timaeus/jaxgmg_al_1e0

OBSOLETE

These models were originally used for RL1, but were trained with previous action, and with the variable learning rate bug. Do not use. Code these were trained with has since been removed.

Result of wandb sweep with sweeps/alpha_1_seed_sweep.py.

Wandb runs

program: /root/timaeus/projects/rl/main_train.py
method: grid
project: jaxgmg_al_1e0
entity: devinterp
command:
  - /root/timaeus/.venv/bin/python
  - ${program}
  - ${args}
  - --use-wandb
  - --use-hf
parameters:
  mask-type:
    value: first_episode
  eval-schedule:
    value: "0:1,250:2,500:5,1000:10,2000:20"
  num-total-env-steps:
    value: 5000000000
  discount-rate:
    value: 0.99
  alpha:
    value: 1.0
  seed:
    values: [200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220]
  cheese-loc:
    value: "any"
  num-levels:
    value: 9600
  grad-acc-per-chunk:
    value: 5
  wandb-project:
    value: jaxgmg_al_1e0
  ckpt-dir:
    value: jaxgmg_al_1e0
  f-str-ckpt:
    value: "al_{alpha}_g_{discount_rate}_seed_{seed}"