docs/experiment_frozen_lake_slipper_sweep.md

FrozenLake Slipper Sweep Runs

This doc covers the experiment script for the FrozenLake slipper-rate sweep.

Scripts Overview

Script	Purpose	Variables
run_frozen_lake_slipper_rate_sweep.sh	Sweep FrozenLake stochasticity while comparing filter vs nofilter	slipper_rate (100,50,20,10,5,2,0 by default), filter/nofilter

The script runs FrozenLake with Qwen2.5-3B, GAE.

---

1. FrozenLake Slipper Sweep (run_frozen_lake_slipper_rate_sweep.sh)

Tracks how filter and nofilter success rates change as FrozenLake stochasticity varies via slipper_rate, using project ragen_release_frozenlake_slipper_rate_sweep.

Goal:

• Test whether RV-style filtering remains helpful as FrozenLake transition randomness changes

Key Details:

• slipper_rate is normalized to a ratio in [0, 1], and the environment is configured with success_rate = 1 - slipper_rate
• Default comparison modes are both filter and nofilter
• This script explicitly fixes rollout_filter_top_p_prob_mode=softmax
• Mode mapping:
  • filter: top_p=0.9 by default and rollout_filter_include_zero=False
  • nofilter: top_p=1.0 by default and rollout_filter_include_zero=True

bash scripts/runs/run_frozen_lake_slipper_rate_sweep.sh

Options:

• --steps (default: 400)
• --slipper-rate (comma list; accepts 100,50,20,10,5,2,0, 1.0,0.5,..., or %-suffixed values)
• --filter-modes (comma list; filter, nofilter, or both; default: both)
• --filter-top-p (default: 0.9)
• --nofilter-top-p (default: 1.0)
• --gpus (comma list; auto-detect if omitted)
• --gpus-per-exp (default: 1)
• --ray-num-cpus (default: 16)
• --cooldown (default: 30)
• --gpu-memory-utilization (default: 0.5)
• --save-freq (default: -1)

Examples:

# Run the full default sweep
bash scripts/runs/run_frozen_lake_slipper_rate_sweep.sh

# Run only `nofilter` on a custom subset of slipper rates
bash scripts/runs/run_frozen_lake_slipper_rate_sweep.sh --slipper-rate 50,20,5 --filter-modes nofilter --gpus 0 --cooldown 30 --ray-num-cpus 8

# Run one `filter` and one `nofilter` 50%-slipper experiment on 4xH100 each
bash scripts/runs/run_frozen_lake_slipper_rate_sweep.sh --slipper-rate 50 --gpus-per-exp 4 --gpus 0,1,2,3,4,5,6,7

Outputs:

• Per-run logs: logs/frozenlake_slipper_rate_sweep_Qwen2.5-3B/<mode>/slip<label>/
• Summary log: logs/frozenlake_slipper_rate_sweep_Qwen2.5-3B.log

---

Common Notes

• Shared setup:
  • Config: _3_frozen_lake
  • Model: Qwen/Qwen2.5-3B
  • algorithm.adv_estimator=gae
  • trainer.total_training_steps=400
  • trainer.save_freq=-1
  • trainer.logger=['console','wandb']
  • trainer.val_before_train=True
  • actor_rollout_ref.actor.loss_agg_mode=token-mean
  • actor_rollout_ref.actor.use_kl_loss=False
  • actor_rollout_ref.actor.kl_loss_type=low-var-kl
  • actor_rollout_ref.actor.kl_loss_coef=0
  • actor_rollout_ref.actor.entropy_coeff=0
  • actor_rollout_ref.actor.entropy_from_logits_with_chunking=True
  • actor_rollout_ref.actor.filter_loss_scaling=none
  • actor_rollout_ref.actor.ppo_mini_batch_size=32
  • actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4
  • critic.ppo_mini_batch_size=32
  • critic.ppo_micro_batch_size_per_gpu=4
  • actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
  • actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
  • actor_rollout_ref.rollout.rollout_filter_strategy=top_p
  • actor_rollout_ref.rollout.rollout_filter_top_p_prob_mode=softmax
  • actor_rollout_ref.rollout.rollout_filter_type=largest
  • actor_rollout_ref.rollout.rollout_filter_metric=reward_variance
  • actor_rollout_ref.rollout.gpu_memory_utilization=0.5
  • actor_rollout_ref.actor.checkpoint.save_contents=[model]
  • critic.checkpoint.save_contents=[model]
• Input and naming conventions:
  • slipper_rate accepts 50, 0.5, and 50% as equivalent inputs
  • Experiment labels use slip<label> with compact decimal formatting
  • Examples: 50% -> slip0p5, 2% -> slip0p02
• Comparison protocol:
  • Each slipper rate is run under both filter and nofilter unless --filter-modes restricts the set
  • With --gpus-per-exp 4, a 4-GPU list runs one experiment at a time; an 8-GPU list can run one filter and one nofilter experiment in parallel
• Base-config inheritance:
  • algorithm.kl_ctrl.kl_coef is not overridden in this script