docs/experiment_main_table.md

Main Table Runs

This doc covers the experiment scripts for the main performance table.

Scripts Overview

Script	Purpose	Variables
run_main_table_diff_algo.sh	Compare algorithms	PPO, DAPO, GRPO, DrGRPO; filter/nofilter
run_main_table_diff_size.sh	Compare model sizes	0.5B, 1.5B, 3B, 7B; filter/nofilter
run_main_table_diff_model.sh	Compare model types	Instruct, Reasoning; filter/nofilter

All scripts run experiments across 5 tasks (sokoban, frozenlake, webshop, metamathqa, countdown) with filter/nofilter settings.

---

1. Different Algorithms (run_main_table_diff_algo.sh)

Compares PPO/DAPO/GRPO/DrGRPO using Qwen2.5-3B.

bash scripts/runs/run_main_table_diff_algo.sh --steps 400

Options:

• --steps (default: 400)
• --tasks (comma list; default: sokoban,frozenlake,webshop,metamathqa,countdown)
• --algos (comma list; default: PPO,DAPO,GRPO,DrGRPO)
• --gpus (comma list; auto-detect if omitted)
• --gpus-per-exp (default: 1)
• --cooldown (seconds; default: 30)
• --gpu-memory-utilization (default: 0.3)
• --filters (comma list; filter, nofilter, or all; default: all)

Examples:

# Run one `filter` and one `nofilter` PPO experiment on 4xH100 each
bash scripts/runs/run_main_table_diff_algo.sh --steps 400 --tasks sokoban --gpus-per-exp 4 --gpu-memory-utilization 0.3 --filters all --gpus 0,1,2,3,4,5,6,7 --algos PPO

Outputs:

• Per-task logs: logs/diff_algo_<task>_Qwen2.5-3B/
• Summary log: logs/diff_algo_Qwen2.5-3B.log

---

2. Different Model Sizes (run_main_table_diff_size.sh)

Compares Qwen2.5 models of different sizes using PPO.

bash scripts/runs/run_main_table_diff_size.sh --steps 400

Options:

• --steps (default: 400)
• --tasks (comma list; default: sokoban,frozenlake,webshop,metamathqa,countdown)
• --models (comma list; default: Qwen2.5-0.5B,Qwen2.5-1.5B,Qwen2.5-3B,Qwen2.5-7B)
• --gpus (comma list; auto-detect if omitted)
• --gpus-per-exp (default: 1)
• --cooldown (seconds; default: 30)
• --gpu-memory-utilization (default: 0.3)
• --filters (comma list; filter, nofilter, or all; default: all)

Examples:

# Run a single 1.5B/filter experiment on 4xH100
bash scripts/runs/run_main_table_diff_size.sh --steps 400 --tasks sokoban --gpus-per-exp 4 --gpu-memory-utilization 0.3 --filters filter --gpus 0,1,2,3 --models Qwen2.5-1.5B

# Quick test with smallest model
bash scripts/runs/run_main_table_diff_size.sh --steps 5 --models Qwen2.5-0.5B --tasks sokoban

Outputs:

• Per-task logs: logs/diff_size_<task>/
• Summary log: logs/diff_size_PPO.log

---

3. Different Model Types (run_main_table_diff_model.sh)

Compares different model types (Instruct, Reasoning) using PPO.

bash scripts/runs/run_main_table_diff_model.sh --steps 400

Options:

• --steps (default: 400)
• --tasks (comma list; default: sokoban,frozenlake,webshop,metamathqa,countdown)
• --models (comma list; default: Qwen2.5-3B-Instruct)
• --gpus (comma list; auto-detect if omitted)
• --gpus-per-exp (default: 1)
• --cooldown (seconds; default: 30)
• --gpu-memory-utilization (default: 0.3)
• --filters (comma list; filter, nofilter, or all; default: all)

Examples:

# Run one `filter` and one `nofilter` Llama-3.2-3B-Instruct experiment on 4xH100 each
bash scripts/runs/run_main_table_diff_model.sh --steps 400 --tasks sokoban --gpus-per-exp 4 --gpu-memory-utilization 0.5 --filters all --gpus 0,1,2,3,4,5,6,7 --models=meta-llama/Llama-3.2-3B-Instruct

Outputs:

• Per-task logs: logs/diff_model_<task>/
• Summary log: logs/diff_model_PPO.log

---

Common Notes

• Effective rollout filter config for main-table runs:
  • rollout_filter_strategy=top_p
  • rollout_filter_top_p_prob_mode=softmax
  • rollout_filter_type=largest
  • rollout_filter_metric=reward_variance
  • rollout_filter_include_zero=True
• Filter mode mapping:
  • filter: top_p=0.9, include_zero=True
  • nofilter: top_p=1.0, include_zero=True
• Because include_zero=True, nofilter (top_p=1.0) keeps all groups; it does not disable the filter code path, but it is effectively "no filtering" for the batch
• You can run a single experiment on 4xH100 by setting --gpus-per-exp 4 and passing a 4-GPU list, or run one filter and one nofilter experiment in parallel by passing an 8-GPU list