docs/experiment_deepcoder.md

DeepCoder Experiment Runs

Command Snippets Overview

Condition	Purpose	Variables
GRPO top-p 1.0	Full-retention baseline under linear top-p filtering	rollout_filter_strategy=top_p, rollout_filter_value=1, rollout_filter_include_zero=True
GRPO top-p 0.9	Stronger reward-variance filtering with adaptive retention	rollout_filter_strategy=top_p, rollout_filter_value=0.9, rollout_filter_include_zero=False
GRPO top-k 0.25	Fixed-budget filtering that keeps the top 25% of train groups	rollout_filter_strategy=top_k, rollout_filter_value=0.25, rollout_filter_include_zero=True

All three command snippets run DeepCoder with Qwen/Qwen2.5-Coder-7B, GRPO, and single-turn code generation.

---

1. Top-p 1.0 (Qwen/Qwen2.5-Coder-7B-200-GRPO-top-p-1)

Uses linear top-p filtering with rollout_filter_value=1.

Goal:

• Establish a full-retention baseline while keeping the same reward-variance ranking machinery as the filtered runs

Key Details:

• Filtering uses top_p, largest, reward_variance, and rollout_filter_top_p_prob_mode=linear
• rollout_filter_value=1 with rollout_filter_include_zero=True keeps the full train-group pool under linear top-p selection, so this is the closest thing to a no-filter baseline in deepcoder_lines
• actor_rollout_ref.actor.use_ref=False and actor_rollout_ref.actor.use_kl_loss=False remove reference-policy KL from training
• The run budget is 200 training steps with checkpoints every 20 steps

Source:

• docs/deepcoder_lines, lines 1-89

Outputs:

• W&B project: deepcoder_RAGEN_final_3
• Run name: Qwen/Qwen2.5-Coder-7B-200-GRPO-top-p-1

---

2. Top-p 0.9 (Qwen/Qwen2.5-Coder-7B-200-GRPO-top-p-0.9)

Uses the same linear top-p filter, but keeps only the highest-variance groups whose score mass reaches 0.9.

Goal:

• Increase reward-variance filtering strength while keeping the rest of the GRPO setup fixed

Key Details:

• Filtering again uses top_p, largest, reward_variance, and rollout_filter_top_p_prob_mode=linear
• rollout_filter_value=0.9 makes retention adaptive: the number of kept groups depends on how reward variance is distributed across the 16 train groups
• rollout_filter_include_zero=False excludes zero-variance groups from selection
• Because rollout_filter_type=largest, the filter prioritizes groups with the highest within-group reward variance

Source:

• docs/deepcoder_lines, lines 93-181

Outputs:

• W&B project: deepcoder_RAGEN_final_3
• Run name: Qwen/Qwen2.5-Coder-7B-200-GRPO-top-p-0.9

---

3. Top-k 0.25 (Qwen/Qwen2.5-Coder-7B-200-GRPO-top-k-0.25)

Switches from adaptive top-p filtering to fixed-fraction top-k filtering.

Goal:

• Compare adaptive top-p filtering against a fixed keep-top-25% regime

Key Details:

• rollout_filter_strategy=top_k with rollout_filter_value=0.25 keeps int(0.25 * 16) = 4 train groups per step
• With es_manager.train.group_size=8, this corresponds to at most 32 kept rollouts per training step after filtering
• rollout_filter_include_zero=True means zero-variance groups are still part of the ranking pool, but only the top 4 groups survive
• rollout_filter_type=largest means those 4 groups are chosen by highest reward variance

Source:

• docs/deepcoder_lines, lines 185-273

Outputs:

• W&B project: deepcoder_RAGEN_final_3
• Run name: Qwen/Qwen2.5-Coder-7B-200-GRPO-top-k-0.25

---

Common Notes

• Source format:
  • docs/deepcoder_lines is a collection of three standalone bash snippets, not a parameterized sweep script
  • The file defines both USE_GRPO and USE_PPO, but all three python train.py commands actually expand $USE_GRPO
• Shared setup across all three conditions:
  • Config: _10_deepcoder
  • Model: Qwen/Qwen2.5-Coder-7B
  • algorithm.adv_estimator=grpo
  • agent_proxy.reward_normalization.method=identity
  • trainer.total_training_steps=200
  • ppo_mini_batch_size=32
  • micro_batch_size_per_gpu=1
  • es_manager.train.env_groups=16, es_manager.train.group_size=8
  • es_manager.val.env_groups=256, es_manager.val.group_size=1
  • trainer.n_gpus_per_node=8
  • system.CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
  • actor_rollout_ref.rollout.tensor_model_parallel_size=4
  • agent_proxy.max_turn=1
  • actor_rollout_ref.actor.use_ref=False
  • actor_rollout_ref.rollout.rollout_filter_type=largest
  • actor_rollout_ref.rollout.rollout_filter_metric=reward_variance by default from config/base.yaml
  • actor_rollout_ref.rollout.rollout_filter_top_p_prob_mode=linear
  • actor_rollout_ref.rollout.rollout_filter_empty_stop_steps=0
  • actor_rollout_ref.rollout.max_model_len=10000
  • actor_rollout_ref.rollout.max_num_batched_tokens=10000
  • actor_rollout_ref.rollout.response_length=4000
  • agent_proxy.fail_on_prompt_too_long=True
  • lora.rank=0, lora.alpha=64, lora.target_modules=all-linear
  • actor_rollout_ref.rollout.gpu_memory_utilization=0.6
  • trainer.save_freq=20
  • trainer.validation_steps=1
  • trainer.val_before_train=True
  • trainer.test_freq=10
  • collapse_detection.first_turn_enabled=False
  • collapse_detection.multi_turn_enabled=False
  • trainer.resume_mode=disable
• Logging and artifacts:
  • Default local log dir remains results/
  • Default logger remains ['console', 'wandb']
  • Checkpoints are saved every 20 steps