|
--- |
|
base_model: budecosystem/genz-13b-v2 |
|
inference: false |
|
language: |
|
- en |
|
library_name: transformers |
|
license: llama2 |
|
model_creator: Bud |
|
model_name: GenZ 13B v2 |
|
model_type: llama |
|
pipeline_tag: text-generation |
|
prompt_template: '### User: |
|
|
|
{prompt} |
|
|
|
|
|
### Assistant: |
|
|
|
' |
|
quantized_by: TheBloke |
|
--- |
|
|
|
<!-- header start --> |
|
<!-- 200823 --> |
|
<div style="width: auto; margin-left: auto; margin-right: auto"> |
|
<img src="https://i.imgur.com/EBdldam.jpg" alt="TheBlokeAI" style="width: 100%; min-width: 400px; display: block; margin: auto;"> |
|
</div> |
|
<div style="display: flex; justify-content: space-between; width: 100%;"> |
|
<div style="display: flex; flex-direction: column; align-items: flex-start;"> |
|
<p style="margin-top: 0.5em; margin-bottom: 0em;"><a href="https://discord.gg/theblokeai">Chat & support: TheBloke's Discord server</a></p> |
|
</div> |
|
<div style="display: flex; flex-direction: column; align-items: flex-end;"> |
|
<p style="margin-top: 0.5em; margin-bottom: 0em;"><a href="https://www.patreon.com/TheBlokeAI">Want to contribute? TheBloke's Patreon page</a></p> |
|
</div> |
|
</div> |
|
<div style="text-align:center; margin-top: 0em; margin-bottom: 0em"><p style="margin-top: 0.25em; margin-bottom: 0em;">TheBloke's LLM work is generously supported by a grant from <a href="https://a16z.com">andreessen horowitz (a16z)</a></p></div> |
|
<hr style="margin-top: 1.0em; margin-bottom: 1.0em;"> |
|
<!-- header end --> |
|
|
|
# GenZ 13B v2 - AWQ |
|
- Model creator: [Bud](https://huggingface.co/budecosystem) |
|
- Original model: [GenZ 13B v2](https://huggingface.co/budecosystem/genz-13b-v2) |
|
|
|
<!-- description start --> |
|
## Description |
|
|
|
This repo contains AWQ model files for [Bud's GenZ 13B v2](https://huggingface.co/budecosystem/genz-13b-v2). |
|
|
|
|
|
### About AWQ |
|
|
|
AWQ is an efficient, accurate and blazing-fast low-bit weight quantization method, currently supporting 4-bit quantization. Compared to GPTQ, it offers faster Transformers-based inference. |
|
|
|
It is also now supported by continuous batching server [vLLM](https://github.com/vllm-project/vllm), allowing use of Llama AWQ models for high-throughput concurrent inference in multi-user server scenarios. |
|
|
|
As of September 25th 2023, preliminary Llama-only AWQ support has also been added to [Huggingface Text Generation Inference (TGI)](https://github.com/huggingface/text-generation-inference). |
|
|
|
Note that, at the time of writing, overall throughput is still lower than running vLLM or TGI with unquantised models, however using AWQ enables using much smaller GPUs which can lead to easier deployment and overall cost savings. For example, a 70B model can be run on 1 x 48GB GPU instead of 2 x 80GB. |
|
<!-- description end --> |
|
<!-- repositories-available start --> |
|
## Repositories available |
|
|
|
* [AWQ model(s) for GPU inference.](https://huggingface.co/TheBloke/genz-13B-v2-AWQ) |
|
* [GPTQ models for GPU inference, with multiple quantisation parameter options.](https://huggingface.co/TheBloke/genz-13B-v2-GPTQ) |
|
* [2, 3, 4, 5, 6 and 8-bit GGUF models for CPU+GPU inference](https://huggingface.co/TheBloke/genz-13B-v2-GGUF) |
|
* [Bud's original unquantised fp16 model in pytorch format, for GPU inference and for further conversions](https://huggingface.co/budecosystem/genz-13b-v2) |
|
<!-- repositories-available end --> |
|
|
|
<!-- prompt-template start --> |
|
## Prompt template: User-Assistant-Newlines |
|
|
|
``` |
|
### User: |
|
{prompt} |
|
|
|
### Assistant: |
|
|
|
``` |
|
|
|
<!-- prompt-template end --> |
|
|
|
|
|
<!-- README_AWQ.md-provided-files start --> |
|
## Provided files, and AWQ parameters |
|
|
|
