ohara 0.1.2

A collection of implementations of LLM, papers, and other models.

Ohara

This is my collection of implimention of llm,paper and things I hand in my mind I hand lot of fragmented code of implimention of diffrent model

this is attempt to make it eveything in one place
This lib is for runing/copying code for expriments
install with

git clone https://github.com/joey00072/ohara.git
pip install -e .

then

## download and pretokenize
python examples/prepare-dataset.py

## train 
## look at train.py its fairly easy
python examples/train_llama.py

# lighting fabric verison is also avalible (recommanded)
python examples/train_llama_fabric.py 

llama-20M trained on tinystores for 1.7B

inferance on phi2

## this will download model from hf and run it in torch.flaot16

python phi_inference.py 

## look at files and you can impliment rest of things easily, 
## I belive in you 😉

The lib to maximize FAFO

papaers and theory is on one side but code is truth, in the end things that matter that works (runs)
If you look into docs you can find some written things. this are mostly copied from my obsidian notes

WORK IS PROGESS (always)

papers

1. SwiGLU | md
2. Alibi Embeddings | md
3. Rotary Embeddings | md
4. LoRA
5. LLAMA | md
6. XPOS
7. GPT | md
8. Mamba
9. renet

More things are not in this repo

1. TinyLora
2. Neural Style Transfer in Pytorch

TODO (A lot)

• make infercer class better
• make training loop better (use lightning fabric maybe)
• Finetuning in structed way (I just rawdoag code when I need it)
• DPO
• make is py modele so I can create expriment folder and put all this in it

Fund My Caffeine Addiction

contribution guidelines

• be nice,
• code explaintions || docs are appricated
• memes on pr recommend