DeciLM 6B
DeciLM 6B is a 5.7 billion parameter decoder-only text generation model. With a context window of 4096 tokens, the highly efficient model uses variable Grouped-Query Attention (GQA) to achieve an optimal balance between performance and computational efficiency. The model's architecture was generated using Deci's proprietary Neural Architecture Search-based technology, AutoNAC.
Model Details
Model Description
Deci developed and publically released the DeciLM 6B large language model, a pretrained, high-efficiency generative text model with 5.7 billion parameters. DeciLM 6B outpaces pretrained models in its class, with a throughput that's up to 15 times that of Llama 2 7B's. DeciLM-6B was further fine-tuned using LoRA for instruction following on a subset of the OpenOrca dataset, creating DeciLM 6B-Instruct
- 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: Llama 2 Community License Agreement with an extention of Deci regarding hosting service providers.
Model Architecture
Parameters | Layers | Heads | Sequence Length | GQA num_key_value_heads* | Hidden Size |
---|---|---|---|---|---|
5.7B | 32 | 32 | 4096 | Variable | 4096 |
*AutoNAC was employed to optimize the selection of the GQA num_key_value_heads for each layer of the model.
- Decoder layer: Varible Grouped Query Attention. Grouped Query Attention (GQA) was introduced in Ainslie et al., 2023
- Position Embeddings: Dynamic NTK Scaling Rotary Position Embeddings Su et al., 2021
Model Sources
- Paper: DeciLM Technical Blog
- Demo: DeciLM 6B Instruct Demo
- Notebook: DeciLM 6B Notebook
Uses
The model is intended for commercial and research use in English and can be fine-tuned for use in other languages.
How to Get Started with the Model
Use the code below to get started with the model.
# pip install -q transformers
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
checkpoint = "Deci/DeciLM-6b"
device = "cuda" # for GPU usage or "cpu" for CPU usage
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint, torch_dtype=torch.bfloat16, trust_remote_code=True).to(device)
inputs = tokenizer.encode("In a shocking finding, scientists discovered a herd of unicorns living in", return_tensors="pt").to(device)
outputs = model.generate(inputs, max_new_tokens=100, do_sample=True, top_p=0.95)
print(tokenizer.decode(outputs[0]))
Training Details
DeciLM 6B underwent training utilizing a subset of the SlimPajamas dataset, leveraging advanced proprietary methodologies allowing for fast training.
Evaluation
Below are DeciLM's 6B evaluation results.
Average | ARC Challenge* | ARC Easy* | BoolQ | HellaSwag* | LAMBDA OpenAI | OpenBookQA | PIQA | TruthfulQA | Winogrande |
---|---|---|---|---|---|---|---|---|---|
60.33 | 42.06 | 70.02 | 71.01 | 74.58 | 69.78 | 34 | 77.09 | 36.19 | 68.03 |
Accuracy-norm score* |
Runtime Benchmarks
Inference Tool/Hardware | A10 (tokens/sec) |
---|---|
PyTorch | 652.49 |
Infery LLM | 2,029.6 |
- Throughput (tokens/sec) - Measured with optimal batch - PyTorch BS 64, Infery LLM BS 128
- In order to replicate the results of the PyTorch benchmark, use this code example
How to Cite
Please cite this model using this format.
@misc{DeciFoundationModels,
title = {DeciLM 6B},
author = {DeciAI Research Team},
year = {2023}
url={[https://huggingface.co/Deci/DeciLM-6b](https://huggingface.co/Deci/DeciLM-6b)},
}
- Downloads last month
- 774
Dataset used to train Deci/DeciLM-6b
Spaces using Deci/DeciLM-6b 3
Collection including Deci/DeciLM-6b
Evaluation results
- ARC Challenge on ai2_arcself-reported42.060
- ARC Easy on ai2_arcself-reported70.020
- BoolQ on boolqself-reported71.010
- HellaSwag on hellaswagself-reported74.580
- LAMBDA on OpenAI LAMBDAself-reported69.780
- OpenBookQA on openbookqaself-reported34.000
- PIQA on piqaself-reported77.090
- TruthfulQA on truthful_qaself-reported36.190
- Winogrande on winograndeself-reported68.030