hbfreed commited on
Commit
6b1cbcf
·
verified ·
1 Parent(s): 2182264

Add files using upload-large-folder tool

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. healed/grid_math/reap_keep25_s1225/args.json +62 -0
  2. healed/grid_math/reap_keep25_s1225/train_log.jsonl +150 -0
  3. healed/grid_math/uniform_keep25_s1226/train_log.jsonl +150 -0
  4. healed/grid_math/uniform_keep75_s1226/args.json +62 -0
  5. healed/grid_math/uniform_keep75_s1226/train_log.jsonl +150 -0
  6. healed/liger_mb8/step0004/chat_template.jinja +9 -0
  7. healed/liger_mb8/step0004/config.json +887 -0
  8. healed/liger_mb8/step0004/configuration_pruned_olmoe.py +27 -0
  9. healed/liger_mb8/step0004/generation_config.json +6 -0
  10. healed/liger_mb8/step0004/model.safetensors.index.json +0 -0
  11. healed/liger_mb8/step0004/modeling_pruned_olmoe.py +66 -0
  12. healed/liger_mb8/step0004/special_tokens_map.json +23 -0
  13. healed/liger_mb8/step0004/tokenizer.json +0 -0
  14. healed/liger_mb8/step0004/tokenizer_config.json +247 -0
  15. healed/liger_smoke/step0006/chat_template.jinja +9 -0
  16. healed/liger_smoke/step0006/config.json +887 -0
  17. healed/liger_smoke/step0006/configuration_pruned_olmoe.py +27 -0
  18. healed/liger_smoke/step0006/generation_config.json +6 -0
  19. healed/liger_smoke/step0006/model.safetensors.index.json +0 -0
  20. healed/liger_smoke/step0006/modeling_pruned_olmoe.py +66 -0
  21. healed/liger_smoke/step0006/special_tokens_map.json +23 -0
  22. healed/liger_smoke/step0006/tokenizer.json +0 -0
  23. healed/liger_smoke/step0006/tokenizer_config.json +247 -0
  24. healed/mixonly_keep50/vllm_live/chat_template.jinja +9 -0
  25. healed/mixonly_keep50/vllm_live/config.json +887 -0
  26. healed/mixonly_keep50/vllm_live/configuration_pruned_olmoe.py +27 -0
  27. healed/mixonly_keep50/vllm_live/generation_config.json +6 -0
  28. healed/mixonly_keep50/vllm_live/model.safetensors.index.json +0 -0
  29. healed/mixonly_keep50/vllm_live/modeling_pruned_olmoe.py +66 -0
  30. healed/mixonly_keep50/vllm_live/special_tokens_map.json +23 -0
  31. healed/mixonly_keep50/vllm_live/tokenizer.json +0 -0
  32. healed/mixonly_keep50/vllm_live/tokenizer_config.json +247 -0
  33. healed/mixonly_keep50/wandb/debug-internal.log +11 -0
  34. healed/mixonly_keep50/wandb/debug.log +19 -0
  35. healed/mixonly_keep50/wandb/offline-run-20260731_011513-gexxktae/files/requirements.txt +130 -0
  36. healed/mixonly_keep50/wandb/offline-run-20260731_011513-gexxktae/logs/debug-core.log +6 -0
  37. healed/mixonly_keep50/wandb/offline-run-20260731_011513-gexxktae/logs/debug-internal.log +11 -0
  38. healed/mixonly_keep50/wandb/offline-run-20260731_011513-gexxktae/logs/debug.log +19 -0
  39. healed/opd_warm_fixed_keep50/step0120/chat_template.jinja +9 -0
  40. healed/opd_warm_fixed_keep50/step0120/config.json +887 -0
  41. healed/opd_warm_fixed_keep50/step0120/configuration_pruned_olmoe.py +27 -0
  42. healed/opd_warm_fixed_keep50/step0120/generation_config.json +6 -0
  43. healed/opd_warm_fixed_keep50/step0120/model.safetensors.index.json +0 -0
  44. healed/opd_warm_fixed_keep50/step0120/modeling_pruned_olmoe.py +66 -0
  45. healed/opd_warm_fixed_keep50/step0120/special_tokens_map.json +23 -0
  46. healed/opd_warm_fixed_keep50/step0120/tokenizer.json +0 -0
  47. healed/opd_warm_fixed_keep50/step0120/tokenizer_config.json +247 -0
  48. healed/opd_warm_fixed_keep50/step0200/chat_template.jinja +9 -0
  49. healed/opd_warm_fixed_keep50/step0200/config.json +887 -0
  50. healed/opd_warm_fixed_keep50/step0200/configuration_pruned_olmoe.py +27 -0
healed/grid_math/reap_keep25_s1225/args.json ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "student": "outputs/pruned/reap-0125inst-math-keep25",
3
+ "teacher": "allenai/OLMoE-1B-7B-0125-Instruct",
4
+ "training_mode": "off-policy",
5
+ "kl_direction": "forward",
6
+ "dataset": "allenai/RLVR-MATH",
7
+ "dataset_sources": null,
8
+ "max_difficulty": null,
9
+ "trajectories": "outputs/teacher_trajectories/dolci_math_curated.jsonl",
10
+ "trajectory_dataset": "allenai/Dolci-Instruct-RL",
11
+ "off_policy_frames": "chat",
12
+ "off_policy_max_seq_len": 2048,
13
+ "topk_targets": "outputs/teacher_trajectories/dolci_math_curated_opd_top128",
14
+ "max_loss_tokens": null,
15
+ "loss_tokens_per_step": 120000,
16
+ "teacher_device": "cuda:0",
17
+ "student_device": "cuda:0",
18
+ "lr": 3e-05,
19
+ "optimizer": "adamw8bit",
20
+ "weight_decay": 0.1,
21
+ "epochs": 3,
22
+ "prompts_per_step": 256,
23
+ "group_size": 1,
24
+ "rollout_batch": 64,
25
+ "micro_batch": 3,
26
+ "max_new_tokens": 256,
27
+ "max_prompt_len": 1024,
28
+ "warmup_steps": 10,
29
+ "max_grad_norm": 1.0,
30
+ "eval_every": 10,
31
+ "gsm8k_every": 0,
32
+ "gsm8k_n": 256,
33
+ "gsm8k_batch": 16,
34
+ "gsm8k_max_new_tokens": 512,
35
+ "gsm8k_frames": "chat",
36
+ "save_every": 50,
37
+ "out_dir": "outputs/healed/grid_math/reap_keep25_s1225",
38
+ "sweep": 150,
39
+ "wandb": true,
40
+ "wandb_project": "glean-grid",
41
+ "wandb_run_name": "reap-math-keep25-s1225",
42
+ "wandb_run_id": null,
43
+ "wandb_resume": null,
44
+ "wandb_mode": "online",
45
+ "no_wandb_sync": false,
46
+ "debug": false,
47
+ "resume_from": null,
48
+ "start_step": 0,
49
+ "no_grad_checkpointing": false,
50
+ "seed": 1225,
51
+ "no_teacher_overlap": false,
52
+ "sync_checkpoints": false,
53
+ "rollout_engine": "hf",
54
+ "vllm_gpu": null,
55
+ "vllm_port": 8377,
56
+ "vllm_refresh_every": 5,
57
+ "vllm_serve_bin": "vllm-plugin/.venv/bin/python",
58
+ "vllm_gpu_mem_util": 0.85,
59
+ "vllm_refresh_mode": "reload",
60
+ "vllm_live_dir": null,
61
+ "resolved_kl_direction": "forward"
62
+ }
healed/grid_math/reap_keep25_s1225/train_log.jsonl ADDED
@@ -0,0 +1,150 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"step": 1, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 5.5052759854316715, "tokens": 120000, "cumulative_loss_tokens": 120000, "grad_norm": 91.0, "lr": 6e-06, "finish_rate": 0.733, "comp_len": 628.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.5, "frames": {"chat": 191}, "mem_gb": 9.93}
2
+ {"step": 2, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 5.34218070195516, "tokens": 120000, "cumulative_loss_tokens": 240000, "grad_norm": 72.5, "lr": 9e-06, "finish_rate": 0.845, "comp_len": 547.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.7, "frames": {"chat": 219}, "mem_gb": 9.99}
3
+ {"step": 3, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 5.2662361368576684, "tokens": 120000, "cumulative_loss_tokens": 360000, "grad_norm": 54.5, "lr": 1.2e-05, "finish_rate": 0.778, "comp_len": 579.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.4, "frames": {"chat": 207}, "mem_gb": 10.0}
4
+ {"step": 4, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 4.323533807410796, "tokens": 120000, "cumulative_loss_tokens": 480000, "grad_norm": 49.75, "lr": 1.5e-05, "finish_rate": 0.755, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.4, "frames": {"chat": 208}, "mem_gb": 9.95}
5
+ {"step": 5, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 3.656937780322631, "tokens": 120000, "cumulative_loss_tokens": 600000, "grad_norm": 42.5, "lr": 1.8e-05, "finish_rate": 0.799, "comp_len": 547.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.2, "frames": {"chat": 219}, "mem_gb": 9.99}
6
+ {"step": 6, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 3.098056636095047, "tokens": 120000, "cumulative_loss_tokens": 720000, "grad_norm": 32.25, "lr": 2.1e-05, "finish_rate": 0.915, "comp_len": 487.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.9, "frames": {"chat": 246}, "mem_gb": 9.87}
7
+ {"step": 7, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 2.628579444358746, "tokens": 120000, "cumulative_loss_tokens": 840000, "grad_norm": 41.0, "lr": 2.4e-05, "finish_rate": 0.704, "comp_len": 582.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.1, "frames": {"chat": 206}, "mem_gb": 10.02}
8
+ {"step": 8, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 2.0468826131433246, "tokens": 120000, "cumulative_loss_tokens": 960000, "grad_norm": 26.75, "lr": 2.7000000000000002e-05, "finish_rate": 0.876, "comp_len": 515.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.1, "frames": {"chat": 233}, "mem_gb": 10.0}
9
+ {"step": 9, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 1.5576237986435493, "tokens": 120000, "cumulative_loss_tokens": 1080000, "grad_norm": 16.125, "lr": 3e-05, "finish_rate": 0.847, "comp_len": 524.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.7, "frames": {"chat": 229}, "mem_gb": 9.86}
10
+ {"step": 10, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 1.2283661879966656, "tokens": 120000, "cumulative_loss_tokens": 1200000, "grad_norm": 9.4375, "lr": 3e-05, "finish_rate": 0.864, "comp_len": 508.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.2, "frames": {"chat": 236}, "mem_gb": 9.9}
11
+ {"step": 11, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.9977921682407459, "tokens": 120000, "cumulative_loss_tokens": 1320000, "grad_norm": 6.375, "lr": 3e-05, "finish_rate": 0.87, "comp_len": 502.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.2, "frames": {"chat": 239}, "mem_gb": 9.78}
