LLAMA-FIT CO.โข - Your LLaMA Will Like The Way It Looks
-
Goal is to finetune a LLaMA2 so that it can impress during a Tech Interview
-
business model
- I got suited up llama interns waiting to work for you
-
operation model
- read the doc (TODO: make a doc); the joke here is that it's basically the code and I have no moat
-
street cred maxxing
-
impress employers/vc/people/ai
-
benchmark llama2 fine-tune against base model
- split training set against validation set (90/10?)
- benchmark fine-tune against base
- goal: achieve higher "cosine similarity" (between llama output and validation set) using fine-tuned version than base version
- (exclaim - oh shit! - I'm a regmonkey)
- run all cells in notebook
- last cell can be ran to view simple haphazard benchmark
- LLaMA2 Paper
- definitive primary source
- LLaMA2 Github
- actual code
- not to be confused with llamacode library
- (aside) spent way to much time trying to figure out why the download code wasn't working ...
- 7b model
- Deepgram Video Analysis
- good take aimed for layman
- Karparthy is a Beast
- look it's clear that I'm not legit until I make it onto the README.md page
- but as a weekend project ... nah ... (at least right now)
- Fiverr
- there's academic integrity which I adhere to
- but I'm also technically an entrepreneur ... had the look
- these rates seems decent
- side hustle oppurtunity?
- Random Paper
- literally a random paper I pulled from arxiv talking about training and finetunning
- insight...
- AlpacaFarm
- is this useful?
- Yannic Kilcher segment
- Original LLaMA
- LLaMA-Accessory (potential fine-tune tool)
- together.ai, openchat, lmsys.org (tools that leverage llama)
- kaggle always have some interesting datasets (interview quetions related)
- able to manually compile a list of interesting interview questions
- potential OCR (future... TODO)
- use GPT4 API to synthetically generate responses
- potential human reinforcement here
- scale.ai/mechanical turk stuff? (TODO)
- potential human reinforcement here
- acquire datasets from kaggle
- random code in data_processing
- Use GTP4 to do data formatting work
- create .env with OPENAI_API_KEY=...
- alternatively, upload/pull datasets into/from huggingface
Code to be executed are found in Notebooks, which contains a bunch of colab notebooks that should be one-click solutions
- good no fluff script
- copy of code to use in pipeline in Notebooks
- this looks like an out of the box solution - one click colab solution
- cons is that it uses some strange sharded model of llama 7b
LlamaIndex Semantic Similarity Evaluator
- compare text
- basis for more rigorous benchmarking
- eval-2-base_llama vs eval-2-llama_finetune
MMLU (Massive Multitask Language Understanding)
- left as an excersize to better understand benchmarking
- choose/create fine-tuning dataset
- prepare dataset
- choose fine-tuning framework
- configure fine-tuning process
- train the model
- evaluate the model
Collab Notesbooks in Repo
- Yo, all instances are reserved
- WTF
- need to containerize application so that it can run on any compute cloud/otherwise/etc.
- this is the actual product
- buried Alpha if you read this far