--- license: apache-2.0 tags: - moe - frankenmoe - merge - mergekit - lazymergekit - cognitivecomputations/dolphin-2_6-phi-2 - rhysjones/phi-2-orange base_model: - cognitivecomputations/dolphin-2_6-phi-2 - rhysjones/phi-2-orange --- # PhiMiX-2x2B

PhiMiX-2x2B is a Mixure of Experts (MoE) made with the following models using mergekit: * [cognitivecomputations/dolphin-2_6-phi-2](https://huggingface.co/cognitivecomputations/dolphin-2_6-phi-2) * [rhysjones/phi-2-orange](https://huggingface.co/rhysjones/phi-2-orange) ## ©️ Credits * [mlabonne's phixtral](https://huggingface.co/mlabonne/phixtral-4x2_8) for the PhiConfig and inference code. * [mergekit](https://github.com/cg123/mergekit) code which I tweaked (you can find the PhiConfig [here](https://github.com/cg123/mergekit/blob/508348ae34be17ea0a95d0a288a6e34491a2558a/mergekit/architecture.py#L289)) by mainly adding the config in the `moe_mixtral.py` script from `mixtral` branch. ## ⏱️ Benchmarks | Model |AGIEval|GPT4All|TruthfulQA|Bigbench|Average| |----------------------------------------------------------------|------:|------:|---------:|-------:|------:| |[**PhiMiX-2x2B**](https://huggingface.co/paulilioaica/PhiMiX-2x2B) | **33.34** | **71.75** | 49.25 | **37.62** | **47.99** | |[phixtral-4x2_8](https://huggingface.co/mlabonne/phixtral-4x2_8)| 33.91| 70.44| 48.78| 37.68| 47.7| |[phixtral-2x2_8](https://huggingface.co/mlabonne/phixtral-2x2_8)| 34.1| 70.44| 48.78| 37.82| 47.78| |[_phi-2-orange_](https://huggingface.co/rhysjones/phi-2-orange)| _33.37_| _71.33_| _49.87_ | _37.3_ | _47.97_ | |[_dolphin-2_6-phi-2_](https://huggingface.co/cognitivecomputations/dolphin-2_6-phi-2)| _33.12_| _69.85_| _47.39_| _37.2_| _46.89_| I have used __bold__ to highlight this merge from the list, and _italics_ to highlight it's base modes used in the merge, and then __bold__ in the cells where it exceeds the performance of either. ## 🧩 Configuration ```yaml base_model: rhysjones/phi-2-orange gate_mode: cheap_embed dtype: float16 experts: - source_model: cognitivecomputations/dolphin-2_6-phi-2 positive_prompts: ["research, logic, math, science"] - source_model: rhysjones/phi-2-orange positive_prompts: ["programming, reasoning"] ``` ## 💻 Usage ```python !pip install -qU transformers bitsandbytes accelerate from transformers import AutoTokenizer import transformers import torch model = "paulilioaica/PhiMiX-2x2B" tokenizer = AutoTokenizer.from_pretrained(model) pipeline = transformers.pipeline( "text-generation", model=model, trust_remote_code=True, model_kwargs={"torch_dtype": torch.float16, "load_in_4bit": True,}, ) prompt="How many continents are there?" input = f"Instruct: {prompt}\nOutput:" outputs = pipeline(input, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95) print(outputs[0]["generated_text"]) ``` ```terminal Instruct: How many continents are there? Output: There are seven continents: Africa, Antarctica, Asia, Europe, North America, Australia, and South America. The total number of continents on Earth is seven, including Antarctica, which is sometimes considered part of the continent of Antarctica or as its own continent. It's important to note that the number of continents in popular education and geography is seven, but some sources may include Antarctica as its own continent, while others include it as part of the continent of Antarctica. Regardless of the exact categorization, there are seven continents that collectively make up the Earth's landmass. The continents can be divided into several subregions, such as islands, archipelagos, and microcontinents, which are smaller land masses surrounded by water. These subregions can be considered part of the continents or their own unique entities, depending on the context. Each continent has its own unique geography, climate, flora, fauna, and human cultures. The continents are interconnected through various landforms, bodies of water, and global trade routes. In summary, there are seven continents on Earth, each with its own distinct characteristics and unique contributions to the world's diversity. While the number may vary depending on the categorization of Antarctica, all seven continents together make ``` ## ♻️ Replicate this repo **beware** this will only work with 2 phis, you might have to tinker in the naming thing for more layers **AFTER** all the file modifications and run, you need to replace `configs.json` with the one from **this repo** **AFTER** that you need to add `modeling_phi.py` and `configurations.phi` from **this repo** to your repo Steps 1. Modify moe_mixtral.py from `/content/mergekit/mergekit/scripts/mixtral_moe.py` to your hf repo [***mixtral_moe.py***](https://github.com/paulilioaica/Phi-MOE/blob/main/mixtral_moe.py) 2. Modify architecture.py `/content/mergekit/mergekit/architecture.py` (this you can take from the link to the commit i have in description) [***architecture.py***](https://github.com/paulilioaica/Phi-MOE/blob/main/architecture.py) 3) replace `configs.json` with the one from **this repo** 4) you need to add `modeling_phi.py` and `configurations.phi` from **this repo** to your repo