training metric train/code_usage_step

[choochooyaya6-blip]

Thanks for your great work.
In stage 1 and stage 2, what is the typical value of the training metric train/code_usage_step?
When I try to reproduce the results on some datasets, this train/code_usage_step metric is around 0.25 and train/code_usage_epoch is around 0.3 — is that considered too low?

[retsuh-bqw]

I've just updated the stage-1 training code to correct the source of the instruction embeddings for the LAM (Latent Action Model) decoder. They are now correctly retrieved from the T5 encoder instead of the LAM encoder.

The previous implementation was an experimental ablation study design and has been proven to be inferior to this updated version. Please use the new code if you plan to train LAM from scratch.

Regarding code usage: with our full diverse dataset, it typically remains above 0.95 for both stages. However, if you are training on a less diverse, single dataset (e.g., BridgeV2 only), expect it to be lower. A value around 0.3, however, is still considered too low and may indicate an issue.

[Matilda-de]

Thanks for the great work!

I'd like to ask what the final value of train/code_usage_uncontrol is approximately? If I need to finetune LAM on my dataset, can I directly train based on the pre-trained stage2 model?

[retsuh-bqw]

Yes, you can directly fine-tune our LAM on your custom dataset. Please note that the train/code_usage_uncontrol metric may fluctuate and is typically around 0.3.