🦙 Llama ParamPal — LLM Parameters Repository

Llama ParamPal — a community-driven digest to help you find the optimal recommended model parameters for running local LLMs using llama.cpp.

📦 What is this?

Finding the recommended sampling parameters to run llms can sometimes be a cumbersome and time-consuming process. This project aims to make this whole process a little bit easier to:

• Avoid guesswork when running llms.
• Contribute references and links to the models recommended parameter documentations.

The project consists of the models.json file that serves as the source of knowledge and a frontend that is available under https://llama-parampal.codecut.de/ which can be used to quickly search for models in this json.

---

✍️ How to Contribute

1. Fork this repo

2. Open the models.json file

3. Add your model or a profile under an existing one

4. Include:

   • A descriptive name for the profile
   • the llama.cpp CLI sampling parameters
   • At least one valid reference to a documentation of the model creators where those settings are documented.

5. Validate the JSON:

   cd validation
   npm install
   npm run validate
   

6. Submit a Pull Request - we'll review and merge!

💡 Make sure your JSON is valid and follows the existing structure. When in doubt, use current entries as examples.

Do's and don't do's when adding a commandline string.

• Make sure you define the max --ctx-size that is defined in the gguf headers as context_length of the model you are referencing. ( You can look it up in the huggingface gguf metadata viewer panel. )
• Don't add any hardware dependend parameters like ngl, sm and such. What this json is trying to accomplish is to document the sampling parameters.

---

💬 Questions or Ideas?

Open an issue or join the discussion at https://github.com/kseyhan/llama-param-pal.

---

🧪 Related Projects

• llama.cpp: An Llm inference engine in pure C/C++

---

🪄 License

MIT — free to use, improve, and share.