docs/experiment_webshop_release.md

WebShop Release Runs

This doc covers the current WebShop release runner.

Overview

Use scripts/runs/run_webshop_release_combos.sh for the main WebShop release experiments.

Fixed setup:

• Task: WebShop
• Config: _6_webshop
• Model: Qwen/Qwen2.5-3B-Instruct
• Algorithm: GRPO
• Train batch: 16 env groups x 8 samples
• Validation batch: 256 env groups x 1 sample
• W&B project: main_webshop

The runner compares four filter modes while keeping the rest of the setup fixed:

Filter	Strategy	Value	Prob Mode	Description
topk25	top_k	0.25	linear	Keep top 25% of groups by reward variance
topp09	top_p	0.9	linear	Keep groups covering 90% cumulative reward variance
topp095	top_p	0.95	linear	Keep groups covering 95% cumulative reward variance
nofilter	top_p	1.0	linear	Keep all groups

All four modes use:

• rollout_filter_type=largest
• rollout_filter_metric=reward_variance
• rollout_filter_include_zero=True

Setup

For WebShop, use the original base setup path:

bash scripts/setup_ragen_old.sh
bash scripts/setup_webshop.sh

Do not use setup_ragen.sh as the primary WebShop setup path.

Notes:

• setup_webshop.sh now treats the Google Drive full-data download as best-effort. A failure there does not mean the Python environment setup failed.
• The runner already pins micro_batch_size_per_gpu=1 and log_prob_micro_batch_size_per_gpu=1 to reduce WebShop rollout/log-prob memory pressure.

Main Commands

4 GPUs

Run one experiment on 4 GPUs:

bash scripts/runs/run_webshop_release_combos.sh \
  --steps 400 \
  --gpus 0,1,2,3 \
  --gpus-per-exp 4 \
  --filters nofilter

Other filter modes:

bash scripts/runs/run_webshop_release_combos.sh \
  --steps 400 \
  --gpus 0,1,2,3 \
  --gpus-per-exp 4 \
  --filters topk25

bash scripts/runs/run_webshop_release_combos.sh \
  --steps 400 \
  --gpus 0,1,2,3 \
  --gpus-per-exp 4 \
  --filters topp09

bash scripts/runs/run_webshop_release_combos.sh \
  --steps 400 \
  --gpus 0,1,2,3 \
  --gpus-per-exp 4 \
  --filters topp095

8 GPUs

Run one experiment on 8 GPUs:

bash scripts/runs/run_webshop_release_combos.sh \
  --steps 400 \
  --gpus 0,1,2,3,4,5,6,7 \
  --gpus-per-exp 8 \
  --filters nofilter

You can swap nofilter for topk25, topp09, or topp095.

Key Flags

Flag	Default	Description
--steps N	100	Total training steps
--gpus LIST	auto-detect	Comma-separated GPU IDs
--gpus-per-exp N	1	GPUs assigned to one experiment
--save-freq N	100	Checkpoint save frequency
--gpu-memory-utilization V	0.3	vLLM rollout memory fraction
--filters LIST	all	topk25, topp09, topp095, nofilter, or all
--cooldown N	30	Delay before reusing the same GPU slot

Outputs

• Logs: logs/webshop_release_combos/<exp_name>.log
• Per-run summaries: logs/webshop_release_combos/<exp_name>.result
• Checkpoints: model_saving/webshop_release_combos/<model>/<algo>/<filter>/<exp_name>/
• W&B project: main_webshop

Experiment names follow:

webshop-release-GRPO-<filter-suffix>-Qwen2.5-3B-Instruct-16x8

Examples:

• webshop-release-GRPO-topk25-Qwen2.5-3B-Instruct-16x8
• webshop-release-GRPO-topp09-linear-Qwen2.5-3B-Instruct-16x8
• webshop-release-GRPO-topp095-linear-Qwen2.5-3B-Instruct-16x8
• webshop-release-GRPO-nofilter-Qwen2.5-3B-Instruct-16x8

Related Files

• scripts/runs/run_webshop_release_combos.sh
• scripts/runs/run_webshop_small_combos.sh
• scripts/runs/README_webshop_small_combos.md