Lyria

KO | EN

gitlite — search

#ai-agents #mcp #ai #typescript #claude-code #rust #claude #developer-tools #python #llm #ai-agent #cli

★ 4 Open GitHub ↗

No description available.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Lyria: A General LLM‑Driven Genetic Algorithm Framework for Problem Solving

This is the repo for the paper: **Lyria: A General LLM‑Driven Genetic Algorithm Framework for Problem Solving**.
All the code, data, results, and analysis are available in this repo.

## Folders
* `packages`: it consists of all the source code for running Lyria;
* `data`: it contains the problem dataset, i.e. Sudoku, GC, TSP;
* `results`: it contains all results; 
* `analysis`: it contains the analysis of Lyria.

**Note**: 
Since result files are quite large around 2.3G exceeding the support of Git, we have to use the Git LFS to store them, which makes the result files are only pointers to Git LFS.
However, to see the results, you can also check `analysis/main.ipynb` which keeps the analysis of the results. In addition, you can follow the `How To Run Lyria` described below to reproduce the results or run any other LLMs.

## How To Run Lyria
1. Clone/Download this repo;
2. Create the environment and install the dependencies using `conda env -f environment.yml` (if the environment name is conflicted with yours, change the value of the name field in `environment.yml`);
3. Activate this environment;
4. Change the `.env.example` to `.env`;
5. (Optional) Provide your OpenAI Key at `.env`, if you want to use OpenAI LLMs;
6. (Optional) Download [Ollama](https://ollama.com/) and pull the models you want to use. If you need to run LLMs with large parameter size or with GPU, you can change `OLLAMA_BASE_URL` in `.env` to your server address;
7. (Optional) Provide your [together.ai](https://www.together.ai/) Key at `.env`, if you want to use LLMs provided by together.ai;
8. (Optional) Provide your [DeepSeek](https://platform.deepseek.com) Key at `.env`, if your want to use DeepSeek LLMs, e.g. deepseek-chat;
9. Run `python main.py --help`, you will see the commands list, in which `algo` is to run algorithms we used in this work while `dgen` is for problem generation. 
10. Run `python main.py algo gc llmga --help`, you will see all the essential, default, and optional parameters for running Lyria. For other problems, you just need to replace `gc` with `sudoku` or `tsp`. For other methods, you just need to change `llmga` with `dp` or `dp-by-sampling-k`;

**Note**:
1. For 5-8, one of them must be provided to process the experiment;
2. For 9-10, you can always run `python main.py ... --help` to show available commands or hints;
3. To run the whole experiment may take around 20 days depending on your GPU. Thus, in the code, when solving problems, an algorithm is always run in parallel with `max_thread` fixed to 5 as a default, which can be changed in `packages/cli/algorithm_evaluator/graph_coloring.py`, for example, for GC problem. Therefore, you can reduce it if your server or LLM provider does not allow you to run in parallel. You can also increase it, if you think the progress is too slow;
4. In addition, the progress is normally very slow for `dp-by-sampling-k` (BoN) and `llmga` (Lyria) depending on their parameter settings, it may take several hours or even 1-2 days to update, especially if you run it in the machine without powerful GPUs. Thus, while we provide a nice progressbar to show the progress, we also log the running information in `logs` folder (but only for Lyria). Thus, when you see the progressbar is stuck for a long time, you can always go to the `logs` folder to check what is happening.

## Additional Explanation of Folders in `packages`:
1. `algorithms`: it consists of all algorithms we used in this work where `llmga` is Lyria;
2. `cli`: it implements command line interface;
3. `problem_generator`: it contains code to generate a single problem with the implementations of verifiers and metrics;
4. `data_generator`: it enables generating problems in batch;
5. `llm`: it consists of OpenAI compatible LLMs implementation;
6. `prompt`: it consists of various designed prompts;
7. `evluator`: it consists of evaluators to get statistics for solutions;
8. `utils`: it includes some util functions;

Note: you can go to `algorithms/llmga/default` to see the implements of _Error Detector_, _External Crossover Operator_, _External Mutation Operator_, etc., for different type of problems.

Lyria

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README