12
+ {"step": 12, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.8474529584680994, "tokens": 120000, "cumulative_loss_tokens": 1440000, "grad_norm": 6.25, "lr": 3e-05, "finish_rate": 0.867, "comp_len": 497.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.9, "frames": {"chat": 241}, "mem_gb": 9.9}
13
+ {"step": 13, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.7915342311960956, "tokens": 120000, "cumulative_loss_tokens": 1560000, "grad_norm": 15.375, "lr": 3e-05, "finish_rate": 0.863, "comp_len": 531.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.8, "frames": {"chat": 226}, "mem_gb": 9.87}
14
+ {"step": 14, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.6635135169697305, "tokens": 120000, "cumulative_loss_tokens": 1680000, "grad_norm": 3.984375, "lr": 3e-05, "finish_rate": 0.893, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.7, "frames": {"chat": 234}, "mem_gb": 10.0}
15
+ {"step": 15, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.6129945642106235, "tokens": 120000, "cumulative_loss_tokens": 1800000, "grad_norm": 2.453125, "lr": 3e-05, "finish_rate": 0.914, "comp_len": 466.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.2, "frames": {"chat": 257}, "mem_gb": 9.99}
16
+ {"step": 16, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.714731409107397, "tokens": 120000, "cumulative_loss_tokens": 1920000, "grad_norm": 3.296875, "lr": 3e-05, "finish_rate": 0.76, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.6, "frames": {"chat": 208}, "mem_gb": 10.04}
17
+ {"step": 17, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.5896475110622744, "tokens": 120000, "cumulative_loss_tokens": 2040000, "grad_norm": 5.53125, "lr": 3e-05, "finish_rate": 0.763, "comp_len": 568.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.6, "frames": {"chat": 211}, "mem_gb": 10.02}
18
+ {"step": 18, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.517651187770317, "tokens": 120000, "cumulative_loss_tokens": 2160000, "grad_norm": 1.390625, "lr": 3e-05, "finish_rate": 0.806, "comp_len": 528.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.0, "frames": {"chat": 227}, "mem_gb": 10.0}
19
+ {"step": 19, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.5277626443808278, "tokens": 120000, "cumulative_loss_tokens": 2280000, "grad_norm": 1.3984375, "lr": 3e-05, "finish_rate": 0.796, "comp_len": 568.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.0, "frames": {"chat": 211}, "mem_gb": 9.98}
20
+ {"step": 20, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.4388222487750153, "tokens": 120000, "cumulative_loss_tokens": 2400000, "grad_norm": 1.1875, "lr": 3e-05, "finish_rate": 0.861, "comp_len": 504.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.8, "frames": {"chat": 238}, "mem_gb": 9.99}
21
+ {"step": 21, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.42888357511336606, "tokens": 120000, "cumulative_loss_tokens": 2520000, "grad_norm": 1.046875, "lr": 3e-05, "finish_rate": 0.835, "comp_len": 506.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.7, "frames": {"chat": 237}, "mem_gb": 10.03}
22
+ {"step": 22, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.44955972477830947, "tokens": 120000, "cumulative_loss_tokens": 2640000, "grad_norm": 1.171875, "lr": 3e-05, "finish_rate": 0.721, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.5, "frames": {"chat": 208}, "mem_gb": 10.03}
23
+ {"step": 23, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.40363404167294503, "tokens": 120000, "cumulative_loss_tokens": 2760000, "grad_norm": 0.9921875, "lr": 3e-05, "finish_rate": 0.801, "comp_len": 543.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.8, "frames": {"chat": 221}, "mem_gb": 10.12}
24
+ {"step": 24, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.36760696430206297, "tokens": 120000, "cumulative_loss_tokens": 2880000, "grad_norm": 0.8828125, "lr": 3e-05, "finish_rate": 0.853, "comp_len": 517.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.4, "frames": {"chat": 232}, "mem_gb": 9.96}
25
+ {"step": 25, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.3675604015878091, "tokens": 120000, "cumulative_loss_tokens": 3000000, "grad_norm": 0.8515625, "lr": 3e-05, "finish_rate": 0.764, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.1, "frames": {"chat": 208}, "mem_gb": 9.99}
26
+ {"step": 26, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.3380553434039156, "tokens": 120000, "cumulative_loss_tokens": 3120000, "grad_norm": 0.7734375, "lr": 3e-05, "finish_rate": 0.837, "comp_len": 528.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.2, "frames": {"chat": 227}, "mem_gb": 9.91}
27
+ {"step": 27, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.3542323801631729, "tokens": 120000, "cumulative_loss_tokens": 3240000, "grad_norm": 0.7578125, "lr": 3e-05, "finish_rate": 0.824, "comp_len": 543.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.9, "frames": {"chat": 221}, "mem_gb": 9.94}
28
+ {"step": 28, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.3140212787228326, "tokens": 120000, "cumulative_loss_tokens": 3360000, "grad_norm": 0.7421875, "lr": 3e-05, "finish_rate": 0.815, "comp_len": 517.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.5, "frames": {"chat": 232}, "mem_gb": 10.0}
29
+ {"step": 29, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.2983292136088014, "tokens": 120000, "cumulative_loss_tokens": 3480000, "grad_norm": 0.71875, "lr": 3e-05, "finish_rate": 0.822, "comp_len": 547.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.8, "frames": {"chat": 219}, "mem_gb": 10.0}
30
+ {"step": 30, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.3193080648737649, "tokens": 120000, "cumulative_loss_tokens": 3600000, "grad_norm": 0.70703125, "lr": 3e-05, "finish_rate": 0.713, "comp_len": 615.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.8, "frames": {"chat": 195}, "mem_gb": 10.09}
31
+ {"step": 31, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.31196173311385017, "tokens": 120000, "cumulative_loss_tokens": 3720000, "grad_norm": 0.80859375, "lr": 3e-05, "finish_rate": 0.833, "comp_len": 555.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.9, "frames": {"chat": 216}, "mem_gb": 10.0}
32
+ {"step": 32, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.2535243794289728, "tokens": 120000, "cumulative_loss_tokens": 3840000, "grad_norm": 0.60546875, "lr": 3e-05, "finish_rate": 0.788, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.9, "frames": {"chat": 208}, "mem_gb": 9.88}
33
+ {"step": 33, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.263915772453323, "tokens": 120000, "cumulative_loss_tokens": 3960000, "grad_norm": 0.70703125, "lr": 3e-05, "finish_rate": 0.919, "comp_len": 510.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.9, "frames": {"chat": 235}, "mem_gb": 9.88}
34
+ {"step": 34, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.27100975296981633, "tokens": 120000, "cumulative_loss_tokens": 4080000, "grad_norm": 0.7265625, "lr": 3e-05, "finish_rate": 0.853, "comp_len": 533.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.9, "frames": {"chat": 225}, "mem_gb": 9.99}
35
+ {"step": 35, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.34947378008762997, "tokens": 120000, "cumulative_loss_tokens": 4200000, "grad_norm": 0.76953125, "lr": 3e-05, "finish_rate": 0.77, "comp_len": 563.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.2, "frames": {"chat": 213}, "mem_gb": 10.08}
36
+ {"step": 36, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.2621212104951342, "tokens": 120000, "cumulative_loss_tokens": 4320000, "grad_norm": 0.71875, "lr": 3e-05, "finish_rate": 0.922, "comp_len": 466.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.1, "frames": {"chat": 257}, "mem_gb": 9.75}
37
+ {"step": 37, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.2858927722416818, "tokens": 120000, "cumulative_loss_tokens": 4440000, "grad_norm": 0.6640625, "lr": 3e-05, "finish_rate": 0.792, "comp_len": 566.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.6, "frames": {"chat": 212}, "mem_gb": 10.02}
38
+ {"step": 38, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.24749040407985448, "tokens": 120000, "cumulative_loss_tokens": 4560000, "grad_norm": 0.55859375, "lr": 3e-05, "finish_rate": 0.833, "comp_len": 543.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.5, "frames": {"chat": 221}, "mem_gb": 10.0}
39
+ {"step": 39, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.24993702379036695, "tokens": 120000, "cumulative_loss_tokens": 4680000, "grad_norm": 0.6328125, "lr": 3e-05, "finish_rate": 0.868, "comp_len": 495.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.7, "frames": {"chat": 242}, "mem_gb": 9.99}
40
+ {"step": 40, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.22095589282835523, "tokens": 120000, "cumulative_loss_tokens": 4800000, "grad_norm": 0.5390625, "lr": 3e-05, "finish_rate": 0.836, "comp_len": 545.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.1, "frames": {"chat": 220}, "mem_gb": 9.96}
41