For my first release of AWQ models, I am releasing 128g models only. I will consider adding 32g as well if there is interest, and once I have done perplexity and evaluation comparisons, but at this time 32g models are still not fully tested with AutoAWQ and vLLM. |
|
|
|
Models are released as sharded safetensors files. |
|
|
|
| Branch | Bits | GS | AWQ Dataset | Seq Len | Size | |
|
| ------ | ---- | -- | ----------- | ------- | ---- | |
|
| [main](https://huggingface.co/TheBloke/genz-13B-v2-AWQ/tree/main) | 4 | 128 | [wikitext](https://huggingface.co/datasets/wikitext/viewer/wikitext-2-v1/test) | 4096 | 7.25 GB |
|
|
|
<!-- README_AWQ.md-provided-files end --> |
|
|
|
<!-- README_AWQ.md-use-from-vllm start --> |
|
## Serving this model from vLLM |
|
|
|
Documentation on installing and using vLLM [can be found here](https://vllm.readthedocs.io/en/latest/). |
|
|
|
Note: at the time of writing, vLLM has not yet done a new release with AWQ support. |
|
|
|
If you try the vLLM examples below and get an error about `quantization` being unrecognised, or other AWQ-related issues, please install vLLM from Github source. |
|
|
|
- When using vLLM as a server, pass the `--quantization awq` parameter, for example: |
|
|
|
```shell |
|
python3 python -m vllm.entrypoints.api_server --model TheBloke/genz-13B-v2-AWQ --quantization awq --dtype half |
|
``` |
|
|
|
When using vLLM from Python code, pass the `quantization=awq` parameter, for example: |
|
|
|
```python |
|
from vllm import LLM, SamplingParams |
|
|
|
prompts = [ |
|
"Hello, my name is", |
|
"The president of the United States is", |
|
"The capital of France is", |
|
"The future of AI is", |
|
] |
|
sampling_params = SamplingParams(temperature=0.8, top_p=0.95) |
|
|
|
llm = LLM(model="TheBloke/genz-13B-v2-AWQ", quantization="awq", dtype="half") |
|
|
|
outputs = llm.generate(prompts, sampling_params) |
|
|
|
# Print the outputs. |
|
for output in outputs: |
|
prompt = output.prompt |
|
generated_text = output.outputs[0].text |
|
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") |
|
``` |
|
<!-- README_AWQ.md-use-from-vllm start --> |
|
|
|
<!-- README_AWQ.md-use-from-tgi start --> |
|
## Serving this model from Text Generation Inference (TGI) |
|
|
|
Use TGI version 1.1.0 or later. The official Docker container is: `ghcr.io/huggingface/text-generation-inference:1.1.0` |
|
|
|
Example Docker parameters: |
|
|
|
```shell |
|
--model-id TheBloke/genz-13B-v2-AWQ --port 3000 --quantize awq --max-input-length 3696 --max-total-tokens 4096 --max-batch-prefill-tokens 4096 |
|
``` |
|
|
|
Example Python code for interfacing with TGI (requires huggingface-hub 0.17.0 or later): |
|
|
|
```shell |
|
pip3 install huggingface-hub |
|
``` |
|
|
|
```python |
|
from huggingface_hub import InferenceClient |
|
|
|
endpoint_url = "https://your-endpoint-url-here" |
|
|
|
prompt = "Tell me about AI" |
|
prompt_template=f'''### User: |
|
{prompt} |
|
|
|
### Assistant: |
|
|
|
''' |
|
|
|
client = InferenceClient(endpoint_url) |
|
response = client.text_generation(prompt, |
|
max_new_tokens=128, |
|
do_sample=True, |
|
temperature=0.7, |
|
top_p=0.95, |
|
top_k=40, |
|
repetition_penalty=1.1) |
|
|
|
print(f"Model output: {response}") |
|
``` |
|
<!-- README_AWQ.md-use-from-tgi end --> |
|
|
|
<!-- README_AWQ.md-use-from-python start --> |
|
## How to use this AWQ model from Python code |
|
|
|
### Install the necessary packages |
|
|
|
Requires: [AutoAWQ](https://github.com/casper-hansen/AutoAWQ) 0.1.1 or later |
|
|
|
```shell |
|
pip3 install autoawq |
|
``` |
|
|
|
If you have problems installing [AutoAWQ](https://github.com/casper-hansen/AutoAWQ) using the pre-built wheels, install it from source instead: |
|
|
|
```shell |
|
pip3 uninstall -y autoawq |
|
git clone https://github.com/casper-hansen/AutoAWQ |
