| # Train 2 TODO | |
| - name the project (like tr2-26B-prompt) | |
| - arch&scale suggests using the same model size as tr1 (13B) but with the model and data changes listed below | |
| - group the tensorboard reports: | |
| ``` | |
| Batch-size | |
| - Batch-size | |
| - Batch-size vs samples | |
| Grad-norm | |
| - Grad norm | |
| - Grad norm vs samples | |
| Learning rate | |
| - Learning rate | |
| - Learning rate vs samples | |
| Lm loss train | |
| - Lm loss | |
| - Lm loss vs samples | |
| Lm loss validation | |
| - Lm loss | |
| - Lm loss vs samples | |
| - Lm loss ppl | |
| - Lm loss ppl vs samples | |
| Loss scale | |
| - Loss scale | |
| - Loss scale vs samples | |
| Num zeros | |
| - Num zeros | |
| - Num zeros vs samples | |
| ``` | |
| that's mostly about changing to | |
| ``` | |
| tb.add_scalar("batch size/batch size", batch_size, iteration) | |
| tb.add_scalar("batch size/batch size vs samples", batch_size, args.consumed_train_samples) | |
| ``` | |
| tracking: https://github.com/bigscience-workshop/Megatron-DeepSpeed/issues/38 | |
| add new metrics: XXX | |
| - Depending on the results from the arch&scale experiments (when do we expect to start this run? we want to make sure we have answers for the following questions by then) | |
| - Rotary embeddings | |
| - Prefix-lm | |
| - Train on multiple languages | |