+ {"step": 41, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.24182025297402093, "tokens": 120000, "cumulative_loss_tokens": 4920000, "grad_norm": 0.62890625, "lr": 3e-05, "finish_rate": 0.896, "comp_len": 500.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.0, "frames": {"chat": 240}, "mem_gb": 9.85}
42
+ {"step": 42, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.24479971672501416, "tokens": 120000, "cumulative_loss_tokens": 5040000, "grad_norm": 0.60546875, "lr": 3e-05, "finish_rate": 0.728, "comp_len": 582.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.6, "frames": {"chat": 206}, "mem_gb": 9.98}
43
+ {"step": 43, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.2458774261167273, "tokens": 120000, "cumulative_loss_tokens": 5160000, "grad_norm": 0.640625, "lr": 3e-05, "finish_rate": 0.867, "comp_len": 531.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.1, "frames": {"chat": 226}, "mem_gb": 10.0}
44
+ {"step": 44, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.31763079155124724, "tokens": 120000, "cumulative_loss_tokens": 5280000, "grad_norm": 1.1640625, "lr": 3e-05, "finish_rate": 0.877, "comp_len": 491.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.2, "frames": {"chat": 244}, "mem_gb": 9.78}
45
+ {"step": 45, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.26981097101569174, "tokens": 120000, "cumulative_loss_tokens": 5400000, "grad_norm": 0.6015625, "lr": 3e-05, "finish_rate": 0.804, "comp_len": 535.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.6, "frames": {"chat": 224}, "mem_gb": 10.0}
46
+ {"step": 46, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.2382314425634841, "tokens": 120000, "cumulative_loss_tokens": 5520000, "grad_norm": 0.625, "lr": 3e-05, "finish_rate": 0.923, "comp_len": 442.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 47.1, "frames": {"chat": 271}, "mem_gb": 9.72}
47
+ {"step": 47, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.21951183261151114, "tokens": 120000, "cumulative_loss_tokens": 5640000, "grad_norm": 0.546875, "lr": 3e-05, "finish_rate": 0.856, "comp_len": 508.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.6, "frames": {"chat": 236}, "mem_gb": 10.01}
48
+ {"step": 48, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.24863433394009868, "tokens": 120000, "cumulative_loss_tokens": 5760000, "grad_norm": 0.63671875, "lr": 3e-05, "finish_rate": 0.841, "comp_len": 517.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.5, "frames": {"chat": 232}, "mem_gb": 9.88}
49
+ {"step": 49, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.19559897173379237, "tokens": 120000, "cumulative_loss_tokens": 5880000, "grad_norm": 0.50390625, "lr": 3e-05, "finish_rate": 0.79, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.5, "frames": {"chat": 210}, "mem_gb": 9.93}
50
+ {"step": 50, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.18844997002581756, "tokens": 120000, "cumulative_loss_tokens": 6000000, "grad_norm": 0.470703125, "lr": 3e-05, "finish_rate": 0.811, "comp_len": 553.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.2, "frames": {"chat": 217}, "mem_gb": 9.9}
51
+ {"step": 51, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.25687755517810584, "tokens": 120000, "cumulative_loss_tokens": 6120000, "grad_norm": 0.58203125, "lr": 3e-05, "finish_rate": 0.839, "comp_len": 535.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.9, "frames": {"chat": 224}, "mem_gb": 10.02}
52
+ {"step": 52, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.28079532670788465, "tokens": 120000, "cumulative_loss_tokens": 6240000, "grad_norm": 0.61328125, "lr": 3e-05, "finish_rate": 0.749, "comp_len": 591.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.1, "frames": {"chat": 203}, "mem_gb": 9.87}
53
+ {"step": 53, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.2312026971814533, "tokens": 120000, "cumulative_loss_tokens": 6360000, "grad_norm": 0.5546875, "lr": 3e-05, "finish_rate": 0.887, "comp_len": 502.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.4, "frames": {"chat": 239}, "mem_gb": 9.97}
54
+ {"step": 54, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.19427155094649642, "tokens": 120000, "cumulative_loss_tokens": 6480000, "grad_norm": 0.5, "lr": 3e-05, "finish_rate": 0.902, "comp_len": 472.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.4, "frames": {"chat": 254}, "mem_gb": 9.88}
55
+ {"step": 55, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.18302697168818363, "tokens": 120000, "cumulative_loss_tokens": 6600000, "grad_norm": 0.474609375, "lr": 3e-05, "finish_rate": 0.876, "comp_len": 497.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.5, "frames": {"chat": 241}, "mem_gb": 9.97}
56
+ {"step": 56, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.24095728847576925, "tokens": 120000, "cumulative_loss_tokens": 6720000, "grad_norm": 0.51953125, "lr": 3e-05, "finish_rate": 0.746, "comp_len": 563.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.0, "frames": {"chat": 213}, "mem_gb": 10.0}
57
+ {"step": 57, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2061793549572428, "tokens": 120000, "cumulative_loss_tokens": 6840000, "grad_norm": 0.74609375, "lr": 3e-05, "finish_rate": 0.864, "comp_len": 543.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.2, "frames": {"chat": 221}, "mem_gb": 10.05}
58
+ {"step": 58, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.20495547246101, "tokens": 120000, "cumulative_loss_tokens": 6960000, "grad_norm": 0.48046875, "lr": 3e-05, "finish_rate": 0.745, "comp_len": 612.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.3, "frames": {"chat": 196}, "mem_gb": 10.01}
59
+ {"step": 59, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.17600186436952403, "tokens": 120000, "cumulative_loss_tokens": 7080000, "grad_norm": 0.498046875, "lr": 3e-05, "finish_rate": 0.926, "comp_len": 444.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 48.0, "frames": {"chat": 270}, "mem_gb": 9.81}
60
+ {"step": 60, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.15102526473104955, "tokens": 120000, "cumulative_loss_tokens": 7200000, "grad_norm": 0.4140625, "lr": 3e-05, "finish_rate": 0.815, "comp_len": 555.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.6, "frames": {"chat": 216}, "mem_gb": 9.99}
61
+ {"step": 61, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.21443645971684405, "tokens": 120000, "cumulative_loss_tokens": 7320000, "grad_norm": 0.57421875, "lr": 3e-05, "finish_rate": 0.775, "comp_len": 600.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 40.6, "frames": {"chat": 200}, "mem_gb": 9.96}
62
+ {"step": 62, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.1430999824684734, "tokens": 120000, "cumulative_loss_tokens": 7440000, "grad_norm": 0.419921875, "lr": 3e-05, "finish_rate": 0.767, "comp_len": 582.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.9, "frames": {"chat": 206}, "mem_gb": 9.91}
63
+ {"step": 63, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.1475067599070569, "tokens": 120000, "cumulative_loss_tokens": 7560000, "grad_norm": 0.40234375, "lr": 3e-05, "finish_rate": 0.902, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.9, "frames": {"chat": 234}, "mem_gb": 9.94}
64
+ {"step": 64, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.17827725302210698, "tokens": 120000, "cumulative_loss_tokens": 7680000, "grad_norm": 0.439453125, "lr": 3e-05, "finish_rate": 0.823, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.3, "frames": {"chat": 215}, "mem_gb": 9.95}
65
+ {"step": 65, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.15364940941271682, "tokens": 120000, "cumulative_loss_tokens": 7800000, "grad_norm": 0.494140625, "lr": 3e-05, "finish_rate": 0.922, "comp_len": 470.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.8, "frames": {"chat": 255}, "mem_gb": 9.94}
66
+ {"step": 66, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.18826355347931384, "tokens": 120000, "cumulative_loss_tokens": 7920000, "grad_norm": 0.546875, "lr": 3e-05, "finish_rate": 0.892, "comp_len": 480.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.2, "frames": {"chat": 250}, "mem_gb": 9.82}
67
+ {"step": 67, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.17189633793290704, "tokens": 120000, "cumulative_loss_tokens": 8040000, "grad_norm": 0.4375, "lr": 3e-05, "finish_rate": 0.884, "comp_len": 495.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.5, "frames": {"chat": 242}, "mem_gb": 9.99}
68
+ {"step": 68, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2132726859041179, "tokens": 120000, "cumulative_loss_tokens": 8160000, "grad_norm": 0.5234375, "lr": 3e-05, "finish_rate": 0.729, "comp_len": 603.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.8, "frames": {"chat": 199}, "mem_gb": 10.0}
69
+ {"step": 69, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.24157150672028463, "tokens": 120000, "cumulative_loss_tokens": 8280000, "grad_norm": 0.578125, "lr": 3e-05, "finish_rate": 0.784, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.0, "frames": {"chat": 208}, "mem_gb": 10.03}
70
