O/B/M入口配置核对

{
  "schema": 1,
  "kind": "credit_arm_config_comparison",
  "observed_at": "2026-10-01T15:22:32+08:00",
  "scope": "Read-only comparison of retained launch logs; no new training or evaluation",
  "arms": {
    "B80": {
      "edit_credit_mode": "\u001b[36m(skyrl_entrypoint pid=197839)\u001b[0m   edit_credit_mode: broadcast",
      "harbor_trial_config_file": "\u001b[36m(skyrl_entrypoint pid=197839)\u001b[0m harbor_trial_config_file: mini_swe_tito_clean",
      "train_batch_size": "\u001b[36m(skyrl_entrypoint pid=197839)\u001b[0m   micro_train_batch_size_per_gpu: 1",
      "n_samples_per_prompt": "\u001b[36m(skyrl_entrypoint pid=197839)\u001b[0m   eval_n_samples_per_prompt: 1",
      "max_training_steps": "\u001b[36m(skyrl_entrypoint pid=197839)\u001b[0m   max_training_steps: 80",
      "resume_mode": "\u001b[36m(skyrl_entrypoint pid=197839)\u001b[0m   resume_mode: none",
      "Length of train_dataloader": "\u001b[36m(skyrl_entrypoint pid=197839)\u001b[0m 2026-09-21 09:13:42.318 | INFO     | skyrl.train.fully_async_trainer:_build_train_dataloader_and_compute_training_steps:428 - Length of train_dataloader: 2400",
      "train_dataloader_length": 2400,
      "environment_line": "IsolatedAgentDockerEnvironment",
      "config_line": "mini_swe_tito_clean",
      "log_sha256": "2b51f98b0da35285e73b6e3e1b8d1f47f70719b57346140dc3e93277c0c8d6cc",
      "log": "rl/runs/skyrl-qwen35-9b/rl9b-cfcredit-clean-B-b16n8-lr1e-6-fe05-p2p025-20260920/logs/launch_20260922_001146.log",
      "run": "rl9b-cfcredit-clean-B-b16n8-lr1e-6-fe05-p2p025-20260920"
    },
    "M80": {
      "edit_credit_mode": "\u001b[36m(skyrl_entrypoint pid=218400)\u001b[0m   edit_credit_mode: localized",
      "harbor_trial_config_file": "\u001b[36m(skyrl_entrypoint pid=218400)\u001b[0m harbor_trial_config_file: mini_swe_tito_clean",
      "train_batch_size": "\u001b[36m(skyrl_entrypoint pid=218400)\u001b[0m   micro_train_batch_size_per_gpu: 1",
      "n_samples_per_prompt": "\u001b[36m(skyrl_entrypoint pid=218400)\u001b[0m   eval_n_samples_per_prompt: 1",
      "max_training_steps": "\u001b[36m(skyrl_entrypoint pid=218400)\u001b[0m   max_training_steps: 80",
      "resume_mode": "\u001b[36m(skyrl_entrypoint pid=218400)\u001b[0m   resume_mode: none",
      "Length of train_dataloader": "\u001b[36m(skyrl_entrypoint pid=218400)\u001b[0m 2026-09-23 19:11:27.802 | INFO     | skyrl.train.fully_async_trainer:_build_train_dataloader_and_compute_training_steps:428 - Length of train_dataloader: 2400",
      "train_dataloader_length": 2400,
      "environment_line": "IsolatedAgentDockerEnvironment",
      "config_line": "mini_swe_tito_clean",
      "log_sha256": "50de812d3cd9b3c5dc2c06f8d0695873c05190f9ad9f7d8cdf77300c393d4a26",
      "log": "rl/runs/skyrl-qwen35-9b/rl9b-cfcredit-clean-M-retry1-b16n8-lr1e-6-fe05-p2p025-20260924/logs/launch_20260924_100931.log",
      "run": "rl9b-cfcredit-clean-M-retry1-b16n8-lr1e-6-fe05-p2p025-20260924"
    },
    "O80_original_clean": {
      "edit_credit_mode": "\u001b[36m(skyrl_entrypoint pid=2500424)\u001b[0m   edit_credit_mode: false",
