Description du poste
The current problem: Language models struggle with basic commonsense and learn dangerous capabilities. If these two problems can be solved at once, the current AI situation could be made much safer.
Premise: A fixed-size model trained on a fixed-amount of data can gain predictive performance if the data is re-selected from a simpler distribution. Also, a model may learn dangerous capabilities from the data it is trained on.
It follows: I recommend constraining language modeling pretraining data in order to build better and safer language models. Small scale results, more theoretical details, further findings, and next steps below.
Small scale results:
I showed order-of-magnitude performance improvements at the small scale when I trained models on a constrained version of the TinyStories dataset (https://zenodo.org/records/23306194) and analyzed reasoning and generalization in this regime (https://zenodo.org/records/24075817). A ~10x smaller model trained on ~10x less data (~100x less compute) (constrained to a smaller world with less unique places, objects, adjectives, etc.) achieved comparable LLM-as-a-judge story output quality scores to the original TinyStories-33m param model, just within the smaller world. Also, what the model knew about (specific places, objects, adjectives, etc.) could be controlled. This is scalable up to trillions of parameters and tokens.
More theoretical details:
I view pretraining data as being sampled from a distribution that could described by a theoretical, very complex, unrealized knowledge graph. I view entities (people, places, things, concepts, actions, adjectives, etc.) as nodes in this knowledge graph, and I view relationships (is-a, has-a, part-of, etc.) as being probabilistic connections between nodes. A knowledge graph can be simplified by pruning nodes; pretraining data can be constrained by deleting entities.
I create a whitelist of allowed lemmas (for example: "run") where Kuperman's age of acquisition (AoA) is (or probabl