+ {"step": 70, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.18809422614208113, "tokens": 120000, "cumulative_loss_tokens": 8400000, "grad_norm": 0.515625, "lr": 3e-05, "finish_rate": 0.764, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.3, "frames": {"chat": 208}, "mem_gb": 9.97}
71
+ {"step": 71, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.19989227577894927, "tokens": 120000, "cumulative_loss_tokens": 8520000, "grad_norm": 0.4765625, "lr": 3e-05, "finish_rate": 0.732, "comp_len": 574.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.1, "frames": {"chat": 209}, "mem_gb": 10.12}
72
+ {"step": 72, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.1725833379857863, "tokens": 120000, "cumulative_loss_tokens": 8640000, "grad_norm": 0.484375, "lr": 3e-05, "finish_rate": 0.855, "comp_len": 510.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.9, "frames": {"chat": 235}, "mem_gb": 9.95}
73
+ {"step": 73, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.1478453546665609, "tokens": 120000, "cumulative_loss_tokens": 8760000, "grad_norm": 0.41015625, "lr": 3e-05, "finish_rate": 0.74, "comp_len": 588.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.8, "frames": {"chat": 204}, "mem_gb": 9.94}
74
+ {"step": 74, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.21500386069975794, "tokens": 120000, "cumulative_loss_tokens": 8880000, "grad_norm": 0.5, "lr": 3e-05, "finish_rate": 0.745, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.1, "frames": {"chat": 208}, "mem_gb": 10.0}
75
+ {"step": 75, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.1554392906052371, "tokens": 120000, "cumulative_loss_tokens": 9000000, "grad_norm": 0.44140625, "lr": 3e-05, "finish_rate": 0.825, "comp_len": 500.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.9, "frames": {"chat": 240}, "mem_gb": 10.0}
76
+ {"step": 76, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.1832204219336311, "tokens": 120000, "cumulative_loss_tokens": 9120000, "grad_norm": 0.50390625, "lr": 3e-05, "finish_rate": 0.89, "comp_len": 487.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.4, "frames": {"chat": 246}, "mem_gb": 9.99}
77
+ {"step": 77, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.17209678962721178, "tokens": 120000, "cumulative_loss_tokens": 9240000, "grad_norm": 0.4453125, "lr": 3e-05, "finish_rate": 0.909, "comp_len": 493.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.1, "frames": {"chat": 243}, "mem_gb": 9.81}
78
+ {"step": 78, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.20465335384116817, "tokens": 120000, "cumulative_loss_tokens": 9360000, "grad_norm": 0.5078125, "lr": 3e-05, "finish_rate": 0.745, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.8, "frames": {"chat": 208}, "mem_gb": 10.01}
79
+ {"step": 79, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.1694449459930261, "tokens": 120000, "cumulative_loss_tokens": 9480000, "grad_norm": 0.3984375, "lr": 3e-05, "finish_rate": 0.817, "comp_len": 547.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.7, "frames": {"chat": 219}, "mem_gb": 10.0}
80
+ {"step": 80, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.22224440845083446, "tokens": 120000, "cumulative_loss_tokens": 9600000, "grad_norm": 0.5234375, "lr": 3e-05, "finish_rate": 0.782, "comp_len": 568.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.6, "frames": {"chat": 211}, "mem_gb": 10.01}
81
+ {"step": 81, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.1509838725623054, "tokens": 120000, "cumulative_loss_tokens": 9720000, "grad_norm": 0.41015625, "lr": 3e-05, "finish_rate": 0.862, "comp_len": 517.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.9, "frames": {"chat": 232}, "mem_gb": 9.97}
82
+ {"step": 82, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.17593013638195892, "tokens": 120000, "cumulative_loss_tokens": 9840000, "grad_norm": 0.47265625, "lr": 3e-05, "finish_rate": 0.804, "comp_len": 560.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.9, "frames": {"chat": 214}, "mem_gb": 10.0}
83
+ {"step": 83, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.15619046537938217, "tokens": 120000, "cumulative_loss_tokens": 9960000, "grad_norm": 0.39453125, "lr": 3e-05, "finish_rate": 0.819, "comp_len": 531.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.5, "frames": {"chat": 226}, "mem_gb": 9.89}
84
+ {"step": 84, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.17479738595858216, "tokens": 120000, "cumulative_loss_tokens": 10080000, "grad_norm": 0.4609375, "lr": 3e-05, "finish_rate": 0.819, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.1, "frames": {"chat": 210}, "mem_gb": 10.01}
85
+ {"step": 85, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.1404316657436701, "tokens": 120000, "cumulative_loss_tokens": 10200000, "grad_norm": 0.404296875, "lr": 3e-05, "finish_rate": 0.858, "comp_len": 550.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.9, "frames": {"chat": 218}, "mem_gb": 9.83}
86
+ {"step": 86, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.1403924273949427, "tokens": 120000, "cumulative_loss_tokens": 10320000, "grad_norm": 0.3828125, "lr": 3e-05, "finish_rate": 0.858, "comp_len": 515.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.0, "frames": {"chat": 233}, "mem_gb": 9.98}
87
+ {"step": 87, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.17982150458991528, "tokens": 120000, "cumulative_loss_tokens": 10440000, "grad_norm": 0.427734375, "lr": 3e-05, "finish_rate": 0.786, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.8, "frames": {"chat": 215}, "mem_gb": 10.0}
88
+ {"step": 88, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.18029257206898183, "tokens": 120000, "cumulative_loss_tokens": 10560000, "grad_norm": 0.462890625, "lr": 3e-05, "finish_rate": 0.845, "comp_len": 515.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.0, "frames": {"chat": 233}, "mem_gb": 9.99}
89
+ {"step": 89, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.14494490633426854, "tokens": 120000, "cumulative_loss_tokens": 10680000, "grad_norm": 0.380859375, "lr": 3e-05, "finish_rate": 0.766, "comp_len": 574.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.6, "frames": {"chat": 209}, "mem_gb": 9.94}
90
+ {"step": 90, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.15547993779405953, "tokens": 120000, "cumulative_loss_tokens": 10800000, "grad_norm": 0.4453125, "lr": 3e-05, "finish_rate": 0.908, "comp_len": 458.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.3, "frames": {"chat": 262}, "mem_gb": 9.87}
91
+ {"step": 91, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.13846087997810294, "tokens": 120000, "cumulative_loss_tokens": 10920000, "grad_norm": 0.384765625, "lr": 3e-05, "finish_rate": 0.9, "comp_len": 481.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.2, "frames": {"chat": 249}, "mem_gb": 9.96}
92
+ {"step": 92, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.20217994603576758, "tokens": 120000, "cumulative_loss_tokens": 11040000, "grad_norm": 0.4765625, "lr": 3e-05, "finish_rate": 0.819, "comp_len": 528.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.6, "frames": {"chat": 227}, "mem_gb": 9.99}
93
+ {"step": 93, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.14583957826520005, "tokens": 120000, "cumulative_loss_tokens": 11160000, "grad_norm": 0.388671875, "lr": 3e-05, "finish_rate": 0.814, "comp_len": 543.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.3, "frames": {"chat": 221}, "mem_gb": 9.99}
94
+ {"step": 94, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.1660599309977765, "tokens": 120000, "cumulative_loss_tokens": 11280000, "grad_norm": 0.4453125, "lr": 3e-05, "finish_rate": 0.859, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.1, "frames": {"chat": 234}, "mem_gb": 10.01}
95
+ {"step": 95, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.12514705213187263, "tokens": 120000, "cumulative_loss_tokens": 11400000, "grad_norm": 0.365234375, "lr": 3e-05, "finish_rate": 0.817, "comp_len": 563.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.6, "frames": {"chat": 213}, "mem_gb": 9.95}
96
+ {"step": 96, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.12465972695561747, "tokens": 120000, "cumulative_loss_tokens": 11520000, "grad_norm": 0.349609375, "lr": 3e-05, "finish_rate": 0.836, "comp_len": 563.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.9, "frames": {"chat": 213}, "mem_gb": 9.89}
97
+ {"step": 97, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.15661951549227038, "tokens": 120000, "cumulative_loss_tokens": 11640000, "grad_norm": 0.443359375, "lr": 3e-05, "finish_rate": 0.906, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.8, "frames": {"chat": 234}, "mem_gb": 9.92}
98
+ {"step": 98, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.14978609310248867, "tokens": 120000, "cumulative_loss_tokens": 11760000, "grad_norm": 0.416015625, "lr": 3e-05, "finish_rate": 0.793, "comp_len": 540.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.3, "frames": {"chat": 222}, "mem_gb": 9.99}
99
+ {"step": 99, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.16640961758115638, "tokens": 120000, "cumulative_loss_tokens": 11880000, "grad_norm": 0.419921875, "lr": 3e-05, "finish_rate": 0.806, "comp_len": 528.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 48.3, "frames": {"chat": 227}, "mem_gb": 10.0}
