ExpertRamonda-7Bx2_MoE
ExpertRamonda-7Bx2_MoE is a Mixure of Experts (MoE) made with the following models using LazyMergekit:
🏆 Benchmarks
Open LLM Leaderboard
| Model | Average | ARC_easy | HellaSwag | MMLU | TruthfulQA_mc2 | Winogrande | GSM8K |
|---|
| mayacinka/ExpertRamonda-7Bx2_MoE | 78.10 | 86.87 | 87.51 | 61.63 | 78.02 | 81.85 | 72.71 |
MMLU
| Groups | Version | Filter | n-shot | Metric | Value | | Stderr |
|---|
| mmlu | N/A | none | 0 | acc | 0.6163 | ± | 0.0039 |
| - humanities | N/A | none | None | acc | 0.5719 | ± | 0.0067 |
| - other | N/A | none | None | acc | 0.6936 | ± | 0.0079 |
| - social_sciences | N/A | none | None | acc | 0.7121 | ± | 0.0080 |
| - stem | N/A | none | None | acc | 0.5128 | ± | 0.0085 |
🧩 Configuration
base_model: mlabonne/AlphaMonarch-7B
gate_mode: hidden
dtype: bfloat16
experts_per_token: 2
experts:
- source_model: mlabonne/AlphaMonarch-7B
positive_prompts:
- "You excel at reasoning skills. For every prompt you think of an answer from 3 different angles"
## (optional)
# negative_prompts:
# - "This is a prompt expert_model_1 should not be used for"
- source_model: bardsai/jaskier-7b-dpo-v5.6
positive_prompts:
- "You excel at logic and reasoning skills. Reply in a straightforward and concise way"
💻 Usage
!pip install -qU transformers bitsandbytes accelerate
from transformers import AutoTokenizer
import transformers
import torch
model = "mayacinka/ExpertRamonda-7Bx2_MoE"
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"])