|
cd AutoAWQ |
|
pip3 install . |
|
``` |
|
|
|
### You can then try the following example code |
|
|
|
```python |
|
from awq import AutoAWQForCausalLM |
|
from transformers import AutoTokenizer |
|
|
|
model_name_or_path = "TheBloke/genz-13B-v2-AWQ" |
|
|
|
# Load model |
|
model = AutoAWQForCausalLM.from_quantized(model_name_or_path, fuse_layers=True, |
|
trust_remote_code=False, safetensors=True) |
|
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=False) |
|
|
|
prompt = "Tell me about AI" |
|
prompt_template=f'''### User: |
|
{prompt} |
|
|
|
### Assistant: |
|
|
|
''' |
|
|
|
print("\n\n*** Generate:") |
|
|
|
tokens = tokenizer( |
|
prompt_template, |
|
return_tensors='pt' |
|
).input_ids.cuda() |
|
|
|
# Generate output |
|
generation_output = model.generate( |
|
tokens, |
|
do_sample=True, |
|
temperature=0.7, |
|
top_p=0.95, |
|
top_k=40, |
|
max_new_tokens=512 |
|
) |
|
|
|
print("Output: ", tokenizer.decode(generation_output[0])) |
|
|
|
""" |
|
# Inference should be possible with transformers pipeline as well in future |
|
# But currently this is not yet supported by AutoAWQ (correct as of September 25th 2023) |
|
from transformers import pipeline |
|
|
|
print("*** Pipeline:") |
|
pipe = pipeline( |
|
"text-generation", |
|
model=model, |
|
tokenizer=tokenizer, |
|
max_new_tokens=512, |
|
do_sample=True, |
|
temperature=0.7, |
|
top_p=0.95, |
|
top_k=40, |
|
repetition_penalty=1.1 |
|
) |
|
|
|
print(pipe(prompt_template)[0]['generated_text']) |
|
""" |
|
``` |
|
<!-- README_AWQ.md-use-from-python end --> |
|
|
|
<!-- README_AWQ.md-compatibility start --> |
|
## Compatibility |
|
|
|
The files provided are tested to work with: |
|
|
|
- [AutoAWQ](https://github.com/casper-hansen/AutoAWQ) |
|
- [vLLM](https://github.com/vllm-project/vllm) |
|
- [Huggingface Text Generation Inference (TGI)](https://github.com/huggingface/text-generation-inference) |
|
|
|
TGI merged AWQ support on September 25th, 2023: [TGI PR #1054](https://github.com/huggingface/text-generation-inference/pull/1054). Use the `:latest` Docker container until the next TGI release is made. |
|
|
|
<!-- README_AWQ.md-compatibility end --> |
|
|
|
<!-- footer start --> |
|
<!-- 200823 --> |
|
## Discord |
|
|
|
For further support, and discussions on these models and AI in general, join us at: |
|
|
|
[TheBloke AI's Discord server](https://discord.gg/theblokeai) |
|
|
|
## Thanks, and how to contribute |
|
|
|
Thanks to the [chirper.ai](https://chirper.ai) team! |
|
|
|
Thanks to Clay from [gpus.llm-utils.org](llm-utils)! |
|
|
|
I've had a lot of people ask if they can contribute. I enjoy providing models and helping people, and would love to be able to spend even more time doing it, as well as expanding into new projects like fine tuning/training. |
|
|
|
If you're able and willing to contribute it will be most gratefully received and will help me to keep providing more models, and to start work on new AI projects. |
|
|
|
Donaters will get priority support on any and all AI/LLM/model questions and requests, access to a private Discord room, plus other benefits. |
|
|
|
* Patreon: https://patreon.com/TheBlokeAI |
|
* Ko-Fi: https://ko-fi.com/TheBlokeAI |
|
|
|
**Special thanks to**: Aemon Algiz. |
|
|
|