100
+ {"step": 100, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.16042816521745795, "tokens": 120000, "cumulative_loss_tokens": 12000000, "grad_norm": 0.427734375, "lr": 3e-05, "finish_rate": 0.835, "comp_len": 550.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.1, "frames": {"chat": 218}, "mem_gb": 10.04}
101
+ {"step": 101, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.17738757199452568, "tokens": 120000, "cumulative_loss_tokens": 12120000, "grad_norm": 0.44921875, "lr": 3e-05, "finish_rate": 0.798, "comp_len": 538.1, "t_data_s": 0.1, "t_rollout_s": 0.0, "t_step_s": 43.9, "frames": {"chat": 223}, "mem_gb": 10.01}
102
+ {"step": 102, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.18987772140322873, "tokens": 120000, "cumulative_loss_tokens": 12240000, "grad_norm": 0.45703125, "lr": 3e-05, "finish_rate": 0.772, "comp_len": 582.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.6, "frames": {"chat": 206}, "mem_gb": 10.0}
103
+ {"step": 103, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.15001419565274068, "tokens": 120000, "cumulative_loss_tokens": 12360000, "grad_norm": 0.41015625, "lr": 3e-05, "finish_rate": 0.784, "comp_len": 563.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.8, "frames": {"chat": 213}, "mem_gb": 9.92}
104
+ {"step": 104, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.19792243093705425, "tokens": 120000, "cumulative_loss_tokens": 12480000, "grad_norm": 0.451171875, "lr": 3e-05, "finish_rate": 0.843, "comp_len": 538.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.1, "frames": {"chat": 223}, "mem_gb": 9.86}
105
+ {"step": 105, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.15784500931380316, "tokens": 120000, "cumulative_loss_tokens": 12600000, "grad_norm": 0.3984375, "lr": 3e-05, "finish_rate": 0.828, "comp_len": 528.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.3, "frames": {"chat": 227}, "mem_gb": 9.97}
106
+ {"step": 106, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.15715095629387846, "tokens": 120000, "cumulative_loss_tokens": 12720000, "grad_norm": 0.4375, "lr": 3e-05, "finish_rate": 0.889, "comp_len": 474.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.9, "frames": {"chat": 253}, "mem_gb": 10.0}
107
+ {"step": 107, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.16428898396513736, "tokens": 120000, "cumulative_loss_tokens": 12840000, "grad_norm": 0.48828125, "lr": 3e-05, "finish_rate": 0.792, "comp_len": 555.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.4, "frames": {"chat": 216}, "mem_gb": 10.0}
108
+ {"step": 108, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.13419136733350653, "tokens": 120000, "cumulative_loss_tokens": 12960000, "grad_norm": 0.39453125, "lr": 3e-05, "finish_rate": 0.766, "comp_len": 585.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.2, "frames": {"chat": 205}, "mem_gb": 9.97}
109
+ {"step": 109, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.17433834038786591, "tokens": 120000, "cumulative_loss_tokens": 13080000, "grad_norm": 0.41796875, "lr": 3e-05, "finish_rate": 0.729, "comp_len": 579.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.7, "frames": {"chat": 207}, "mem_gb": 10.06}
110
+ {"step": 110, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.14855496705075105, "tokens": 120000, "cumulative_loss_tokens": 13200000, "grad_norm": 0.388671875, "lr": 3e-05, "finish_rate": 0.814, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.6, "frames": {"chat": 215}, "mem_gb": 9.98}
111
+ {"step": 111, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.11281266368478536, "tokens": 120000, "cumulative_loss_tokens": 13320000, "grad_norm": 0.43359375, "lr": 3e-05, "finish_rate": 0.86, "comp_len": 526.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.7, "frames": {"chat": 228}, "mem_gb": 10.0}
112
+ {"step": 112, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.14154021223733823, "tokens": 120000, "cumulative_loss_tokens": 13440000, "grad_norm": 0.4453125, "lr": 3e-05, "finish_rate": 0.747, "comp_len": 543.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.8, "frames": {"chat": 221}, "mem_gb": 10.04}
113
+ {"step": 113, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.11087060506536314, "tokens": 120000, "cumulative_loss_tokens": 13560000, "grad_norm": 0.36328125, "lr": 3e-05, "finish_rate": 0.882, "comp_len": 472.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.2, "frames": {"chat": 254}, "mem_gb": 9.84}
114
+ {"step": 114, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.09634949580542743, "tokens": 120000, "cumulative_loss_tokens": 13680000, "grad_norm": 0.37109375, "lr": 3e-05, "finish_rate": 0.843, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.5, "frames": {"chat": 210}, "mem_gb": 9.96}
115
+ {"step": 115, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.11305949703895797, "tokens": 120000, "cumulative_loss_tokens": 13800000, "grad_norm": 0.361328125, "lr": 3e-05, "finish_rate": 0.827, "comp_len": 531.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.1, "frames": {"chat": 226}, "mem_gb": 9.92}
116
+ {"step": 116, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.11892387241683901, "tokens": 120000, "cumulative_loss_tokens": 13920000, "grad_norm": 0.359375, "lr": 3e-05, "finish_rate": 0.802, "comp_len": 566.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.9, "frames": {"chat": 212}, "mem_gb": 9.99}
117
+ {"step": 117, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.14203277846990775, "tokens": 120000, "cumulative_loss_tokens": 14040000, "grad_norm": 0.388671875, "lr": 3e-05, "finish_rate": 0.754, "comp_len": 568.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.0, "frames": {"chat": 211}, "mem_gb": 9.92}
118
+ {"step": 118, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1415941094346655, "tokens": 120000, "cumulative_loss_tokens": 14160000, "grad_norm": 0.3984375, "lr": 3e-05, "finish_rate": 0.776, "comp_len": 612.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 40.4, "frames": {"chat": 196}, "mem_gb": 9.97}
119
+ {"step": 119, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.10169251247107362, "tokens": 120000, "cumulative_loss_tokens": 14280000, "grad_norm": 0.33203125, "lr": 3e-05, "finish_rate": 0.811, "comp_len": 566.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.5, "frames": {"chat": 212}, "mem_gb": 9.99}
120
+ {"step": 120, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.10335419424610834, "tokens": 120000, "cumulative_loss_tokens": 14400000, "grad_norm": 0.326171875, "lr": 3e-05, "finish_rate": 0.877, "comp_len": 491.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.8, "frames": {"chat": 244}, "mem_gb": 9.91}
121
+ {"step": 121, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1078583001211596, "tokens": 120000, "cumulative_loss_tokens": 14520000, "grad_norm": 0.333984375, "lr": 3e-05, "finish_rate": 0.838, "comp_len": 540.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.2, "frames": {"chat": 222}, "mem_gb": 9.95}
122
+ {"step": 122, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1223276870971856, "tokens": 120000, "cumulative_loss_tokens": 14640000, "grad_norm": 0.3671875, "lr": 3e-05, "finish_rate": 0.78, "comp_len": 550.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.0, "frames": {"chat": 218}, "mem_gb": 10.0}
123
+ {"step": 123, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.12229978259069224, "tokens": 120000, "cumulative_loss_tokens": 14760000, "grad_norm": 0.349609375, "lr": 3e-05, "finish_rate": 0.913, "comp_len": 476.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.8, "frames": {"chat": 252}, "mem_gb": 9.87}
124
+ {"step": 124, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1320776972546242, "tokens": 120000, "cumulative_loss_tokens": 14880000, "grad_norm": 0.3984375, "lr": 3e-05, "finish_rate": 0.728, "comp_len": 594.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.2, "frames": {"chat": 202}, "mem_gb": 10.05}
125
+ {"step": 125, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1445485326328315, "tokens": 120000, "cumulative_loss_tokens": 15000000, "grad_norm": 0.408203125, "lr": 3e-05, "finish_rate": 0.835, "comp_len": 506.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.8, "frames": {"chat": 237}, "mem_gb": 10.0}
126
+ {"step": 126, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1339294173414198, "tokens": 120000, "cumulative_loss_tokens": 15120000, "grad_norm": 0.388671875, "lr": 3e-05, "finish_rate": 0.868, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.6, "frames": {"chat": 234}, "mem_gb": 9.98}
127
+ {"step": 127, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.09875204861378298, "tokens": 120000, "cumulative_loss_tokens": 15240000, "grad_norm": 0.330078125, "lr": 3e-05, "finish_rate": 0.809, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.2, "frames": {"chat": 215}, "mem_gb": 10.0}
128
