README.md

---
license: apache-2.0
tags:
- moe
- frankenmoe
- merge
- mergekit
- lazymergekit
- DeepMount00/Minerva-3B-base-RAG
- FairMind/Minerva-3B-Instruct-v1.0
base_model:
- DeepMount00/Minerva-3B-base-RAG
- FairMind/Minerva-3B-Instruct-v1.0
---

# Minerva-MoE-3x3B

Minerva-MoE-3x3B is a Mixture of Experts (MoE) made with the following models using [LazyMergekit](https://colab.research.google.com/drive/1obulZ1ROXHjYLn6PPZJwRR6GzgQogxxb?usp=sharing):
* [DeepMount00/Minerva-3B-base-RAG](https://huggingface.co/DeepMount00/Minerva-3B-base-RAG)
* [FairMind/Minerva-3B-Instruct-v1.0](https://huggingface.co/FairMind/Minerva-3B-Instruct-v1.0)

## Evaluation
arc_it acc_norm: 31.91
hellaswag_it acc_norm: 52.20
mmmlu_it: 25.72

## 🧩 Configuration

```yaml
base_model: sapienzanlp/Minerva-3B-base-v1.0
experts:
  - source_model: DeepMount00/Minerva-3B-base-RAG
    positive_prompts:
    - "rispondi a domande"
    - "cosa è"
    - "chi è"
    - "dove è"
    - "come si"
    - "spiegami"
    - "definisci"
  - source_model: FairMind/Minerva-3B-Instruct-v1.0
    positive_prompts:
    - "istruzione"
    - "input"
    - "risposta"
    - "scrivi"
    - "sequenza"
    - "istruzioni"
dtype: bfloat16
```

## 💻 Usage

```python
!pip install -qU transformers bitsandbytes accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "ludocomito/Minerva-MoE-3x3B"

tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    model_kwargs={"torch_dtype": torch.float16, "load_in_4bit": True},
)

messages = [{"role": "user", "content": "Explain what a Mixture of Experts is in less than 100 words."}]
prompt = pipeline.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])
```