Looking for easy ways to do fine-tuning of Generative AI models.
These are the options for (text only?) LLMs:
| Aspect / Option | Axolotl | Unsloth | TorchTune |
|---|---|---|---|
| Link | https://github.com/OpenAccess-AI-Collective/axolotl | https://github.com/unslothai/unsloth | https://github.com/pytorch/torchtune |
| License | Open source: Apache 2.0 | Open source: Apache 2.0 | Open source: BSD-3 |
| USP | Just works, because community-maintained best practices implementation | Finetune Llama 3, Mistral & Gemma 2-5x faster with 80% less memoryBut upsell to paid Unsloth Pro which is 30x faster | From Meta + Native PyTorch + no other dependencies (no trainers, no frameworks) |
| See design principles | |||
| Thoughts 🚧 | More number of models & techniques supported? | Prefer for popular models such as Llama, Mistral, Gemma? | Prefer for Llama 3? |
| Prefer for full customization? |
Also see https://wandb.ai/augmxnt/train-bench/reports/torchtune-vs-axolotl-vs-unsloth-Trainer-Performance-Comparison--Vmlldzo4MzU3NTAx
<aside> ⚠️
Newbie here, looking for feedback on how to think about this!
(yes, I’m invoking Cunningham’s Law)
</aside>
How to use these?