**Patreon special mentions**: Pierre Kircher, Stanislav Ovsiannikov, Michael Levine, Eugene Pentland, Andrey, μ€κ΅ κΉ, Randy H, Fred von Graf, Artur Olbinski, Caitlyn Gatomon, terasurfer, Jeff Scroggin, James Bentley, Vadim, Gabriel Puliatti, Harry Royden McLaughlin, Sean Connelly, Dan Guido, Edmond Seymore, Alicia Loh, subjectnull, AzureBlack, Manuel Alberto Morcote, Thomas Belote, Lone Striker, Chris Smitley, Vitor Caleffi, Johann-Peter Hartmann, Clay Pascal, biorpg, Brandon Frisco, sidney chen, transmissions 11, Pedro Madruga, jinyuan sun, Ajan Kanaga, Emad Mostaque, Trenton Dambrowitz, Jonathan Leane, Iucharbius, usrbinkat, vamX, George Stoitzev, Luke Pendergrass, theTransient, Olakabola, Swaroop Kallakuri, Cap'n Zoog, Brandon Phillips, Michael Dempsey, Nikolai Manek, danny, Matthew Berman, Gabriel Tamborski, alfie_i, Raymond Fosdick, Tom X Nguyen, Raven Klaugh, LangChain4j, Magnesian, Illia Dulskyi, David Ziegler, Mano Prime, Luis Javier Navarrete Lozano, Erik BjΓ€reholt, ιΏζ, Nathan Dryer, Alex, Rainer Wilmers, zynix, TL, Joseph William Delisle, John Villwock, Nathan LeClaire, Willem Michiel, Joguhyik, GodLy, OG, Alps Aficionado, Jeffrey Morgan, ReadyPlayerEmma, Tiffany J. Kim, Sebastain Graf, Spencer Kim, Michael Davis, webtim, Talal Aujan, knownsqashed, John Detwiler, Imad Khwaja, Deo Leter, Jerry Meng, Elijah Stavena, Rooh Singh, Pieter, SuperWojo, Alexandros Triantafyllidis, Stephen Murray, Ai Maven, ya boyyy, Enrico Ros, Ken Nordquist, Deep Realms, Nicholas, Spiking Neurons AB, Elle, Will Dee, Jack West, RoA, Luke @flexchar, Viktor Bowallius, Derek Yates, Subspace Studios, jjj, Toran Billups, Asp the Wyvern, Fen Risland, Ilya, NimbleBox.ai, Chadd, Nitin Borwankar, Emre, Mandus, Leonard Tan, Kalila, K, Trailburnt, S_X, Cory Kujawski |
|
|
|
|
|
Thank you to all my generous patrons and donaters! |
|
|
|
And thank you again to a16z for their generous grant. |
|
|
|
<!-- footer end --> |
|
|
|
# Original model card: Bud's GenZ 13B v2 |
|
|
|
--- |
|
|
|
<div align="center"><h1 align="center">~ GenZ ~</h1><img src="https://raw.githubusercontent.com/BudEcosystem/GenZ/main/assets/genz-logo.png" width=150></div> |
|
|
|
|
|
<p align="center"><i>Democratizing access to LLMs for the open-source community.<br>Let's advance AI, together. </i></p> |
|
|
|
--- |
|
|
|
|
|
## Introduction π |
|
|
|
Welcome to **GenZ**, an advanced Large Language Model (LLM) fine-tuned on the foundation of Meta's open-source Llama V2 13B parameter model. At Bud Ecosystem, we believe in the power of open-source collaboration to drive the advancement of technology at an accelerated pace. Our vision is to democratize access to fine-tuned LLMs, and to that end, we will be releasing a series of models across different parameter counts (7B, 13B, and 70B) and quantizations (32-bit and 4-bit) for the open-source community to use, enhance, and build upon. |
|
|
|
<p align="center"><img src="https://raw.githubusercontent.com/BudEcosystem/GenZ/main/assets/MTBench_CompareChart_28July2023.png" width="500"></p> |
|
|
|
The smaller quantization version of our models makes them more accessible, enabling their use even on personal computers. This opens up a world of possibilities for developers, researchers, and enthusiasts to experiment with these models and contribute to the collective advancement of language model technology. |
|
|
|
GenZ isn't just a powerful text generatorβit's a sophisticated AI assistant, capable of understanding and responding to user prompts with high-quality responses. We've taken the robust capabilities of Llama V2 and fine-tuned them to offer a more user-focused experience. Whether you're seeking informative responses or engaging interactions, GenZ is designed to deliver. |
|
|
|
And this isn't the end. It's just the beginning of a journey towards creating more advanced, more efficient, and more accessible language models. We invite you to join us on this exciting journey. π |
|
|
|
--- |
|
|
|
<h2>Milestone Releases οΈπ</h2> |
|
|
|
**[27 July 2023]** |
|
[_GenZ-13B V2 (ggml)_](https://huggingface.co/budecosystem/genz-13b-v2-ggml) : Announcing our GenZ-13B v2 with ggml. This variant of GenZ can run inferencing using only CPU and without the need of GPU. Download the model from [HuggingFace](https://huggingface.co/budecosystem/genz-13b-v2-ggml). |
