Google Gemini 1.5 Vibe Check
Today, during the I/O event, Google announced that they are updating Gemini 1.5 Pro, introducing 1.5 Flash, rolling out new developer features and adding two new Gemma models. 1
So I tried Gemini as my Pixel Assistant today. I migrated from Google Assistant to Gemini on Android.
As I’m currently fine-tuning large language models, I need to compare fine-tuning Transformer tools. So I asked Gemini. Here’s the link to the entire chat below:
Prompt:
You are an large language models expert. You have trained deep learning Transformer models. I want to fine-tune a base model with my domain-specific dataset.
What do you think about Axolotl vs Unsloth vs TorchTune?
Output:
I can help you with comparing Axolotl, Unsloth, and TorchTune for fine-tuning a base model with your domain-specific dataset. Here's a brief comparison:
- Axolotl: A lightweight library designed specifically for fine-tuning transformers on small datasets. It excels in tasks where efficiency and ease of use are priorities.
- Unsloth: A comprehensive library that offers various functionalities for fine-tuning transformers, including hyperparameter tuning, regularization techniques, and integration with other libraries. It caters to more complex fine-tuning scenarios.
- TorchTune: A powerful framework built on PyTorch for training and fine-tuning deep learning models, including transformers. It provides a high degree of flexibility and customization but requires more in-depth PyTorch knowledge.
Choosing the best option depends on your specific needs:
- For smaller datasets and ease of use, Axolotl is a strong choice.
- For more complex fine-tuning with additional functionalities, consider Unsloth.
- If you have experience with PyTorch and need maximum flexibility, TorchTune is an option.
Prompt:
What and why Axolotl win the ease of use?
Output:
Axolotl wins in ease of use due to its streamlined design focused on fine-tuning transformers with minimal configuration. Here's a breakdown of its advantages:
- Simpler API: Axolotl offers a concise API that simplifies the fine-tuning process, requiring less code compared to other libraries.
- Automatic Configuration: Axolotl handles many configurations automatically, reducing the need for manual parameter tuning, which can be error-prone and time-consuming.
- Focus on Small Datasets: Axolotl is optimized for smaller datasets, common in domain-specific fine-tuning, making it easier to achieve good results without extensive hyperparameter tuning.
A side-by-side comparison:
| Axolotl | Unsloth | TorchTune | |
|---|---|---|---|
| Strengths | Smaller datasets and ease of use. Automatic configuration. It just works, because community-maintained best practices implementation. | More complex fine-tuning with additional functionalities. Finetune Llama 3, Mistral & Gemma 2-5x faster with 80% less memory but upsell to paid Unsloth Pro which is 30x faster. | If you have experience with PyTorch and need maximum flexibility. Low dependencies. |
| Opinions | Support more models & techniques. | Prefer to support popular models such as Llama, Mistral | |
| License | Open source: Apache 2.0 | Open source: Apache 2.0 | Open source: BSD-3 |
The latest Gemini 1.5 Pro model is not bad anymore. Interesting. I plan to test it further.