docs/experiment_sokoban_gradient_analysis.md

Sokoban Gradient Analysis Runs

This doc covers the helper scripts for the Sokoban top-p=0.9 gradient-analysis experiments.

For the internal execution order, metric definitions, and plotting workflow, see guide_gradient_analysis.md.

Scripts Overview

Script	Purpose	When to use
run_sokoban_ppo_filter_grad_analysis.sh	Train Sokoban with periodic gradient-analysis passes	Run this first to produce training logs and checkpoints
run_sokoban_ppo_filter_grad_analysis_probe_ckpt.sh	Resume a saved checkpoint and run one analysis-only probe	Run this after the training script when you want to inspect a specific checkpoint

Both scripts run Sokoban with Qwen2.5-3B, reward-variance top-p filtering at 0.9, and a separate gradient-analysis batch of 128x16.

---

Recommended Workflow

1. Start with run_sokoban_ppo_filter_grad_analysis.sh.

• This is the script that actually trains the policy, runs periodic gradient analysis, and writes the checkpoint layout that the probe script expects by default.

2. Choose a saved global_step_* checkpoint.

• The probe helper defaults to global_step_101 under the checkpoint directory layout produced by the training script.
• If your run saved a different step, pass --checkpoint-step or --resume-from-path.

3. Run run_sokoban_ppo_filter_grad_analysis_probe_ckpt.sh.

• This reloads the checkpoint, runs one gradient-analysis pass, optionally performs a validation first, and then exits.

---

1. Periodic Training + Analysis (run_sokoban_ppo_filter_grad_analysis.sh)

Trains Sokoban and inserts gradient-analysis passes during training.

Goal:

• Follow the filtered Sokoban setup while logging gradient-analysis metrics at a fixed cadence on a larger analysis batch.

Key Details:

• Validation runs once before training and then every 10 steps.
• Gradient analysis runs every 50 steps. With the default 101 steps, the trigger points are 1, 51, and 101.
• The normal training batch is 8 env groups x 16 samples.
• Gradient analysis uses a separate batch of 128 env groups x 16 samples.
• The run continues after analysis because trainer.exit_after_gradient_analysis=False.
• This script uses top_p=0.9, rollout_filter_top_p_prob_mode=linear, rollout_filter_type=largest, rollout_filter_metric=reward_variance, and rollout_filter_include_zero=False.
• --algo PPO selects algorithm.adv_estimator=gae and actor_rollout_ref.actor.loss_agg_mode=token-mean.
• --algo GRPO selects algorithm.adv_estimator=grpo, algorithm.norm_adv_by_std_in_grpo=True, and actor_rollout_ref.actor.loss_agg_mode=seq-mean-token-mean.
• The training helper keeps actor_rollout_ref.actor.use_kl_loss=False, so it is meant for filtered training with periodic analysis rather than a KL-regularized sweep.

Examples:

# Default PPO run
bash scripts/runs/run_sokoban_ppo_filter_grad_analysis.sh

# GRPO run on four GPUs
bash scripts/runs/run_sokoban_ppo_filter_grad_analysis.sh --algo GRPO --gpus 0,1,2,3

# Short smoke test
bash scripts/runs/run_sokoban_ppo_filter_grad_analysis.sh --steps 5 --gpus 0,1,2,3

Options:

• --algo NAME (PPO or GRPO; default: PPO)
• --steps (default: 101)
• --gpus (comma list; auto-detect if omitted)
• --gpu-memory-utilization (default: 0.3)
• --ray-num-cpus (default: 16)
• --ppo-micro-batch-size-per-gpu (default: 4)
• --log-prob-micro-batch-size-per-gpu (default: 4)
• --save-freq (default: 100)

Outputs:

• Per-run log: logs/gradient_analysis_sokoban_Qwen2.5-3B/<exp_name>.log
• Checkpoints: model_saving/gradient_analysis/sokoban/<ALGO>/filter/<exp_name>/
• W&B project: ragen_gradient_analysis

---

2. Checkpoint Probe (run_sokoban_ppo_filter_grad_analysis_probe_ckpt.sh)

Resumes a saved checkpoint and runs one gradient-analysis-only probe.