|
|
|
**[27 July 2023]** |
|
[_GenZ-13B V2 (4-bit)_](https://huggingface.co/budecosystem/genz-13b-v2-4bit) : Announcing our GenZ-13B v2 with 4-bit quantisation. Enabling inferencing with much lesser GPU memory than the 32-bit variant. Download the model from [HuggingFace](https://huggingface.co/budecosystem/genz-13b-v2-4bit). |
|
|
|
**[26 July 2023]** |
|
[_GenZ-13B V2_](https://huggingface.co/budecosystem/genz-13b-v2) : We're excited to announce the release of our Genz 13B v2 model, a step forward with improved evaluation results compared to v1. Experience the advancements by downloading the model from [HuggingFace](https://huggingface.co/budecosystem/genz-13b-v2). |
|
|
|
**[20 July 2023]** |
|
[_GenZ-13B_](https://huggingface.co/budecosystem/genz-13b) : We marked an important milestone with the release of the Genz 13B model. The journey began here, and you can partake in it by downloading the model from [Hugging Face](https://huggingface.co/budecosystem/genz-13b). |
|
|
|
--- |
|
|
|
|
|
<img src="https://raw.githubusercontent.com/BudEcosystem/GenZ/main/assets/screenshot_genz13bv2.png" width="100%"> |
|
|
|
|  |  |  |
|
|:--:|:--:|:--:| |
|
| *Code Generation* | *Poem Generation* | *Email Generation* | |
|
|
|
<!-- |
|
<p align="center"><img src="https://raw.githubusercontent.com/adrot-dev/git-test/blob/main/assets/Python.gif" width="33%" alt="Python Code"><img src="https://raw.githubusercontent.com/adrot-dev/git-test/blob/main/assets/Poem.gif" width="33%"><img src="https://raw.githubusercontent.com/adrot-dev/git-test/blob/main/assets/Email.gif" width="33%"></p> |
|
--> |
|
|
|
|
|
<h2>Getting Started on Hugging Face π€</h2> |
|
|
|
Getting up and running with our models on Hugging Face is a breeze. Follow these steps: |
|
|
|
<h3>1οΈβ£ : Import necessary modules</h3> |
|
|
|
|
|
Start by importing the necessary modules from the βtransformersβ library and βtorchβ. |
|
|
|
```python |
|
import torch |
|
from transformers import AutoTokenizer, AutoModelForCausalLM |
|
``` |
|
|
|
<h3>2οΈβ£ : Load the tokenizer and the model</h3> |
|
|
|
Next, load up the tokenizer and the model for βbudecosystem/genz-13b-v2β from Hugging Face using the βfrom_pretrainedβ method. |
|
|
|
```python |
|
tokenizer = AutoTokenizer.from_pretrained("budecosystem/genz-13b-v2", trust_remote_code=True) |
|
model = AutoModelForCausalLM.from_pretrained("budecosystem/genz-13b-v2", torch_dtype=torch.bfloat16) |
|
``` |
|
|
|
<h3>3οΈβ£ : Generate responses</h3> |
|
|
|
|
|
Now that you have the model and tokenizer, you're ready to generate responses. Here's how you can do it: |
|
|
|
```python |
|
inputs = tokenizer("The meaning of life is", return_tensors="pt") |
|
sample = model.generate(**inputs, max_length=128) |
|
print(tokenizer.decode(sample[0])) |
|
``` |
|
|
|
In this example, "The meaning of life is" is the prompt template used for inference. You can replace it with any string you like. |
|
|
|
Want to interact with the model in a more intuitive way? We have a Gradio interface set up for that. Head over to our GitHub page, clone the repository, and run the βgenerate.pyβ script to try it out. Happy experimenting! π |
|
|
|
<h2>Fine-tuning π―</h2> |
|
|
|
|
|
It's time to upgrade the model by fine-tuning the model. You can do this using our provided finetune.py script. Here's an example command: |
|
|
|
```bash |
|
python finetune.py \ |
|
--model_name meta-llama/Llama-2-13b \ |
|
--data_path dataset.json \ |
|
--output_dir output \ |
|
--trust_remote_code \ |
|
--prompt_column instruction \ |
|
--response_column output \ |
|
--pad_token_id 50256 |
|
``` |
|
|
|
--- |
|
|
|
<h2 >Bonus: Colab Notebooks π <b><i>(WIP)</i></b></h2> |
|
|
|
|
|
