Posted by volotat 1 day ago
So, first of all it does work and you can see the sample from the whole training run here: https://raw.githubusercontent.com/volotat/mini-AGI/refs/head...
Here is the scaling law graph I have so far, and it looks very promising: https://github.com/volotat/mini-AGI/blob/main/assets/scaling...
The model was built under my deep dissatisfaction so we cannot really train even moderately big models (1B+ scale) on the consumer's hardware. We can inference and fine-tune them for sure, but I would like to have full control over what the model sees over the training run, so it is fully aligned with my interests, not some corporations.
I was thinking about for some time and come up with two interesting ideas I thought worth pursuing: MoE with a lot of experts that gets added and pruned from the model while it trains, where only a small subset of of experts are actually in use at any particular moment + batch 1 training on the single continuous stream of data.
First allows us to be bounded only by the disk space in terms of number of parameters and load and unload experts only when they are needed. The second (if figured out and it turns out to be doable) allows us to get aways with small VRAM capacity because we do not need to store big randomized batches and their respective gradients.
I started brainstorming with Claude and after some time we found an approach that seems to be promising, and low and behold, a few weeks pass and you can see the results yourself.
Obviously, I did use AI in the process of making this project and I am pretty sure it would be completely impossible for me to do something like this without it, so I hope it is more than justified.
The model is still running over the first of 7.8B characters corpus I selected for training, so the weights are not out yet, and it's about a couple weeks of waiting until they are cooked at the current reading speed. And yeah, the model just read continuous interleaved passages from the dataset, each by 32K characters long each as a single stream. Just as you or I would do.
The set up seems to be really simple so you can git clone the project, run it and observe everything for yourself.
Thanks for your attention.
The difference might be smaller on a CPU which has limited parallelism.
But it's basically equivalent to a very deep model which might be problematic for training.
The concept is as follows: You train a critic to mimic the datastream and then you train against the critic instead of training against the data. The idea behind this is that the critic will memorize the training data so you do not need to store the full training data anymore. One of the biggest issues with current online stochastic gradient descent is that it is inherently a memory-less technique where the training data acts as the memory.
You can spin this further by going deeper with the nesting and then dropping the supervised critic. I forgot how to put it in words but the goal is that by having a model train against a critic of the critic, you can then drop the top level critic and instead use the mid level critic itself as your meta learning objective to train the actor against an unlabeled data stream.
Top level critic: learns to mimic the labeled training data via online SGD, then you add a simple hand written loss function to compare the predicted output with a given input. Basically you build a model specifically for distillation. Mid level critic: learns a reward function that mimics the top level critic directly but only gets to see the unlabeled training data and the result of the top level critic. Actor: The actor is exclusively trained against the mid level critic
Through this concept you end up with the existing training data stored as objective inside the mid level critic so you end up training not only against the latest data but also the already memorized data which should lower catastrophic forgetting. Of course at some point you might need to update the mid level critic again and to avoid that you might get away with just adding a very very wide Linear RNN / State Space Model / Mamba / Gated Delta Net as the middle critic (shower thought: use internal RNN states to represent LoRA vectors).
