In the realm of natural language processing, Stable Beluga 2 stands out as a significant advancement. Crafted by Stability AI, this model represents cutting-edge technology in text generation. Its foundation is on the impressive Llama2 70B structure but further refined with an Orca-style dataset, elevating its capabilities beyond its predecessors.
Using Stable Beluga 2 is relatively straightforward once you've set up your coding environment. Below is a basic guide on how to interact with the model using Python:
Firstly, you’ll need to import the necessary libraries and load the model:
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
tokenizer = AutoTokenizer.from_pretrained("stabilityai/StableBeluga2", use_fast=False)
model = AutoModelForCausalLM.from_pretrained("stabilityai/StableBeluga2",
torch_dtype=torch.float16,
low_cpu_mem_usage=True,
device_map="auto")
Next, set up your prompt structure:
system_prompt = """
#### System:
You are Stable Beluga, an AI that follows instructions extremely well. Help as much as you can. Remember, be safe, and don't do anything illegal.
#### User:
Write me a poem please
#### Assistant:
"""
Then you can generate a response from the model:
prompt = f"{system_prompt}"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
output = model.generate(**inputs, do_sample=True, top_p=0.95, top_k=0, max_new_tokens=256)
print(tokenizer.decode(output[0], skip_special_tokens=True))
For those interested in different capacities or variations of the model, Stability AI offers a series of Beluga options such as:
· StableBeluga 1
· DeltaStableBeluga 13B
· StableBeluga 7B
Each of these models brings its unique attributes to the table, catering to diverse needs within text generation tasks.
Stable Beluga 2's architecture is based on an auto-regressive framework, enabling it to provide coherent and contextually relevant text. The model performs best with English, tested extensively to ensure quality outputs.
The training process of Stable Beluga 2 hinges on a robust internal dataset inspired by Orca. This ensures the model's versatility and adaptability to various text generation scenarios. The training involves:
· Mixed-precision learning for efficiency
· AdamW optimizer for improved convergence
· Batch sizes and learning rates fine-tuned for best performance
While Stable Beluga 2 represents a leap forward in language models, it's essential to remember it's still a tool with certain limitations. It