Looking for an even simpler way to get started with GenZ? We've got you covered. We've prepared a pair of detailed Colab notebooks - one for Inference and one for Fine-tuning. These notebooks come pre-filled with all the information and code you'll need. All you'll have to do is run them! |
|
|
|
Keep an eye out for these notebooks. They'll be added to the repository soon! |
|
|
|
--- |
|
|
|
<h2>Why Use GenZ? π‘</h2> |
|
|
|
|
|
You might be wondering, "Why should I choose GenZ over a pretrained model?" The answer lies in the extra mile we've gone to fine-tune our models. |
|
|
|
While pretrained models are undeniably powerful, GenZ brings something extra to the table. We've fine-tuned it with curated datasets, which means it has additional skills and capabilities beyond what a pretrained model can offer. Whether you need it for a simple task or a complex project, GenZ is up for the challenge. |
|
|
|
What's more, we are committed to continuously enhancing GenZ. We believe in the power of constant learning and improvement. That's why we'll be regularly fine-tuning our models with various curated datasets to make them even better. Our goal is to reach the state of the art and beyond - and we're committed to staying the course until we get there. |
|
|
|
But don't just take our word for it. We've provided detailed evaluations and performance details in a later section, so you can see the difference for yourself. |
|
|
|
Choose GenZ and join us on this journey. Together, we can push the boundaries of what's possible with large language models. |
|
|
|
--- |
|
|
|
<h2>Model Card for GenZ 13B π</h2> |
|
|
|
Here's a quick overview of everything you need to know about GenZ 13B. |
|
|
|
<h3>Model Details:</h3> |
|
|
|
|
|
- Developed by: Bud Ecosystem |
|
- Base pretrained model type: Llama V2 13B |
|
- Model Architecture: GenZ 13B, fine-tuned on Llama V2 13B, is an auto-regressive language model that employs an optimized transformer architecture. The fine-tuning process for GenZ 13B leveraged Supervised Fine-Tuning (SFT) |
|
- License: The model is available for commercial use under a custom commercial license. For more information, please visit: [Meta AI Model and Library Downloads](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) |
|
|
|
--- |
|
|
|
<h2>Intended Use πΌ</h2> |
|
|
|
When we created GenZ 13B, we had a clear vision of how it could be used to push the boundaries of what's possible with large language models. We also understand the importance of using such models responsibly. Here's a brief overview of the intended and out-of-scope uses for GenZ 13B. |
|
|
|
<h3>Direct Use</h3> |
|
|
|
GenZ 13B is designed to be a powerful tool for research on large language models. It's also an excellent foundation for further specialization and fine-tuning for specific use cases, such as: |
|
- Text summarization |
|
- Text generation |
|
- Chatbot creation |
|
- And much more! |
|
|
|
<h3>Out-of-Scope Use π©</h3> |
|
|
|
While GenZ 13B is versatile, there are certain uses that are out of scope: |
|
|
|
- Production use without adequate assessment of risks and mitigation |
|
- Any use cases which may be considered irresponsible or harmful |
|
- Use in any manner that violates applicable laws or regulations, including trade compliance laws |
|
- Use in any other way that is prohibited by the Acceptable Use Policy and Licensing Agreement for Llama 2 |
|
|
|
Remember, GenZ 13B, like any large language model, is trained on a large-scale corpora representative of the web, and therefore, may carry the stereotypes and biases commonly encountered online. |
|
|
|
<h3>Recommendations π§ </h3> |
|
|
|
We recommend users of GenZ 13B to consider fine-tuning it for the specific set of tasks of interest. Appropriate precautions and guardrails should be taken for any production use. Using GenZ 13B responsibly is key to unlocking its full potential while maintaining a safe and respectful environment. |
