Reproducibility issue in dreamer v1 finger-spin

[stablegradients]

I tried to drproduce dreamer v1 results but they dont seem to track the original paper for different seeds, @ 80k eval steps different seeds are resulting in 199,0,0,6 as rewards for 4 different seeds. whereas other implementations state it shoudl be around 100 w/ std dev of ~70
commant used: python sheeprl.py exp=dreamer_v1 env=dmc env.wrapper.domain_name=finger env.wrapper.task_name=spin algo.cnn_keys.encoder=[rgb] fabric.accelerator=gpu metric.log_every=100 env.action_repeat=2 env.num_envs=1