https://huggingface.co/kalomaze/MiniSymposium-Demo
MiniSymposium is an experimental model that I created based on Mistral 7b. I created it attempting to test these goals:
- Demonstrate the untapped potential of using a small, focused dataset of handwritten examples instead of training on a large amount of synthetic GPT outputs, by lowering the learning rate and doing many passes over the small dataset
- Create a dataset that allows the model to explore different possible answers from multiple perspectives before reaching a final conclusion ('Socratic prompting'?)
- Develop a model that performs well across various pseudo-markdown prompt formats, rather than overfitting to a specific kind of format such as ChatML, which should naturally benefit other general purpose use cases
The current trend in QLora/Lora-based finetuning (and finetuning in general for local LLMs) is to use large synthetic datasets. These are typically GPT-generated datasets trained with higher learning rates.
However, I believe there is a lot of potential in using small, hand-written datasets with low learning rates, even if it's for general-purpose instruction following, as long as you train it for many epochs on a learning rate low enough to avoid overfitting.
This approach, I hypothesize, helps the model to learn the deeper patterns of instruction following , including the small details. This should help to avoid shallow data biases (like "As an AI made by OpenAI" and other GPT-isms) that are irrelevant to deeper instruction following patterns, especially in long context and multiturn scenarios.
My initial configuration for this QLora model used a constant learning rate of 1e-6 (0.000001), which resulted in obvious, massive overfitting after about 100 epochs. The model started reproducing the original dataset almost verbatim, and exhibited poor generalization across different prompt formats, including obvious hallucinations & also Chinese language outputs for some reason.
However, turning down the learning rate to 1/10th of (1e-7, which is 0.0000001) significantly improved the model with the same exact small dataset. I trained for about ~10 hours on my RTX 3060 to 600 epochs; I think it's still a little undertrained, but I encourage people to try the demo model out in the meantime.
It's designed to be very adaptable to different prompt formats and playing roles, and I've gotten some fun and sometimes surprisingly good outputs so far.
A few samples of the training data are formatted like this to help avoid blatant overconfidence in its outputs, to serve as a sort of self-correction mechanism:
Let me know how this model goes. There's lots of merges of models that are all sort of doing the same thing, so I figured a more experimental approach would be appreciated. I think there is still more optimization for LR/epoch balance, and I'll probably add some more examples of specific tasks like Summarization in the dataset so that it's not *too* small (but still lightweight enough to generalize well).
Multiple passes at lower learning rates isn't supposed to produce different results.
(Assuming your mini batching etc is all setup correctly) none the less I love exploration and can't wait to learn more, thanks for sharing dude!
> Multiple passes at lower learning rates isn't supposed to produce different results.
Oh, I was wrong on this, then, my bad.
So would my interpretation be correct that this is essentially causing the overfitting to still happen, just significantly slower, and that a higher LR would work? The problem is at first the average loss tanked in the span of like a single epoch to near zero which overfit, but this LR didn't have the same effect.
gpt4 is claiming this comment's claim is wrong, but I can't trust it blindly ofc, i'll look into my initial claim to verify
GPT-4:
The statement contains several inaccuracies:
Overall, while the intention of the statement is to describe overfitting and the effects of learning rates, it conflates different concepts and could benefit from clearer differentiation between them.
I am inclined to believe gpt4 since it consistently claims this across both the API and your comment... but I'm not sure
Much simpler than GPT-4 – the person above seems to be referring to gradient accumulation (since they mentioned minibatches), where you add up gradients until you reach the target batch size, then apply them. This is perfectly equivalent to training on a larger batch.
Actually training on small batches with a low learning rate, however, and applying the gradients immediately, is definitely not equivalent to a bigger batch with a bigger learning rate, especially if you're in a particularly unstable part of parameter space, where large learning rates might overshoot. On the other hand, the tiny batches would tend to make the direction your model moves somewhat random, which might be good, might be bad.
Whether or not this actually does what OP wants it to is really just an empirical question. If they did it, and it worked better than bigger batches with the same data, then I guess it helped (in this case with this model and this data), haha
Good info thanks for that!
I think that's not the whole story. The smaller increments can lead to "course changes" that would not have happened otherwise. Might let things slip into other local minima and all that. It's not just several small steps instead of one big one. The straight line that is the big step will become a curve, capable of bringing you into an entirely different place. The whole dataset can have its impact before some giant leaps jump into a single direction. As a laymen, maybe I've got this wrong, but I really don't see how you can categorically dismiss the possiblity of creating a much more robust and effective architecture instead of essentially jumping to conclusions and then somewhat fixing it up.
That makes sense cheers
Oh yes it is. The whole point of gradient descent is to slowly explore the dimensions of the gradient. With smaller steps you have a totally different trajectory than with bigger steps. And every pass makes you move.
If you choose a too small learning rate you often will indeed just move slower on the same path but a too big learning rate makes you skip entire paths.
OP seems to have been in that case with their first attempt.
So you're saying my intuition isn't wrong, necessarily, that slow training to learn the small subtle details could work as long as the dataset wasn't *too* limited in scope?
You are correct. Small learning rate allows to do fine adjustments to parameters and thereby learning subtle features. However, initially learning subtle features is useless, since you need to learn the coarse features first. That's why learning rate schedulers go from large learning rate to small learning rate. The tricky bit is doing the minimal amount of training on a large learning rate. That is where various optimizers come in, which try do automate these kinds of things.
You could try to do this by hand by saving checkpoints periodically, and try to find the point where you go from undertrained to overtrained. Then pick a checkpoint which is slightly undertrained, and start training from there with a lower learning rate.
Considering there's an implementation of the cosine scheduler with warmup steps, is there any implementation of a scheduler that starts slow, then rapidly accelerates, and finally stabilizes to learn the subtle features (like a sigmoidal function?) To avoid starting too high in the first place.
https://preview.redd.it/qb1z0n7oci2c1.png?width=1200&format=png&auto=webp&s=15dbab7b3a18ab918defbbbe2ab6816aaa46b489
Honestly, no idea. I have more theoretical than practical understanding. But my idea of the warmup phase is to arrange the initial totally random weights of a network into something where you can optimize on. When finetuning you don't start from randomness, you start from a trained checkpoint, so I expect that the warmup phase is pointless (at least for SGD, no idea if it helps adaptive optimizers). So believe you should go from high learning rate to low learning rate, unless somebody knows better.
Oh, and when training Loras, remember that changing alpha also changes the learning rate by the same factor if I remember right. So many tests about optimal alpha are probably invalid, because people didn't adjust the learning rate.
Good to know thank you