~243 RL agent checkpoints trained on the JaxGMG maze environment with a mixture training distribution: a blend of uniform state visitation and a biased distribution where the mouse starts east of the cheese. Used to study how training distribution mixtures affect learned behaviour and phase transitions.

WandB: https://wandb.ai/devinterp/jaxgmg2_mixture

This repository contains checkpoints from several distinct experiments with different mixture ratios:

Note: The 90/10 group uses alpha=0.0 and cheese_loc=corner, making it from an earlier/different experiment. The 100uniform control group is also from an earlier experiment with uncertain hyperparams. All other groups use alpha=1.0 and cheese_loc=any.

rl_action=train
alpha=1.0
discount_rate=0.98
lr=5e-05
num_total_env_steps=2000000000
num_rollout_steps=64
num_levels=9600
cheese_loc=any
env_layout=open
env_size=13
log_optimizer_state=True
ckpt_dir=jaxgmg2_mixture_east
wandb_project=jaxgmg2_mixture
use_wandb=True
use_hf=True

The env_rule_mixture parameter varies per group (see individual yamls).

Checkpoints are named mixture_{p}uniform_{100-p}east_seed_{seed}.

Yamls from rl/einar/pattern-merge-runpod branch:

from the timaeus monorepo.