Instructions to use SteelStorage/Umbra-v2.1-MoE-4x10.7 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use SteelStorage/Umbra-v2.1-MoE-4x10.7 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="SteelStorage/Umbra-v2.1-MoE-4x10.7")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("SteelStorage/Umbra-v2.1-MoE-4x10.7")
model = AutoModelForCausalLM.from_pretrained("SteelStorage/Umbra-v2.1-MoE-4x10.7")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use SteelStorage/Umbra-v2.1-MoE-4x10.7 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "SteelStorage/Umbra-v2.1-MoE-4x10.7"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "SteelStorage/Umbra-v2.1-MoE-4x10.7",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/SteelStorage/Umbra-v2.1-MoE-4x10.7

SGLang

How to use SteelStorage/Umbra-v2.1-MoE-4x10.7 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "SteelStorage/Umbra-v2.1-MoE-4x10.7" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "SteelStorage/Umbra-v2.1-MoE-4x10.7",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "SteelStorage/Umbra-v2.1-MoE-4x10.7" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "SteelStorage/Umbra-v2.1-MoE-4x10.7",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use SteelStorage/Umbra-v2.1-MoE-4x10.7 with Docker Model Runner:
```
docker model run hf.co/SteelStorage/Umbra-v2.1-MoE-4x10.7
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Umbra-v2.1-MoE-4x10.7

The [Umbra Series] is an offshoot of the [Lumosia Series] With the goal to be a General assistant that has a knack for story telling and RP/ERP

-What's New in v2.1?

Umbra v2.1 isn't just a simple update; it's like giving the model a double shot of espresso. Ive changed the models and prompts, in an attempt to make Umbra not only your go-to assistant for general knowledge but also a great storyteller and RP/ERP companion.

-Longer Positive, Shorter Negative

In an effort to trick the gates into being less uptight, Ive added more positive prompts and snappier negative ones. These changes are based on the model's strengths and, frankly, my whimsical preferences.

-Experimental, As Always

Remember, folks, "v2.1" doesn't mean it's superior to its predecessors – it's just another step in the quest. It's the 'Empire Strikes Back' of our series – could be better, could be worse, but definitely more dramatic.

-Base Context and Coherence

Umbra v2.1 has a base context of 8k scrolling window.

-The Tavern Card

Just for fun - the Umbra Personality Tavern Card. It's your gateway to immersive storytelling experiences, a little like having a 'Choose Your Own Adventure' book, but way cooler because it's digital and doesn't get lost under your bed.

-Token Error? Fixed!

Umbra-v2 had a tokenizer error but was removed faster than you can say "Cops love Donuts"

So, give Umbra v2.1 a whirl and let me know how it goes. Your feedback is like the secret sauce in my development burger.

### System:

### USER:{prompt}

### Assistant:

Settings:

Temp: 1.0
min-p: 0.02-0.1

Evals:

Avg: 73.59
ARC: 69.11
HellaSwag: 87.57
MMLU: 66.48
T-QA: 66.75
Winogrande: 83.11
GSM8K: 68.69

Examples:

posted soon

posted soon

🧩 Configuration

base_model: vicgalle/CarbonBeagle-11B
gate_mode: hidden
dtype: bfloat16
experts:
  - source_model: vicgalle/CarbonBeagle-11B
    positive_prompts: [Revamped]

  - source_model: Sao10K/Fimbulvetr-10.7B-v1
    positive_prompts: [Revamped]

  - source_model: bn22/Nous-Hermes-2-SOLAR-10.7B-MISALIGNED
    positive_prompts: [Revamped]

  - source_model: Yhyu13/LMCocktail-10.7B-v1
    positive_prompts: [Revamed]

Umbra-v2-MoE-4x10.7 is a Mixure of Experts (MoE) made with the following models:
* [vicgalle/CarbonBeagle-11B](https://huggingface.co/vicgalle/CarbonBeagle-11B)
* [Sao10K/Fimbulvetr-10.7B-v1](https://huggingface.co/Sao10K/Fimbulvetr-10.7B-v1)
* [bn22/Nous-Hermes-2-SOLAR-10.7B-MISALIGNED](https://huggingface.co/bn22/Nous-Hermes-2-SOLAR-10.7B-MISALIGNED)
* [Yhyu13/LMCocktail-10.7B-v1](https://huggingface.co/Yhyu13/LMCocktail-10.7B-v1)

💻 Usage

!pip install -qU transformers bitsandbytes accelerate
from transformers import AutoTokenizer
import transformers
import torch

model = "Steelskull/Umbra-v2-MoE-4x10.7"

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"])

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	73.59
AI2 Reasoning Challenge (25-Shot)	69.11
HellaSwag (10-Shot)	87.57
MMLU (5-Shot)	66.48
TruthfulQA (0-shot)	66.57
Winogrande (5-shot)	83.11
GSM8k (5-shot)	68.69

Downloads last month: 30

Safetensors

Model size

36B params

Tensor type

BF16

Model tree for SteelStorage/Umbra-v2.1-MoE-4x10.7

Quantizations

2 models

Evaluation results

normalized accuracy on AI2 Reasoning Challenge (25-Shot)
test set Open LLM Leaderboard

69.110
normalized accuracy on HellaSwag (10-Shot)
validation set Open LLM Leaderboard

87.570
accuracy on MMLU (5-Shot)
test set Open LLM Leaderboard

66.480
mc2 on TruthfulQA (0-shot)
validation set Open LLM Leaderboard

66.570
accuracy on Winogrande (5-shot)
validation set Open LLM Leaderboard

83.110
accuracy on GSM8k (5-shot)
test set Open LLM Leaderboard

68.690