Fine tuning contd
- For finetuning we will be using unsloth
- unsloth has models which we will be using to tune
- unsloth has decent documentation and notebooks shared.
- unsloth can be used in two ways
- on your vm (system) with GPU
- Google Colab (GPUs for free)
- Goal:
- to digest steps
Dataset format
| Format | Use Case | Native Support |
| ——————- | ——————– | —————— |
| Alpaca | Instruction tuning | |
| ChatML | Chat models | |
| ShareGPT | Public chat data | |
| Plain Text | Completion tasks | Manual template |
| JSONL | Large-scale training | |
| Custom Prompt | Advanced control | |
| HuggingFace Dataset | Any dataset | |
{
"instruction": "What is Kubernetes?",
"input": "",
"output": "Kubernetes is a container orchestration platform..."
}
{"instruction":"Explain TCP","output":"TCP is a transport protocol"}
{"instruction":"Explain UDP","output":"UDP is a connectionless protocol"}
Hyperparameters
Steps
