# jaxgmg2_3phase_optim_state_alpha1

> **Note:** Einar trained these models and the description below is uncertain.

Single RL agent checkpoint from jaxgmg2_3phase_optim_state resumed with alpha changed from 0.6 to 1.0,
exploring how changing the value function loss weight mid-training affects learning dynamics.

**WandB:** https://wandb.ai/devinterp/jaxgmg2_patt

## Hyperparams

```
rl_action=train
alpha=1.0
discount_rate=0.98
lr=5e-05
num_total_env_steps=7372800000
num_rollout_steps=64
num_levels=9600
cheese_loc=any
env_layout=open
env_size=13
mask_type=first_episode
use_prev_action=False
grad_acc_per_chunk=4
log_optimizer_state=True
resume=jaxgmg2_3phase_optim_state/al_0.6_g_0.98_id_17_seed_980617
resume_id=3810
resume_optim=True
deterministic=True
seed=42
checkpoint=al_0.6_g_0.98_id_17_seed_980617_resume_alpha1
ckpt_dir=jaxgmg2_3phase_optim_state_alpha1
wandb_project=jaxgmg2_patt
use_wandb=True
use_hf=True
```

## Naming Schema

Checkpoint: `al_0.6_g_0.98_id_17_seed_980617_resume_alpha1`

## Reproduced with

See [`train.yaml`](./train.yaml) in this repository. Run from the
[timaeus monorepo](https://github.com/timaeus-research/timaeus).