https://github.com/armbues/SiLLM-examples

Examples for using the SiLLM framework for training and running Large Language Models (LLMs) on Apple Silicon
https://github.com/armbues/SiLLM-examples

apple-silicon dpo large-language-models llm llm-inference llm-training lora mlx

Last synced: about 2 months ago
JSON representation

Examples for using the SiLLM framework for training and running Large Language Models (LLMs) on Apple Silicon

• Host: GitHub
• URL: https://github.com/armbues/SiLLM-examples
• Owner: armbues
• License: mit
• Created: 2024-04-23T09:16:17.000Z (8 months ago)
• Default Branch: main
• Last Pushed: 2024-09-10T08:45:58.000Z (3 months ago)
• Last Synced: 2024-09-10T11:01:55.383Z (3 months ago)
• Topics: apple-silicon, dpo, large-language-models, llm, llm-inference, llm-training, lora, mlx
• Language: Python
• Homepage:
• Size: 17.6 KB
• Stars: 11
• Watchers: 2
• Forks: 2
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Examples for using SiLLM

Make sure to install the requirements for these examples:

``` sh

pip install -r requirements-examples.txt

```

## Control

Examples for training control vectors from a model's hidden states that can be used to influence the behavior and generated output during inference.

## HelpSteer LoRA

LoRA training [Mistral-7B-Instruct-v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2) with the Nvidia [HelpSteer](https://huggingface.co/datasets/nvidia/HelpSteer) dataset.

Run `train.sh` in the `helpsteer` directory to download the dataset & model from HuggingFace and start the LoRA training. You can customize the training configuration by editing `config.yml`.

## DPO-Mix-7K

DPO training [Qwen1.5-7B-Chat](https://huggingface.co/Qwen/Qwen1.5-7B-Chat) with the [DPO Mix 7K](https://huggingface.co/datasets/argilla/dpo-mix-7k) dataset. The training consists of a supervised fine tuning (SFT) followed by direct preference optimization (DPO).

Run `train.sh` in the `dpo-mix-7k` directory to download the dataset & model from HuggingFace and start the training. You can customize the training configuration by editing the config files `sft.yml` and `dpo.yml`.

## MMLU Benchmark

Implementation of the "Massive Multitask Language Understanding" benchmark using the [MMLU](https://huggingface.co/datasets/cais/mmlu) dataset.

Run `mmlu.py` with the model you would like to evaluate.

## MMLU-Pro Benchmark

Implementation of the MMLU-Pro benchmark using the [MMLU-Pro](https://huggingface.co/datasets/TIGER-Lab/MMLU-Pro) dataset.

Run `mmlu-pro.py` with the model you would like to evaluate.

## Perplexity

Calculating perplexity scores for a sample [dataset](https://huggingface.co/datasets/Cohere/wikipedia-2023-11-embed-multilingual-v3) of entry paragraphs from Wikipedia articles.

Run `perplexity.py` with the model you would like to evaluate. Add quantization options to evaluate perplexity with quantized models.