Will loss curves on Pythia models of different sizes trained on the same data in the same order be similar? | Manifold

Will loss curves on Pythia models of different sizes trained on the same data in the same order be similar?

Mechanistic interpretability

Basic

11

Ṁ748

Nov 30

76%

chance

1D

1W

1M

ALL

Someone in the EleutherAI discord is reporting that finetuning Pythia models of different sizes on the same data in the same order is giving spookily similar loss curves, just vertically shifted.

Will training Pythia models from scratch in the same way produce similar behaviour?
Resolves N/A if it turns out the original result was just a bug or something like that.

Get Ṁ600 play money

Sort by:

Yes

Some evidence https://arxiv.org/abs/2305.18411

Feature-Learning Networks Are Consistent Across Widths At Realistic Scales

We study the effect of width on the dynamics of feature-learning neural networks across a variety of architectures and datasets. Early in training, wide neural networks trained on online data have not only identical loss curves but also agree in their point-wise test predictions throughout training.…

Won't bet in case it's less obious but for the record mi prediction is like 30%.

Have a vague model about how the behaviour happens because the finetuning is affecting the models in the same way but not something else that lowers the loss significantly.

But I'm very confused and don't really get wtf is happening here tbh.

Related in Mechanistic interpretability

By the end of 2026, will we have transparency into any useful internal pattern within a Large Language Model whose semantics would have been unfamiliar to AI and cognitive science in 2006?

Is attention all you need? (transformers SOTA in 2027)

See more Mechanistic interpretability questions

More related questions

Will GPT-4 be trained (roughly) compute-optimally using the best-known scaling laws at the time?

Will any language model trained without large number arithmetic be able to generalize to large number arithmetic by 2026?

Do scaling laws happen because models experience a ton of tiny phase changes which average out to a smooth curve?

Will it be possible to disentangle most of the features learned by a model comparable to GPT-3 this decade? (1k subsidy)

Will any transfer learning model, trained for any amount of time on one Atari environment, outperform the median human learning curve on most other Atari environments when transferred by 2026?

Will GPT-4 improve on the Chinchilla scaling law?

Do scaling laws happen because models experience a ton of tiny phase changes which average out to a smooth curve?

Are Mixture of Expert (MoE) transformer models generally more human interpretable than dense transformers?

Will it be possible to disentangle most of the features learned by a model comparable to GPT-4 this decade?

Will we be able to estimate the feature importance curve or feature sparsity curve of real models? (2024 end)

Mechanistic interpretability questions

By the end of 2026, will we have transparency into any useful internal pattern within a Large Language Model whose semantics would have been unfamiliar to AI and cognitive science in 2006?

Is attention all you need? (transformers SOTA in 2027)

Related questions

Will GPT-4 be trained (roughly) compute-optimally using the best-known scaling laws at the time?

Will GPT-4 improve on the Chinchilla scaling law?

Will any language model trained without large number arithmetic be able to generalize to large number arithmetic by 2026?

Do scaling laws happen because models experience a ton of tiny phase changes which average out to a smooth curve?

Do scaling laws happen because models experience a ton of tiny phase changes which average out to a smooth curve?

Are Mixture of Expert (MoE) transformer models generally more human interpretable than dense transformers?

Will it be possible to disentangle most of the features learned by a model comparable to GPT-3 this decade? (1k subsidy)

Will it be possible to disentangle most of the features learned by a model comparable to GPT-4 this decade?

Will any transfer learning model, trained for any amount of time on one Atari environment, outperform the median human learning curve on most other Atari environments when transferred by 2026?

Will we be able to estimate the feature importance curve or feature sparsity curve of real models? (2024 end)