There are lots of reasons to read through source code you never edit or recompile: security audits, interoperability, learning from their techniques, etc. And I think many of those same ideas apply to seeing the training data of a LLM. It will help you understand quickly (without as much experimentation) what it's likely to be good at, where its biases may be, where some kind of supplement (transfer learning? RAG? whatever) might be needed. And the why.
> security audits
If you are unable to run the multimillion training, then any kind of security audit of the training code is absolutely meaningless, because you have no way to verify that the weights were actually produced by this code.
Also, the analogy with source code/binary code fails really fast, considering that model training process is non-deterministic, so even if are able to run the training, then you get different weights than those that were released by the model developers, then... then what?
I probably shouldn't have led with that example because yeah, reproducible (and cheap) builds would be best for security audits. But I wouldn't say it's absolutely meaningless. At least it can guide your experimentation, and if results start differing radically from what you'd expect from the training data, that raises interesting questions.
If you're going through the effort to be open source you can probably set up fixed batch sizes and deterministic combination of batches without too much more effort. At least I hope it's not super hard.
> considering that model training process is non-deterministic
Why would it have to be? Just use PRNG with published seeds and then anyone can reproduce it.
I have zero actual experience in training models, but in general, when parallelizing work: there can be fundamental nondeterminism (e.g., some race conditions) that is tolerated, whose recording/reproduction can be prohibitive performance-wise.
Agree, this feels like a distinction that needs formalising...
Passive transparency: training data, technical report that tells you what the model learned and why it behaves the way it does. Useful for auditing, AI safety, interoperability.
Active transparency: being able to actually reproduce and augment the model. For that you need the training stack, curriculum, loss weighting decisions, hyperparameter search logs, synthetic data pipeline, RLHF/RLAIF methodology, reward model architecture, what behaviours were targeted and how success was measured, unpublished evals, known failure modes. The list goes on!
I'd also add training checkpoints to the list for active transparency. I think the Olmo models do a decent job, but it would be cool to see it for bigger models and for ones that are closer to state-of-the-art in terms of both architecture and algorithms.
Security audits, etc, are possible because binary code closely implements what the source code says.
In this case, you have no idea what the weights are going to "do", from looking at the source materials --- the training data and algorithm --- without running the training on the data.