|
|
|
--- |
|
|
|
<h2>Training Details π</h2> |
|
|
|
When fine-tuning GenZ 13B, we took a meticulous approach to ensure we were building on the solid base of the pretrained Llama V2 13B model in the most effective way. Here's a look at the key details of our training process: |
|
|
|
<h3>Fine-Tuning Training Data</h3> |
|
|
|
For the fine-tuning process, we used a carefully curated mix of datasets. These included data from OpenAssistant, an instruction fine-tuning dataset, and Thought Source for the Chain Of Thought (CoT) approach. This diverse mix of data sources helped us enhance the model's capabilities across a range of tasks. |
|
|
|
<h3>Fine-Tuning Procedure</h3> |
|
|
|
We performed a full-parameter fine-tuning using Supervised Fine-Tuning (SFT). This was carried out on 4 A100 80GB GPUs, and the process took under 100 hours. To make the process more efficient, we used DeepSpeed's ZeRO-3 optimization. |
|
|
|
<h3>Tokenizer</h3> |
|
|
|
We used the SentencePiece tokenizer during the fine-tuning process. This tokenizer is known for its capability to handle open-vocabulary language tasks efficiently. |
|
|
|
|
|
<h3>Hyperparameters</h3> |
|
|
|
|
|
Here are the hyperparameters we used for fine-tuning: |
|
|
|
| Hyperparameter | Value | |
|
| -------------- | ----- | |
|
| Warmup Ratio | 0.04 | |
|
| Learning Rate Scheduler Type | Cosine | |
|
| Learning Rate | 2e-5 | |
|
| Number of Training Epochs | 3 | |
|
| Per Device Training Batch Size | 4 | |
|
| Gradient Accumulation Steps | 4 | |
|
| Precision | FP16 | |
|
| Optimizer | AdamW | |
|
|
|
--- |
|
|
|
<h2>Evaluations π―</h2> |
|
|
|
Evaluating our model is a key part of our fine-tuning process. It helps us understand how our model is performing and how it stacks up against other models. Here's a look at some of the key evaluations for GenZ 13B: |
|
|
|
<h3>Benchmark Comparison</h3> |
|
|
|
We've compared GenZ V1 with V2 to understand the improvements our fine-tuning has achieved. |
|
|
|
| Model Name | MT Bench | Vicuna Bench | MMLU | Human Eval | Hellaswag | BBH | |
|
|:----------:|:--------:|:------------:|:----:|:----------:|:---------:|:----:| |
|
| Genz 13B | 6.12 | 86.1 | 53.62| 17.68 | 77.38 | 37.76| |
|
| Genz 13B v2| 6.79 | 87.2 | 53.68| 21.95 | 77.48 | 38.1 | |
|
|
|
<h3>MT Bench Score</h3> |
|
|
|
A key evaluation metric we use is the MT Bench score. This score provides a comprehensive assessment of our model's performance across a range of tasks. |
|
|
|
We're proud to say that our model performs at a level that's close to the Llama-70B-chat model on the MT Bench and top of the list among 13B models. |
|
|
|
<p align="center"><img src="https://raw.githubusercontent.com/BudEcosystem/GenZ/main/assets/mt_bench_score.png" width="500"></p> |
|
|
|
In the transition from GenZ V1 to V2, we noticed some fascinating performance shifts. While we saw a slight dip in coding performance, two other areas, Roleplay and Math, saw noticeable improvements. |
|
|
|
--- |
|
|
|
<h2>Looking Ahead π</h2> |
|
|
|
We're excited about the journey ahead with GenZ. We're committed to continuously improving and enhancing our models, and we're excited to see what the open-source community will build with them. We believe in the power of collaboration, and we can't wait to see what we can achieve together. |
|
|
|
Remember, we're just getting started. This is just the beginning of a journey that we believe will revolutionize the world of large language models. We invite you to join us on this exciting journey. Together, we can push the boundaries of what's possible with AI. π |
|
|
|
--- |
|
|
|
|
|
Check the GitHub for the code -> [GenZ](https://raw.githubusercontent.com/BudEcosystem/GenZ) |
|
|
