Hugging Face's logo

pL-Community
/
Llama-3.1-SauerkrautLM-70b-Instruct-FP8-Dynamic

Safetensors
llama
spectrum
Model card Files Community
1
Llama-3.1-SauerkrautLM-70b-Instruct-FP8-Dynamic / README.md
flozi00's picture
flozi00
Update config.json (#1)
c94e85d verified
preview code
|
raw
history blame contribute delete
7.06 kB
metadata
license: llama3.1
language:
  - de
  - en
  - it
  - fr
  - pt
  - es
  - ar
  - nl
tags:
  - spectrum

Llama-3.1-SauerkrautLM-70b-Instruct

VAGO solutions Llama-3.1-SauerkrautLM-70b-Instruct quantized by Florian Zimmermeister for fp8 usage

Fine-tuned Model - to showcase the potential of resource-efficient Fine-Tuning of Large Language Models using Spectrum Fine-Tuning

Introducing Llama-3.1-SauerkrautLM-70b-Instruct – our Sauerkraut version of the powerful meta-llama/Meta-Llama-3.1-70B-Instruct!

  • Fine-tuning on German-English data with Spectrum Fine-Tuning targeting 15% of the layers.
  • Utilized unique German-English Sauerkraut Mix v2 dataset for efficient cross-lingual transfer learning
  • Implemented bespoke, precision-engineered fine-tuning approach to enhance multilingual capabilities
  • Achieved improved performance in multiple languages (including Arabic, Italian, French, Spanish, Dutch, Portuguese) through cross-lingual knowledge transfer

Table of Contents

  1. Overview of all Llama-3.1-SauerkrautLM-70b-Instruct
  2. Model Details
  3. Evaluation
  4. Disclaimer
  5. Contact
  6. Collaborations
  7. Acknowledgement

All Llama-3.1-SauerkrautLM-70b-Instruct

Model HF EXL2 GGUF AWQ
Llama-3.1-SauerkrautLM-70b-Instruct Link coming soon coming soon coming soon

Model Details

Llama-3.1-SauerkrautLM-70b-Instruct

Training Procedure

This model showcases the potential of resource-efficient fine-tuning of large language models using Spectrum Fine-Tuning. Here's a brief on the procedure:

Fine-tuning on German-English Data:

  • Utilized Spectrum Fine-Tuning, targeting 15% of the model's layers
  • Introduced the model to a unique German-English Sauerkraut Mix v2
  • Implemented a bespoke, precision-engineered fine-tuning approach

Cross-lingual Transfer Learning using Sauerkraut Mix v2:

  • Leveraged the Sauerkraut Mix v2 dataset as the foundation for cross-lingual transfer
  • This unique dataset, primarily focused on German and English, enabled the model to transfer knowledge to other languages
  • Improved capabilities in Arabic, Italian, French, Spanish, Dutch, and Portuguese without extensive training data in each language
  • Demonstrated the effectiveness of using a bilingual dataset for multilingual improvement

Sauerkraut Mix v2:

  • Premium Dataset for Language Models, focusing on German and English
  • Meticulously selected, high-quality dataset combinations
  • Cutting-edge synthetic datasets created using proprietary, high-precision generation techniques
  • Serves as the core resource for both fine-tuning and cross-lingual transfer

Objective and Results

The primary goal of this training was twofold:

  1. To demonstrate that Spectrum Fine-Tuning, targeting just 15% of the layers, can significantly enhance a 70 billion parameter model's capabilities while using only a fraction of the resources required by classic fine-tuning approaches.

  2. To showcase the effectiveness of cross-lingual transfer learning using the Sauerkraut Mix v2 dataset, enabling multilingual improvement without extensive language-specific training data.

The results have been remarkable:

  • The model has substantially improved its multilingual skills, as demonstrated by impressive benchmarks on MMLU Multilingual.

Key Findings:

  • Spectrum Fine-Tuning can efficiently enhance a large language model's capabilities in multiple languages while preserving the majority of its previously acquired knowledge.
  • The Sauerkraut Mix v2 dataset proves to be an effective foundation for cross-lingual transfer, allowing for multilingual improvements from a bilingual base.
  • This approach demonstrates a resource-efficient method for creating powerful multilingual models without the need for extensive training data in each target language.

Evaluation

AGIEVAL Llama-3.1-SauerkrautLM-70b-Instruct-AGIEVAL

GPT4ALL Llama-3.1-SauerkrautLM-70b-Instruct-GPT4ALL

TRUTHFULQA Llama-3.1-SauerkrautLM-70b-Instruct-TRUTHFULQA

BBH-HF Llama-3.1-SauerkrautLM-70b-Instruct-bbh

MMLU-Multilingual Llama-3.1-SauerkrautLM-70b-Instruct-mmlu

Disclaimer

We must inform users that despite our best efforts in data cleansing, the possibility of uncensored content slipping through cannot be entirely ruled out. However, we cannot guarantee consistently appropriate behavior. Therefore, if you encounter any issues or come across inappropriate content, we kindly request that you inform us through the contact information provided. Additionally, it is essential to understand that the licensing of these models does not constitute legal advice. We are not held responsible for the actions of third parties who utilize our models.

Contact

If you are interested in customized LLMs for business applications, please get in contact with us via our website. We are also grateful for your feedback and suggestions.

Collaborations

We are also keenly seeking support and investment for our startup, VAGO solutions where we continuously advance the development of robust language models designed to address a diverse range of purposes and requirements. If the prospect of collaboratively navigating future challenges excites you, we warmly invite you to reach out to us at VAGO solutions

Acknowledgement

Many thanks to meta-llama for providing such a valuable model to the Open-Source community.