stanford_alpaca VS llama

Alpaca is an instruction-oriented LLM derived from LLaMA, enhanced by Stanford researchers with a dataset of 52,000 examples of following instructions, sourced from OpenAI’s InstructGPT through the self-instruct method. The extensive self-instruct dataset, details of data generation, and the model refinement code were publicly disclosed. This model complies with the licensing requirements of its base model. Due to the utilization of InstructGPT for data generation, it also adheres to OpenAI’s usage terms, which prohibit the creation of models competing with OpenAI. This illustrates how dataset restrictions can indirectly affect the resulting fine-tuned model.

In this article, we're going to experiment with LoRA and fine-tune Llama Alpaca using commercial hardware.

Most papers I did read showed at least a thousand, even 10000 at several cases, so I assumed that to be the trend in the case of Low rank adapter(PEFT) training.(source: [2305.14314] QLoRA: Efficient Finetuning of Quantized LLMs (arxiv.org) , Stanford CRFM (Alpaca) and the minimum being openchat/openchat · Hugging Face ; There are a lot more examples)

I'm not talking about OpenAI ChatGPT I'm talking about things ALPACA, and where did they train these models? Off the existing models for a fraction of a fraction of a fraction of the cost: https://crfm.stanford.edu/2023/03/13/alpaca.html

I think there's a new approach for “How do you get the data?” that wasn't available when this article was written in 2015. The new text and image generative models can now be used to synthesize training datasets.

I was working on an typing autocorrect project and needed a corpus of "text messages". Most of the traditional NLP corpuses like those available through NLTK [0] aren't suitable. But it was easy to script ChatGPT to generate thousands of believable text messages by throwing random topics at it.

Similarly, you can synthesize a training dataset by giving GPT the outputs/labels and asking it to generate a variety of inputs. For sentiment analysis... "Give me 1000 negative movie reviews" and "Now give me 1000 positive movie reviews".

The Alpaca folks used GPT-3 to generate high-quality instruction-following datasets [1] based on a small set of human samples.

Etc.

[0] https://www.nltk.org/nltk_data/

[1] https://crfm.stanford.edu/2023/03/13/alpaca.html

AFAIK, the original alpaca repo was a full finetune. https://github.com/tatsu-lab/stanford_alpaca

One thing I wanted to add and call attention to is the importance of licensing in open models. This is often overlooked when we blindly accept the vague branding of models as “open”, but I am noticing that many open weight models are actually using encumbered proprietary licenses rather than standard open source licenses that are OSI approved (https://opensource.org/licenses). As an example, Databricks’s DBRX model has a proprietary license that forces adherence to their highly restrictive Acceptable Use Policy by referencing a live website hosting their AUP (https://github.com/databricks/dbrx/blob/main/LICENSE), which means as they change their AUP, you may be further restricted in the future. Meta’s Llama is similar (https://github.com/meta-llama/llama/blob/main/LICENSE ). I’m not sure who can depend on these models given this flaw.

It looks like Llama 2 7B took 184,320 A100-80GB GPU-hours to train[1]. This one says it used a 96×H100 GPU cluster for 2 weeks, for 32,256 hours. That's 17.5% of the number of hours, but H100s are faster than A100s [2] and FP16/bfloat16 performance is ~3x better.

If they had tried to replicate Llama 2 identically with their hardware setup, it'd cost a little bit less than twice their MoE model.

[1] https://github.com/meta-llama/llama/blob/main/MODEL_CARD.md#...

Ironically, the LLaMA license text [1] this is lifted verbatim from is itself copyrighted [2] and doesn't grant you the permission to copy it or make changes like s/meta/dbrx/g lol.

[1] https://github.com/meta-llama/llama/blob/main/LICENSE#L65

This is kind of an epistemological debate at this level, and I make an effort to link to some source code [1] any time it seems contentious.

LLMs (of the decoder-only, generative-pretrained family everyone means) are next token predictors in a literal implementation sense (there are some caveats around batching and what not, but none that really matter to the philosophy of the thing).

But, they have some emergent behaviors that are a trickier beast. Probably the best way to think about a typical Instruct-inspired “chat bot” session is of them sampling from a distribution with a KL-style adjacency to the training corpus (sidebar: this is why shops that do and don’t train/tune on MMLU get ranked so differently than e.g. the arena rankings) at a response granularity, the same way a diffuser/U-net/de-noising model samples at the image batch (NCHW/NHWC) level.

The corpus is stocked with everything from sci-fi novels with computers arguing their own sentience to tutorials on how to do a tricky anti-derivative step-by-step.

This mental model has adequate explanatory power for anything a public LLM has ever been shown to do, but that only heavily implies it’s what they’re doing.

There is active research into whether there is more going on that is thus far not conclusive to the satisfaction of an unbiased consensus. I personally think that research will eventually show it’s just sampling, but that’s a prediction not consensus science.

They might be doing more, there is some research that represents circumstantial evidence they are doing more.

[1] https://github.com/meta-llama/llama/blob/54c22c0d63a3f3c9e77...

Predicting subsequent text is pretty much exactly what they do. Lots of very cool engineering that’s a real feat, but at its core it’s argmax(P(token|token,corpus)):

https://github.com/facebookresearch/llama/blob/main/llama/ge...

The engineering feats are up there with anything, but it’s a next token predictor.

https://github.com/facebookresearch/llama/pull/947/

> Instead, it turns out a few hundred lines of Python is genuinely enough to train a basic version!

actually its not just a basic version. Llama 1/2's model.py is 500 lines: https://github.com/facebookresearch/llama/blob/main/llama/mo...

Mistral (is rumored to have) forked llama and is 369 lines: https://github.com/mistralai/mistral-src/blob/main/mistral/m...

and both of these are SOTA open source models.

In transformers, they tried really hard to have a single function or method to deal with both self and cross attention mechanisms, masking, positional and relative encodings, interpolation etc. While it allows a user to use the same function/method for any model, it has led to severe parameter bloat. Just compare the original implementation of llama by FAIR with the implementation by HF to get an idea.

(myvenv) root@a176328806c7:/home# apt-get install git (myvenv) root@a176328806c7:/home# apt-get install wget (myvenv) root@a176328806c7:/home# git clone https://github.com/facebookresearch/llama.git (myvenv) root@a176328806c7:/home# cd llama/ (myvenv) root@a176328806c7:/home/llama# pip install -e . (myvenv) root@a176328806c7:/home/llama# ./download.sh