      "harbor_trial_config_file": "\u001b[36m(skyrl_entrypoint pid=1542805)\u001b[0m harbor_trial_config_file: mini_swe_tito_clean_scm",
      "train_batch_size": "\u001b[36m(skyrl_entrypoint pid=1542805)\u001b[0m   micro_train_batch_size_per_gpu: 1",
      "n_samples_per_prompt": "\u001b[36m(skyrl_entrypoint pid=1542805)\u001b[0m   eval_n_samples_per_prompt: 1",
      "max_training_steps": "\u001b[36m(skyrl_entrypoint pid=1542805)\u001b[0m   max_training_steps: 80",
      "resume_mode": "\u001b[36m(skyrl_entrypoint pid=1542805)\u001b[0m   resume_mode: none",
      "Length of train_dataloader": "\u001b[36m(skyrl_entrypoint pid=1542805)\u001b[0m 2026-09-27 08:40:13.775 | INFO     | skyrl.train.fully_async_trainer:_build_train_dataloader_and_compute_training_steps:428 - Length of train_dataloader: 2384",
      "train_dataloader_length": 2400,
      "environment_line": "IsolatedAgentDockerEnvironment",
      "config_line": "mini_swe_tito_clean",
      "log_sha256": "d6678ae1a8fadddff010c696692a4a867328db265447ae46c7d6b7421fb7f26c",
      "log": "rl/runs/skyrl-qwen35-9b/rl9b-cfcredit-clean-O-b16n8-lr1e-6-fe05-p2p025-20260920/logs/launch_20260920_212653.log",
      "run": "rl9b-cfcredit-clean-O-b16n8-lr1e-6-fe05-p2p025-20260920"
    },
    "O42_retry1_scm": {
      "edit_credit_mode": "\u001b[36m(skyrl_entrypoint pid=1542805)\u001b[0m   edit_credit_mode: false",
      "harbor_trial_config_file": "\u001b[36m(skyrl_entrypoint pid=1542805)\u001b[0m harbor_trial_config_file: mini_swe_tito_clean_scm",
      "train_batch_size": "\u001b[36m(skyrl_entrypoint pid=1542805)\u001b[0m   micro_train_batch_size_per_gpu: 1",
      "n_samples_per_prompt": "\u001b[36m(skyrl_entrypoint pid=1542805)\u001b[0m   eval_n_samples_per_prompt: 1",
      "max_training_steps": "\u001b[36m(skyrl_entrypoint pid=1542805)\u001b[0m   max_training_steps: 80",
      "resume_mode": "\u001b[36m(skyrl_entrypoint pid=1542805)\u001b[0m   resume_mode: none",
      "Length of train_dataloader": "\u001b[36m(skyrl_entrypoint pid=1542805)\u001b[0m 2026-09-27 08:40:13.775 | INFO     | skyrl.train.fully_async_trainer:_build_train_dataloader_and_compute_training_steps:428 - Length of train_dataloader: 2384",
      "train_dataloader_length": 2384,
      "environment_line": "KeepImagesIsolatedDockerEnvironment",
      "config_line": "mini_swe_tito_clean_scm",
      "log_sha256": "1e92ef6a9d5ee09023d1c9af6377658c20f2e456427c8041f4281155c86f49ff",
      "log": "rl/runs/skyrl-qwen35-9b/rl9b-cfcredit-scm-O-s42-b16n8-lr1e-6-20260927-retry1/logs/launch_20260927_233816.log",
      "run": "rl9b-cfcredit-scm-O-s42-b16n8-lr1e-6-20260927-retry1"
    }
  },
  "finding": "Original O80, B80, and M80 all use the same mini_swe_tito_clean/IsolatedAgentDockerEnvironment profile with a 2400-row dataloader, so this comparison does not explain the B-vs-M entropy difference. The later O42 retry1 uses the newer clean_scm/KeepImagesIsolatedDockerEnvironment profile with 2384 rows; that is a separate recovery-profile difference and must not be conflated with the original O/B/M comparison."
}