Goal:

• Inspect one checkpoint without continuing the normal training run.

Key Details:

• The script resumes from an existing global_step_* directory with trainer.resume_mode=resume_path.
• It runs in probe mode with trainer.gradient_analysis_only=True.
• It exits after the analysis pass because trainer.exit_after_gradient_analysis=True.
• By default it does not run validation first; add --with-val if you want a pre-probe validation.
• It uses the same Sokoban task, model, filter setup, and analysis batch shape as the training helper.
• Unlike the training helper, this probe sets actor_rollout_ref.actor.use_kl_loss=True together with kl_loss_coef=0.001 and entropy_coeff=0.001, so the checkpoint probe explicitly logs KL and entropy gradient components.
• If --resume-from-path is given, that exact checkpoint directory is used. Otherwise the script resolves <checkpoint-root>/global_step_<checkpoint-step>.

Examples:

# Probe the default checkpoint layout produced by the training helper
bash scripts/runs/run_sokoban_ppo_filter_grad_analysis_probe_ckpt.sh

# Probe a specific saved step with validation
bash scripts/runs/run_sokoban_ppo_filter_grad_analysis_probe_ckpt.sh \
  --checkpoint-step 51 \
  --with-val \
  --gpus 0,1,2,3

# Probe an exact checkpoint path
bash scripts/runs/run_sokoban_ppo_filter_grad_analysis_probe_ckpt.sh \
  --resume-from-path model_saving/gradient_analysis/sokoban/PPO/filter/<exp_name>/global_step_101 \
  --gpus 0,1,2,3

Options:

• --algo NAME (PPO or GRPO; default: PPO)
• --checkpoint-step (default: 101)
• --checkpoint-root DIR (default: derived from the training helper's checkpoint layout)
• --resume-from-path DIR (exact global_step_* directory; overrides root + step resolution)
• --with-val (flag; default: off)
• --gpus (comma list; auto-detect if omitted)
• --gpu-memory-utilization (default: 0.3)
• --ray-num-cpus (default: 16)
• --ppo-micro-batch-size-per-gpu (default: 4)
• --log-prob-micro-batch-size-per-gpu (default: 4)

Outputs:

• Per-run log: logs/gradient_analysis_probe_sokoban_Qwen2.5-3B/<exp_name>.log
• Probe output dir: model_saving/gradient_analysis_probe/sokoban/<ALGO>/filter/<exp_name>/
• W&B project: ragen_gradient_analysis_probe

---

Common Notes

• Shared fixed setup:
  • config: _2_sokoban
  • model: Qwen/Qwen2.5-3B
  • training batch: es_manager.train.env_groups=8, es_manager.train.group_size=16
  • analysis batch: trainer.gradient_analysis_env_groups=128, trainer.gradient_analysis_group_size=16
  • trainer.gradient_analysis_log_prefilter=True
  • actor_rollout_ref.rollout.gradient_analysis_num_buckets=6
  • actor_rollout_ref.rollout.gradient_analysis_bucket_mode=quantile
• Shared rollout filter setup:
  • actor_rollout_ref.rollout.rollout_filter_value=0.9
  • actor_rollout_ref.rollout.rollout_filter_strategy=top_p
  • actor_rollout_ref.rollout.rollout_filter_top_p_prob_mode=linear
  • actor_rollout_ref.rollout.rollout_filter_type=largest
  • actor_rollout_ref.rollout.rollout_filter_metric=reward_variance
  • actor_rollout_ref.rollout.rollout_filter_include_zero=False
• GPU behavior:
  • if --gpus is omitted, the scripts try to auto-detect GPUs with nvidia-smi
  • if auto-detection fails, they fall back to 0,1,2,3,4,5,6,7
• Directory relationship:
  • the training helper writes checkpoints under model_saving/gradient_analysis/...
  • the probe helper reads from that layout by default and writes its own outputs under model_saving/gradient_analysis_probe/...
• If you need the meaning of bucket metrics, prefilter logging, or the plotting commands after the run finishes, use guide_gradient_analysis.md.