92 lines
4.6 KiB
Markdown
92 lines
4.6 KiB
Markdown
---
|
||
license: mit
|
||
language:
|
||
- de
|
||
- en
|
||
---
|
||
|
||
## VAGO solutions SauerkrautLM-Phi-3-medium
|
||
Introducing **SauerkrautLM-Phi-3-medium** – our Sauerkraut version of the powerful [unsloth/Phi-3-medium-4k-instruct](https://huggingface.co/unsloth/Phi-3-medium-4k-instruct)!
|
||
|
||
|
||
- Aligned with DPO using [**Spectrum**](https://github.com/cognitivecomputations/spectrum) QLoRA (by Eric Hartford, Lucas Atkins, Fernando Fernandes Neto and David Golchinfar) **targeting 50% of the layers.**
|
||
|
||
# Table of Contents
|
||
1. [Overview of all SauerkrautLM-Phi-3-medium](#all-SauerkrautLM-Phi-3-medium)
|
||
2. [Model Details](#model-details)
|
||
- [Training procedure](#training-procedure)
|
||
3. [Evaluation](#evaluation)
|
||
5. [Disclaimer](#disclaimer)
|
||
6. [Contact](#contact)
|
||
7. [Collaborations](#collaborations)
|
||
8. [Acknowledgement](#acknowledgement)
|
||
|
||
|
||
## All SauerkrautLM-Phi-3-medium
|
||
|
||
| Model | HF | EXL2 | GGUF | AWQ |
|
||
|-------|-------|-------|-------|-------|
|
||
| SauerkrautLM-Phi-3-medium | [Link](https://huggingface.co/VAGOsolutions/SauerkrautLM-Phi-3-medium) | coming soon | coming soon | coming soon |
|
||
|
||
## Model Details
|
||
**SauerkrautLM-Phi-3-medium**
|
||
- **Model Type:** SauerkrautLM-Phi-3-medium is a finetuned Model based on [unsloth/Phi-3-medium-4k-instruct](https://huggingface.co/unsloth/Phi-3-medium-4k-instruct)
|
||
- **Language(s):** German, English
|
||
- **License:** MIT
|
||
- **Contact:** [VAGO solutions](https://vago-solutions.ai)
|
||
|
||
### Training procedure:
|
||
- We trained this model with [**Spectrum**](https://github.com/cognitivecomputations/spectrum) QLoRA DPO Fine-Tuning for 1 epoch with 70k samples targeting 50% of the layers with a high Learningrate of 5e-04.
|
||
This relatively high learning rate was feasible due to the selective targeting of layers; had we applied this rate to all layers, the gradients would have exploded.
|
||
|
||
**Fine-Tuning Details**
|
||
|
||
Epochs: 1
|
||
Data Size: 70,000 samples
|
||
Targeted Layers: 50%
|
||
Learning Rate: 5e-04
|
||
Warm-up Ratio: 0.03
|
||
The strategy of targeting only half of the layers also enabled us to use a very low warm-up ratio of 0.03, contributing to the overall stability of the fine-tuning process.
|
||
|
||
|
||
**Results**
|
||
|
||
This fine-tuning approach resulted in a noticeable improvement in the model's reasoning capabilities.
|
||
The model's performance was evaluated using a variety of benchmark suites, including the newly introduced [MixEval](https://mixeval.github.io/), which shows a 96% correlation with Chatbot Arena.
|
||
MixEval uses regular updated test data, providing a reliable benchmark for model performance.
|
||
|
||
|
||
|
||
|
||
## Evaluation
|
||
|
||
**Open LLM Leaderboard and German RAG:**
|
||
|
||
|
||
|
||
|
||
**Mix Eval Hard**
|
||
|
||
|
||
|
||
**GPT4ALL**
|
||
|
||
|
||
|
||
**AGIEval**
|
||
|
||
|
||
|
||
## 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 websites. 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](https://vago-solutions.ai/#Kontakt)
|
||
|
||
## Acknowledgement
|
||
Many thanks to [unsloth](https://huggingface.co/unsloth/) and [Microsoft](https://huggingface.co/microsoft) for providing such valuable model to the Open-Source community. |