+ {"step": 128, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.10042380276505525, "tokens": 120000, "cumulative_loss_tokens": 15360000, "grad_norm": 0.333984375, "lr": 3e-05, "finish_rate": 0.876, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.8, "frames": {"chat": 234}, "mem_gb": 9.93}
129
+ {"step": 129, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.0915526939183784, "tokens": 120000, "cumulative_loss_tokens": 15480000, "grad_norm": 0.330078125, "lr": 3e-05, "finish_rate": 0.801, "comp_len": 555.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.1, "frames": {"chat": 216}, "mem_gb": 9.98}
130
+ {"step": 130, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.10634777345014736, "tokens": 120000, "cumulative_loss_tokens": 15600000, "grad_norm": 0.326171875, "lr": 3e-05, "finish_rate": 0.805, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.7, "frames": {"chat": 210}, "mem_gb": 9.95}
131
+ {"step": 131, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.12015755086155452, "tokens": 120000, "cumulative_loss_tokens": 15720000, "grad_norm": 0.353515625, "lr": 3e-05, "finish_rate": 0.719, "comp_len": 603.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.8, "frames": {"chat": 199}, "mem_gb": 10.0}
132
+ {"step": 132, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1101692976130483, "tokens": 120000, "cumulative_loss_tokens": 15840000, "grad_norm": 0.337890625, "lr": 3e-05, "finish_rate": 0.824, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.7, "frames": {"chat": 210}, "mem_gb": 10.01}
133
+ {"step": 133, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.11059571424775447, "tokens": 120000, "cumulative_loss_tokens": 15960000, "grad_norm": 0.357421875, "lr": 3e-05, "finish_rate": 0.902, "comp_len": 533.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.9, "frames": {"chat": 225}, "mem_gb": 9.95}
134
+ {"step": 134, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.10813538532971094, "tokens": 120000, "cumulative_loss_tokens": 16080000, "grad_norm": 0.328125, "lr": 3e-05, "finish_rate": 0.913, "comp_len": 474.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.3, "frames": {"chat": 253}, "mem_gb": 9.85}
135
+ {"step": 135, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.12084263433534652, "tokens": 120000, "cumulative_loss_tokens": 16200000, "grad_norm": 0.36328125, "lr": 3e-05, "finish_rate": 0.903, "comp_len": 485.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.0, "frames": {"chat": 247}, "mem_gb": 9.97}
136
+ {"step": 136, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.13201027247390398, "tokens": 120000, "cumulative_loss_tokens": 16320000, "grad_norm": 0.388671875, "lr": 3e-05, "finish_rate": 0.836, "comp_len": 504.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.3, "frames": {"chat": 238}, "mem_gb": 9.97}
137
+ {"step": 137, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.13818833304295938, "tokens": 120000, "cumulative_loss_tokens": 16440000, "grad_norm": 0.40625, "lr": 3e-05, "finish_rate": 0.86, "comp_len": 510.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.4, "frames": {"chat": 235}, "mem_gb": 9.99}
138
+ {"step": 138, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1345420230248322, "tokens": 120000, "cumulative_loss_tokens": 16560000, "grad_norm": 0.423828125, "lr": 3e-05, "finish_rate": 0.805, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.8, "frames": {"chat": 215}, "mem_gb": 9.96}
139
+ {"step": 139, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.10943617184103156, "tokens": 120000, "cumulative_loss_tokens": 16680000, "grad_norm": 0.365234375, "lr": 3e-05, "finish_rate": 0.925, "comp_len": 447.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.5, "frames": {"chat": 268}, "mem_gb": 9.97}
140
+ {"step": 140, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.12807247791942208, "tokens": 120000, "cumulative_loss_tokens": 16800000, "grad_norm": 0.380859375, "lr": 3e-05, "finish_rate": 0.825, "comp_len": 526.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.3, "frames": {"chat": 228}, "mem_gb": 10.0}
141
+ {"step": 141, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.12706402302297454, "tokens": 120000, "cumulative_loss_tokens": 16920000, "grad_norm": 0.38671875, "lr": 3e-05, "finish_rate": 0.881, "comp_len": 476.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.3, "frames": {"chat": 252}, "mem_gb": 9.93}
142
+ {"step": 142, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.12039303982133667, "tokens": 120000, "cumulative_loss_tokens": 17040000, "grad_norm": 0.36328125, "lr": 3e-05, "finish_rate": 0.821, "comp_len": 538.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.1, "frames": {"chat": 223}, "mem_gb": 10.01}
143
+ {"step": 143, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.14205731437917177, "tokens": 120000, "cumulative_loss_tokens": 17160000, "grad_norm": 0.3828125, "lr": 3e-05, "finish_rate": 0.805, "comp_len": 531.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.7, "frames": {"chat": 226}, "mem_gb": 9.99}
144
+ {"step": 144, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1410405628043848, "tokens": 120000, "cumulative_loss_tokens": 17280000, "grad_norm": 0.40625, "lr": 3e-05, "finish_rate": 0.731, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.9, "frames": {"chat": 208}, "mem_gb": 10.04}
145
+ {"step": 145, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.09851235140593101, "tokens": 120000, "cumulative_loss_tokens": 17400000, "grad_norm": 0.31640625, "lr": 3e-05, "finish_rate": 0.883, "comp_len": 500.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.6, "frames": {"chat": 240}, "mem_gb": 9.93}
146
+ {"step": 146, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.13880995498470342, "tokens": 120000, "cumulative_loss_tokens": 17520000, "grad_norm": 0.396484375, "lr": 3e-05, "finish_rate": 0.842, "comp_len": 540.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.3, "frames": {"chat": 222}, "mem_gb": 9.92}
147
+ {"step": 147, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.10342208843190843, "tokens": 120000, "cumulative_loss_tokens": 17640000, "grad_norm": 0.328125, "lr": 3e-05, "finish_rate": 0.881, "comp_len": 508.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.5, "frames": {"chat": 236}, "mem_gb": 9.99}
148
+ {"step": 148, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1016613077900062, "tokens": 120000, "cumulative_loss_tokens": 17760000, "grad_norm": 0.353515625, "lr": 3e-05, "finish_rate": 0.834, "comp_len": 553.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.6, "frames": {"chat": 217}, "mem_gb": 9.96}
149
+ {"step": 149, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.12922203347664327, "tokens": 120000, "cumulative_loss_tokens": 17880000, "grad_norm": 0.48046875, "lr": 3e-05, "finish_rate": 0.921, "comp_len": 476.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.9, "frames": {"chat": 252}, "mem_gb": 9.87}
150
+ {"step": 150, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.10404405277933304, "tokens": 120000, "cumulative_loss_tokens": 18000000, "grad_norm": 0.359375, "lr": 3e-05, "finish_rate": 0.847, "comp_len": 540.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.8, "frames": {"chat": 222}, "mem_gb": 9.99}
healed/grid_math/uniform_keep25_s1226/train_log.jsonl ADDED
@@ -0,0 +1,150 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"step": 1, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 1.2495578979462385, "tokens": 120000, "cumulative_loss_tokens": 120000, "grad_norm": 14.5625, "lr": 6e-06, "finish_rate": 0.902, "comp_len": 472.4, "t_data_s": 0.2, "t_rollout_s": 0.0, "t_step_s": 36.8, "frames": {"chat": 254}, "mem_gb": 9.82}
2
+ {"step": 2, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 1.2842795796026787, "tokens": 120000, "cumulative_loss_tokens": 240000, "grad_norm": 14.25, "lr": 9e-06, "finish_rate": 0.876, "comp_len": 497.9, "t_data_s": 0.1, "t_rollout_s": 0.0, "t_step_s": 29.4, "frames": {"chat": 241}, "mem_gb": 9.98}
3
+ {"step": 3, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 1.2860378812308113, "tokens": 120000, "cumulative_loss_tokens": 360000, "grad_norm": 13.3125, "lr": 1.2e-05, "finish_rate": 0.746, "comp_len": 563.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.7, "frames": {"chat": 213}, "mem_gb": 10.01}
4
+ {"step": 4, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 1.0296244965081414, "tokens": 120000, "cumulative_loss_tokens": 480000, "grad_norm": 8.625, "lr": 1.5e-05, "finish_rate": 0.864, "comp_len": 543.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.3, "frames": {"chat": 221}, "mem_gb": 10.05}
5
+ {"step": 5, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 1.0173526370272041, "tokens": 120000, "cumulative_loss_tokens": 600000, "grad_norm": 6.9375, "lr": 1.8e-05, "finish_rate": 0.745, "comp_len": 612.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.0, "frames": {"chat": 196}, "mem_gb": 10.01}
6
+ {"step": 6, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.7652708411070208, "tokens": 120000, "cumulative_loss_tokens": 720000, "grad_norm": 4.1875, "lr": 2.1e-05, "finish_rate": 0.926, "comp_len": 444.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 31.7, "frames": {"chat": 270}, "mem_gb": 9.82}
7
