alpha_codium.solve_problem \
--dataset_name /path/to/dataset \
--split_name test \
--problem_number 0

The dataset_name is the path to the dataset folder you downloaded in the installation step.
Note that the validation set contains 117 problems, and the test set contains 165 problems, so the problem_number parameter should be accordingly (zero-based)
The split_name can be either valid or test.
The following sections in the configuration file: solve, self_reflection,possible_solutions,generate_ai_tests,initial_code_generation,public_tests, ai_tests
enable to adjust possible configurations for the different stages of the flow.
Each run logs the results to a file named alpha_codium/example.log. Reviewing the log file is a good way to understand what is going on in each stage of the flow.

Example problem (test set, problem number 12):

Solving an entire CodeContest dataset split

to solve the entire dataset with AlphaCodium, from the root folder run:

python -m alpha_codium.solve_dataset \
--dataset_name /path/to/dataset \
--split_name test \
--database_solution_path /path/to/output/dir/dataset_output.json

The split_name can be either valid or test.
database_solution_path is the path to the directory where the solutions will be saved.
The dataset section in the configuration file contains the configuration for the running and evaluation of a dataset.
Note that this is a long process, and it may take a few days to complete with large models (e.g. GPT-4) and several iterations per problem.
dataset.num_iterations defines the number of iterations for each problem (pass@K). For a large number of iterations, it is recommended to introduce some randomness and different options for each iteration to achieve top results.

Running the evaluation

Once you generate a solution for the entire dataset (valid or test), you can evaluate it by running:

python -m alpha_codium.evaluate_dataset \
--dataset_name /path/to/dataset \
--split_name test \
--database_solution_path /path/to/output/dir/dataset_output.json

Solving a new problem (CodeContest format)

To solve a custom problem with AlphaCodium, first create a json file that includes the CodeContest problem fields, and then from the root folder run:

python -m alpha_codium.solve_my_problem \
--my_problem_json_file /path/to/my_problem.json

The my_problem_json_file is the path to to the custom problem json file.

See the my_problem_example.json to see an example of a custom problem. The json file should include the following fields:

name is the name of the problem.
description is a description of the problem.
(optional) public_tests with the following fields:
- input is a list of strings that represent the input.
- output is a list of strings that represent the output.
(optional) private_tests, that follows the same structure as public_tests
(optional) generated_tests, that follows the same structure as public_tests

Technical Q&A

Aggregating some technical questions we received about this project:

Q: How much time did you spend on "prompt engineering" compared to "flow engineering"?<br><br> A: Structured output almost completely eliminates the need for simple prompt engineering. We estimate that ~95% of the time we did more high-level design, reasoning, and injecting data at the correct places, ..., a.k.a. "flow engineering".

Q: How do you know that there wasn't a data leakage? <br><br> A: The test set of CodeContests dataset comprises problems published after September 2021, while the GPT-4 model variant we used (gpt-4-0613) has a data cutoff of September 2021. Hence, there is no data leakage for GPT4, on the test set. For other models like DeepSeek, we cannot be sure. However, note that our main result is a comparison of "direct prompt" vs. "AlphaCodium flow". Data leakage would help both approaches, so the relative improvement of AlphaCodium flow is still valid.

Q: Is this project relevant only to specific programming languages?<br><br> A: No. The proposed flow is language agnostic. We generated solutions in Python, but the flow can be applied to any language.

Q: How did you manage the context window? <br><br> A: We used models with a context window of 8192 tokens, and we did not encounter cases where it did not suffice. However, we clearly observed that as the context we used in practice grows larger (let's say, above 4000 tokens), the model starts to "ignore" some of the information in the context. Hence, there is a clear tradeoff:

Injecting the results of previous stages into the context, may help the model to generate better code.
However, it may also cause the model to ignore specific details and nuances from the problem description.

Q: Is this work "realistic" in terms of the number of LLM calls? <br><br> A: In comparison to AlphaCode, we do four orders of magnitude (!) fewer calls (per solution AlphaCodium does 15-20 calls). Yet we acknowledge that for some applications, this may still be too much, and more optimizations are needed. We however believe that many of the ideas and principles we acquired in this work are broadly applicable, even when the number of calls is further limited.

Q: Why do you iterate only on the generated code, and not on the AI-generated tests? <br><br> A: For code problems in CodeContests, the tests are a list of input-output pairs. Hence, you don't really learn anything new when you "fix" a test - you just change its output to the prediction of the generated code. Instead of fixing tests, we preferred to always try and fix the code, while using "test anchors". (see the paper for more details). However, for other code generation tasks, where the tests are more complex and contain runnable code, iterating on the tests, in addition to iterating on the generated code, may be beneficial.

Broader Applicability

While this work presents results on CodeContests dataset, we believe that it has a broader applicability.

First and foremost, we feel that the proposed AlphaCodium flow, with reasonable adjustments, can be used as a more general framework for other code generation tasks.

Secondly, many of the design concepts, prin

AlphaCodium

Install / Use

README