About Training

Thanks for your work.
I have some questions related to training.
I tried to train the model with a small portion of the data, but when I tried to train using dataset online like:
https://huggingface.co/datasets/imageomics/TreeOfLife-10M/blob/main/dataset/EOL/image_set_01.tar.gz,
and download the dataset in local
``` python
python -m src.training.main \
--train-data 'https://huggingface.co/datasets/imageomics/TreeOfLife-10M/resolve/main/dataset/EOL/image_set_01.tar.gz' \
--val-data 'https://huggingface.co/datasets/imageomics/TreeOfLife-10M/resolve/main/dataset/EOL/image_set_01.tar.gz' \
--dataset-type 'webdataset' \
--pretrained 'openai' \
--text_type 'random' \
--warmup 100 \
--batch-size 1 \
--accum-freq 1 \
--epochs 10 \
--workers 1 \
--model ViT-B-16 \
--lr 1e-4 \
--log-every-n-steps 1 \
--dataset-resampled \
--local-loss \
--gather-with-grad \
--grad-checkpointing \
--logs '../storage/log/' \
--train-num-samples 98000 \
```

it always gets stuck at the following position

```shell
2024-12-11,23:16:02 | INFO | wandb_notes:
2024-12-11,23:16:02 | INFO | wandb_project_name: open-clip
2024-12-11,23:16:02 | INFO | warmup: 100
2024-12-11,23:16:02 | INFO | wd: 0.2
2024-12-11,23:16:02 | INFO | workers: 1
2024-12-11,23:16:02 | INFO | world_size: 1
2024-12-11,23:16:02 | INFO | zeroshot_frequency: 2
2024-12-11,23:16:02 | INFO | Finish counting shard total size: 98000.
2024-12-11,23:16:02 | INFO | Finish counting shard total size: 0.
2024-12-11,23:16:02 | INFO | Start epoch 0

```

In addition, I found the missing "data/resolved.jsonl" file when creating the data,
```shell
python scripts/evobio10m/make_metadata.py --db /fs/ess/PAS2136/open_clip/data/evobio10m-v3.3/mapping.sqlite
```
and the ToL-EDA HF Repo mentioned in the readme has disappeared

Can you provide me with some help to solve these problems
Or where can I find the details about training

Thank you very much

Start with src/training/main and the reported WebLoader output, then inspect scripts/evobio10m/make_metadata.py and the referenced README instructions. Reproduce the online dataset run and metadata-generation command; done means the training setup no longer stalls and the expected data/resolved.jsonl guidance is documented or corrected.

Written by the indexing model from the issue text.