+ {"step": 7, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.7675839649726948, "tokens": 120000, "cumulative_loss_tokens": 840000, "grad_norm": 3.734375, "lr": 2.4e-05, "finish_rate": 0.815, "comp_len": 555.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.0, "frames": {"chat": 216}, "mem_gb": 10.0}
8
+ {"step": 8, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.7070816349441806, "tokens": 120000, "cumulative_loss_tokens": 960000, "grad_norm": 3.28125, "lr": 2.7000000000000002e-05, "finish_rate": 0.775, "comp_len": 600.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 26.7, "frames": {"chat": 200}, "mem_gb": 9.96}
9
+ {"step": 9, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.6169270259405176, "tokens": 120000, "cumulative_loss_tokens": 1080000, "grad_norm": 2.75, "lr": 3e-05, "finish_rate": 0.767, "comp_len": 582.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.5, "frames": {"chat": 206}, "mem_gb": 9.91}
10
+ {"step": 10, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.5323196953435739, "tokens": 120000, "cumulative_loss_tokens": 1200000, "grad_norm": 1.9609375, "lr": 3e-05, "finish_rate": 0.902, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.3, "frames": {"chat": 234}, "mem_gb": 9.95}
11
+ {"step": 11, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.5534217075144251, "tokens": 120000, "cumulative_loss_tokens": 1320000, "grad_norm": 1.640625, "lr": 3e-05, "finish_rate": 0.823, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.2, "frames": {"chat": 215}, "mem_gb": 9.96}
12
+ {"step": 12, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.470129220243295, "tokens": 120000, "cumulative_loss_tokens": 1440000, "grad_norm": 1.1953125, "lr": 3e-05, "finish_rate": 0.922, "comp_len": 470.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.1, "frames": {"chat": 255}, "mem_gb": 9.94}
13
+ {"step": 13, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.4660646371759474, "tokens": 120000, "cumulative_loss_tokens": 1560000, "grad_norm": 1.09375, "lr": 3e-05, "finish_rate": 0.892, "comp_len": 480.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 30.0, "frames": {"chat": 250}, "mem_gb": 9.82}
14
+ {"step": 14, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.4519369126672546, "tokens": 120000, "cumulative_loss_tokens": 1680000, "grad_norm": 1.0234375, "lr": 3e-05, "finish_rate": 0.884, "comp_len": 495.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.8, "frames": {"chat": 242}, "mem_gb": 10.0}
15
+ {"step": 15, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.48208508322536947, "tokens": 120000, "cumulative_loss_tokens": 1800000, "grad_norm": 0.97265625, "lr": 3e-05, "finish_rate": 0.729, "comp_len": 603.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.1, "frames": {"chat": 199}, "mem_gb": 10.0}
16
+ {"step": 16, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.505666505916665, "tokens": 120000, "cumulative_loss_tokens": 1920000, "grad_norm": 0.98046875, "lr": 3e-05, "finish_rate": 0.784, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.4, "frames": {"chat": 208}, "mem_gb": 10.04}
17
+ {"step": 17, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.42821578803832333, "tokens": 120000, "cumulative_loss_tokens": 2040000, "grad_norm": 1.0078125, "lr": 3e-05, "finish_rate": 0.764, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.3, "frames": {"chat": 208}, "mem_gb": 9.97}
18
+ {"step": 18, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.4488865850756566, "tokens": 120000, "cumulative_loss_tokens": 2160000, "grad_norm": 0.8359375, "lr": 3e-05, "finish_rate": 0.732, "comp_len": 574.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.5, "frames": {"chat": 209}, "mem_gb": 10.12}
19
+ {"step": 19, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.3713087936893105, "tokens": 120000, "cumulative_loss_tokens": 2280000, "grad_norm": 12.4375, "lr": 3e-05, "finish_rate": 0.855, "comp_len": 510.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.6, "frames": {"chat": 235}, "mem_gb": 9.96}
20
+ {"step": 20, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.3715890112740298, "tokens": 120000, "cumulative_loss_tokens": 2400000, "grad_norm": 0.80078125, "lr": 3e-05, "finish_rate": 0.74, "comp_len": 588.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.4, "frames": {"chat": 204}, "mem_gb": 9.95}
21
+ {"step": 21, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.44148291802903017, "tokens": 120000, "cumulative_loss_tokens": 2520000, "grad_norm": 0.9375, "lr": 3e-05, "finish_rate": 0.745, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.5, "frames": {"chat": 208}, "mem_gb": 10.01}
22
+ {"step": 22, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.36412479583832125, "tokens": 120000, "cumulative_loss_tokens": 2640000, "grad_norm": 0.80859375, "lr": 3e-05, "finish_rate": 0.825, "comp_len": 500.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 30.2, "frames": {"chat": 240}, "mem_gb": 10.0}
23
+ {"step": 23, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.35719255141379935, "tokens": 120000, "cumulative_loss_tokens": 2760000, "grad_norm": 0.69921875, "lr": 3e-05, "finish_rate": 0.89, "comp_len": 487.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.5, "frames": {"chat": 246}, "mem_gb": 9.99}
24
+ {"step": 24, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.3621041040317466, "tokens": 120000, "cumulative_loss_tokens": 2880000, "grad_norm": 0.67578125, "lr": 3e-05, "finish_rate": 0.909, "comp_len": 493.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.1, "frames": {"chat": 243}, "mem_gb": 9.82}
25
+ {"step": 25, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.38283104023163517, "tokens": 120000, "cumulative_loss_tokens": 3000000, "grad_norm": 0.68359375, "lr": 3e-05, "finish_rate": 0.745, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.5, "frames": {"chat": 208}, "mem_gb": 10.01}
26
+ {"step": 26, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.3491992857553065, "tokens": 120000, "cumulative_loss_tokens": 3120000, "grad_norm": 0.67578125, "lr": 3e-05, "finish_rate": 0.817, "comp_len": 547.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.1, "frames": {"chat": 219}, "mem_gb": 10.0}
27
+ {"step": 27, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.3949156841669232, "tokens": 120000, "cumulative_loss_tokens": 3240000, "grad_norm": 0.80859375, "lr": 3e-05, "finish_rate": 0.782, "comp_len": 568.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.0, "frames": {"chat": 211}, "mem_gb": 10.01}
28
+ {"step": 28, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.3246795868317286, "tokens": 120000, "cumulative_loss_tokens": 3360000, "grad_norm": 0.69140625, "lr": 3e-05, "finish_rate": 0.862, "comp_len": 517.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.0, "frames": {"chat": 232}, "mem_gb": 9.98}
29
+ {"step": 29, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.3552298259820789, "tokens": 120000, "cumulative_loss_tokens": 3480000, "grad_norm": 0.73046875, "lr": 3e-05, "finish_rate": 0.804, "comp_len": 560.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.1, "frames": {"chat": 214}, "mem_gb": 10.01}
30
+ {"step": 30, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.3321069305093338, "tokens": 120000, "cumulative_loss_tokens": 3600000, "grad_norm": 0.65625, "lr": 3e-05, "finish_rate": 0.819, "comp_len": 531.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.7, "frames": {"chat": 226}, "mem_gb": 9.9}
31
+ {"step": 31, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.32228971752213936, "tokens": 120000, "cumulative_loss_tokens": 3720000, "grad_norm": 0.64453125, "lr": 3e-05, "finish_rate": 0.819, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.2, "frames": {"chat": 210}, "mem_gb": 10.01}
32
+ {"step": 32, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.2798056041251868, "tokens": 120000, "cumulative_loss_tokens": 3840000, "grad_norm": 0.58203125, "lr": 3e-05, "finish_rate": 0.858, "comp_len": 550.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.4, "frames": {"chat": 218}, "mem_gb": 9.83}
33
+ {"step": 33, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.29189739949926735, "tokens": 120000, "cumulative_loss_tokens": 3960000, "grad_norm": 0.61328125, "lr": 3e-05, "finish_rate": 0.858, "comp_len": 515.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.5, "frames": {"chat": 233}, "mem_gb": 9.99}
34
+ {"step": 34, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.35060337477376063, "tokens": 120000, "cumulative_loss_tokens": 4080000, "grad_norm": 0.64453125, "lr": 3e-05, "finish_rate": 0.786, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.8, "frames": {"chat": 215}, "mem_gb": 10.01}
35
+ {"step": 35, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.33954473176573713, "tokens": 120000, "cumulative_loss_tokens": 4200000, "grad_norm": 0.640625, "lr": 3e-05, "finish_rate": 0.845, "comp_len": 515.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.3, "frames": {"chat": 233}, "mem_gb": 9.99}
36
