Add LLaMA2 + LoRA Fine-Tuning Documentation for Greene HPC - #94
Conversation
|
| To avoid exceeding home directory quotas during large model downloads: | ||
|
|
||
| ```bash | ||
| export HF_HOME=/scratch/ml7612/.cache/huggingface |
There was a problem hiding this comment.
remove your netid and change it to NetID to be more generic for readers.
|
|
||
| --- | ||
|
|
||
| ## Model and Dataset Selection Rationale |
There was a problem hiding this comment.
Move this section to the top and add a few lines on what the dataset is and how the fine-tuning helps the LLM perform better at a particular task. It would be great to have a response from an LLM before and after fine-tuning to demonstrate the effect of fine-tuning.
There was a problem hiding this comment.
Feel free to switch to a different fine-tuning task it makes the task of demonstrating the benefits of fine-tuning easier.
| /scratch/work/public/singularity/cuda11.2.2-cudnn8-devel-ubuntu20.04.sif \ | ||
| /bin/bash -c " | ||
| source /ext3/miniconda3/bin/activate | ||
| cd /scratch/ml7612/fine-tune |
There was a problem hiding this comment.
Same comment as above on removing personal NetID.
| /bin/bash -c " | ||
| source /ext3/miniconda3/bin/activate | ||
| cd /scratch/ml7612/fine-tune | ||
| python train_phi4.py |
There was a problem hiding this comment.
If train_phi4.py is a small file, include it here. If not link to the repo.
|
Also, add a couple of lines on what LORA is. Someone new to this topic would not know the acronym so spell it out on first use and link to some references like https://huggingface.co/learn/llm-course/en/chapter11/4 |
|
Since the dataset we're fine tuning on is instruction, can we use a model with a known instruction tuned variant to compare against? How about using: |
|
|
||
| **Current Applications**: Drug discovery, financial modeling, cryptography, and artificial intelligence research. | ||
|
|
||
| This comparison demonstrates how our custom fine-tuning can achieve similar or better instruction-following capabilities compared to the official instruction-tuned variant. |
There was a problem hiding this comment.
This line should be outside the output block.
9d44ba8 to
3503412
Compare
2 participants
This PR introduces a complete documentation guide for fine-tuning LLaMA2 using LoRA (PEFT) on NYU Greene HPC. The document includes: