DeciLM-7B-instruct

DeciLM-7B-instruct is a model for short-form instruction following. It is built by LoRA fine-tuning on the SlimOrca dataset.

Model Details

Model Description

DeciLM-7B-instruct is a derivative of the recently released DeciLM-7B language model, a pre-trained, high-efficiency generative text model with 7 billion parameters. DeciLM-7B-instruct is one the best 7B instruct models obtained using simple LoRA fine-tuning, without relying on preference optimization techniques such as RLHF and DPO.

• Developed by: Deci
• Model type: DeciLM is an auto-regressive language model using an optimized transformer decoder architecture that includes variable Grouped-Query Attention.
• Language(s) (NLP): English
• License: Apache 2.0

Model Architecture

Parameters	Layers	Heads	Sequence Length	GQA num_key_value_heads*
7.04 billion	32	32	8192	Variable

*AutoNAC was employed to optimize the selection of the GQA num_key_value_heads for each model layer.

Model Sources

• Blog: DeciLM-7B Technical Blog
• Demo: DeciLM-7B-instruct Demo
• Finetuning Notebook: DeciLM-7B Finetuning Notebook
• Text Generation Notebook: DeciLM-7B-instruct Text Generation Notebook

Prompt Template

### System:
{system_prompt}
### User:
{user_prompt}
### Assistant:

Uses

The model is intended for commercial and research use in English.

How to Get Started with the Model

Use the code below to get started with the model.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig, pipeline

model_name = "Deci/DeciLM-7B-instruct"

device = "cuda" # for GPU usage or "cpu" for CPU usage

quantize = False  # Optional. Useful for GPUs with less than 24GB memory

if quantize:
    dtype_kwargs = dict(quantization_config=BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_compute_dtype=torch.bfloat16
    ))
else:
    dtype_kwargs = dict(torch_dtype="auto")

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    trust_remote_code=True,
    **dtype_kwargs
)

tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token

deci_generator = pipeline("text-generation",
                          model=model,
                          tokenizer=tokenizer,
                          temperature=0.1,
                          device_map="auto",
                          max_length=4096,
                          return_full_text=False)

system_prompt = "You are an AI assistant that follows instruction extremely well. Help as much as you can."

user_prompt = "How do I make the most delicious pancakes the world has ever tasted?"

prompt = tokenizer.apply_chat_template([
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": user_prompt},
], tokenize=False, add_generation_prompt=True)

response = deci_generator(prompt)[0]['generated_text']
print(prompt + response)

Evaluation

Below are DeciLM-7B and DeciLM-7B-instruct's evaluation results.

Model	Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K
DecilLM-7B	61.55	59.39	82.51	59.76	40.33	79.95	47.38
DecilLM-7B-instruct	63.19	61.01	82.37	60.24	49.75	79.72	46.02

Runtime Benchmarks

Inference Tool	Hardware	Prompt length	Generation length	Generated tokens/sec	Batch Size	Number of Prompts
HuggingFace (PyTorch)	A100 (SXM4-80GB-400W)	512	512	1174	352	352
HuggingFace (PyTorch)	A100 (SXM4-80GB-400W)	2048	2048	328	72	72
Infery-LLM	A100 (SXM4-80GB-400W)	512	512	4559	1024	4096
Infery-LLM	A100 (SXM4-80GB-400W)	2048	2048	3997	512	2048
Infery-LLM	A10	512	512	1345	128	512
Infery-LLM	A10	2048	2048	599	32	128

• In order to replicate the results of the Hugging Face benchmarks, you can use this code example.
• Infery-LLM, Deci's inference engine, features a suite of optimization algorithms, including selective quantization, optimized beam search, continuous batching, and custom CUDA kernels. To explore the full capabilities of Infery-LLM, schedule a live demo.

Ethical Considerations and Limitations

DeciLM-7B-instruct is a new technology that comes with inherent risks associated with its use. The testing conducted so far has been primarily in English and does not encompass all possible scenarios. Like those of all large language models, DeciLM-7B's outputs are unpredictable, and the model may generate responses that are inaccurate, biased, or otherwise objectionable. Consequently, developers planning to use DeciLM-7B should undertake thorough safety testing and tuning designed explicitly for their intended applications of the model before deployment.

How to Cite

Please cite this model using this format.

@misc{DeciFoundationModels,
title = {DeciLM-7B-instruct},
author = {DeciAI Research Team},
year = {2023}
url={https://huggingface.co/Deci/DeciLM-7B-instruct},
}