+ {"step": 36, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.2843651156159739, "tokens": 120000, "cumulative_loss_tokens": 4320000, "grad_norm": 1.828125, "lr": 3e-05, "finish_rate": 0.766, "comp_len": 574.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.9, "frames": {"chat": 209}, "mem_gb": 9.94}
37
+ {"step": 37, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.2947497258403649, "tokens": 120000, "cumulative_loss_tokens": 4440000, "grad_norm": 0.72265625, "lr": 3e-05, "finish_rate": 0.908, "comp_len": 458.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 30.6, "frames": {"chat": 262}, "mem_gb": 9.88}
38
+ {"step": 38, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.28674538816077016, "tokens": 120000, "cumulative_loss_tokens": 4560000, "grad_norm": 0.56640625, "lr": 3e-05, "finish_rate": 0.9, "comp_len": 481.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.9, "frames": {"chat": 249}, "mem_gb": 9.96}
39
+ {"step": 39, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.33914187191314993, "tokens": 120000, "cumulative_loss_tokens": 4680000, "grad_norm": 1.4921875, "lr": 3e-05, "finish_rate": 0.819, "comp_len": 528.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.3, "frames": {"chat": 227}, "mem_gb": 10.0}
40
+ {"step": 40, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.26380810491734497, "tokens": 120000, "cumulative_loss_tokens": 4800000, "grad_norm": 0.5234375, "lr": 3e-05, "finish_rate": 0.814, "comp_len": 543.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.4, "frames": {"chat": 221}, "mem_gb": 10.0}
41
+ {"step": 41, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.302417300863564, "tokens": 120000, "cumulative_loss_tokens": 4920000, "grad_norm": 0.5546875, "lr": 3e-05, "finish_rate": 0.859, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.4, "frames": {"chat": 234}, "mem_gb": 10.01}
42
+ {"step": 42, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.24419383710200587, "tokens": 120000, "cumulative_loss_tokens": 5040000, "grad_norm": 0.52734375, "lr": 3e-05, "finish_rate": 0.817, "comp_len": 563.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.3, "frames": {"chat": 213}, "mem_gb": 9.96}
43
+ {"step": 43, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.24537141441050916, "tokens": 120000, "cumulative_loss_tokens": 5160000, "grad_norm": 0.51953125, "lr": 3e-05, "finish_rate": 0.836, "comp_len": 563.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.5, "frames": {"chat": 213}, "mem_gb": 9.9}
44
+ {"step": 44, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.27504100236129014, "tokens": 120000, "cumulative_loss_tokens": 5280000, "grad_norm": 0.53125, "lr": 3e-05, "finish_rate": 0.906, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.8, "frames": {"chat": 234}, "mem_gb": 9.93}
45
+ {"step": 45, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.2705650013284758, "tokens": 120000, "cumulative_loss_tokens": 5400000, "grad_norm": 0.5, "lr": 3e-05, "finish_rate": 0.793, "comp_len": 540.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.4, "frames": {"chat": 222}, "mem_gb": 10.0}
46
+ {"step": 46, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.31124431811695297, "tokens": 120000, "cumulative_loss_tokens": 5520000, "grad_norm": 0.59375, "lr": 3e-05, "finish_rate": 0.806, "comp_len": 528.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 31.3, "frames": {"chat": 227}, "mem_gb": 10.01}
47
+ {"step": 47, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.2672547916886707, "tokens": 120000, "cumulative_loss_tokens": 5640000, "grad_norm": 0.51171875, "lr": 3e-05, "finish_rate": 0.835, "comp_len": 550.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.2, "frames": {"chat": 218}, "mem_gb": 10.04}
48
+ {"step": 48, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.3164061588189254, "tokens": 120000, "cumulative_loss_tokens": 5760000, "grad_norm": 0.57421875, "lr": 3e-05, "finish_rate": 0.798, "comp_len": 538.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.8, "frames": {"chat": 223}, "mem_gb": 10.01}
49
+ {"step": 49, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.31090367152442533, "tokens": 120000, "cumulative_loss_tokens": 5880000, "grad_norm": 0.56640625, "lr": 3e-05, "finish_rate": 0.772, "comp_len": 582.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.8, "frames": {"chat": 206}, "mem_gb": 10.01}
50
+ {"step": 50, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.2863034582992395, "tokens": 120000, "cumulative_loss_tokens": 6000000, "grad_norm": 0.5625, "lr": 3e-05, "finish_rate": 0.784, "comp_len": 563.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.9, "frames": {"chat": 213}, "mem_gb": 9.92}
51
+ {"step": 51, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.33238482810370623, "tokens": 120000, "cumulative_loss_tokens": 6120000, "grad_norm": 0.5859375, "lr": 3e-05, "finish_rate": 0.843, "comp_len": 538.1, "t_data_s": 0.1, "t_rollout_s": 0.0, "t_step_s": 30.3, "frames": {"chat": 223}, "mem_gb": 9.86}
52
+ {"step": 52, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.30147026825944584, "tokens": 120000, "cumulative_loss_tokens": 6240000, "grad_norm": 0.5703125, "lr": 3e-05, "finish_rate": 0.828, "comp_len": 528.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.3, "frames": {"chat": 227}, "mem_gb": 9.97}
53
+ {"step": 53, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.2694238721400499, "tokens": 120000, "cumulative_loss_tokens": 6360000, "grad_norm": 1.2109375, "lr": 3e-05, "finish_rate": 0.889, "comp_len": 474.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 30.2, "frames": {"chat": 253}, "mem_gb": 10.0}
54
+ {"step": 54, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2599986071868489, "tokens": 120000, "cumulative_loss_tokens": 6480000, "grad_norm": 0.609375, "lr": 3e-05, "finish_rate": 0.792, "comp_len": 555.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.7, "frames": {"chat": 216}, "mem_gb": 10.01}
55
+ {"step": 55, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.26703055294280253, "tokens": 120000, "cumulative_loss_tokens": 6600000, "grad_norm": 0.53125, "lr": 3e-05, "finish_rate": 0.766, "comp_len": 585.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.4, "frames": {"chat": 205}, "mem_gb": 9.98}
56
+ {"step": 56, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2845016695648432, "tokens": 120000, "cumulative_loss_tokens": 6720000, "grad_norm": 0.6328125, "lr": 3e-05, "finish_rate": 0.729, "comp_len": 579.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 30.1, "frames": {"chat": 207}, "mem_gb": 10.06}
57
+ {"step": 57, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2710578287235151, "tokens": 120000, "cumulative_loss_tokens": 6840000, "grad_norm": 0.53515625, "lr": 3e-05, "finish_rate": 0.814, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.5, "frames": {"chat": 215}, "mem_gb": 9.98}
58
+ {"step": 58, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2025449087051054, "tokens": 120000, "cumulative_loss_tokens": 6960000, "grad_norm": 0.4765625, "lr": 3e-05, "finish_rate": 0.86, "comp_len": 526.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.6, "frames": {"chat": 228}, "mem_gb": 10.0}
59
+ {"step": 59, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2637373869329691, "tokens": 120000, "cumulative_loss_tokens": 7080000, "grad_norm": 0.5625, "lr": 3e-05, "finish_rate": 0.747, "comp_len": 543.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.7, "frames": {"chat": 221}, "mem_gb": 10.04}
60
+ {"step": 60, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.20766283342484385, "tokens": 120000, "cumulative_loss_tokens": 7200000, "grad_norm": 0.53515625, "lr": 3e-05, "finish_rate": 0.882, "comp_len": 472.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.7, "frames": {"chat": 254}, "mem_gb": 9.84}
61
+ {"step": 61, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.17684483035219212, "tokens": 120000, "cumulative_loss_tokens": 7320000, "grad_norm": 0.5078125, "lr": 3e-05, "finish_rate": 0.843, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.7, "frames": {"chat": 210}, "mem_gb": 9.97}
62
+ {"step": 62, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.21151079111825674, "tokens": 120000, "cumulative_loss_tokens": 7440000, "grad_norm": 0.48046875, "lr": 3e-05, "finish_rate": 0.827, "comp_len": 531.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.8, "frames": {"chat": 226}, "mem_gb": 9.92}
63
+ {"step": 63, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.22429327983123562, "tokens": 120000, "cumulative_loss_tokens": 7560000, "grad_norm": 0.48046875, "lr": 3e-05, "finish_rate": 0.802, "comp_len": 566.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.1, "frames": {"chat": 212}, "mem_gb": 9.99}
64
+ {"step": 64, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.23901226744391024, "tokens": 120000, "cumulative_loss_tokens": 7680000, "grad_norm": 0.4921875, "lr": 3e-05, "finish_rate": 0.754, "comp_len": 568.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.3, "frames": {"chat": 211}, "mem_gb": 9.93}
65
+ {"step": 65, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2521503586698324, "tokens": 120000, "cumulative_loss_tokens": 7800000, "grad_norm": 0.53125, "lr": 3e-05, "finish_rate": 0.776, "comp_len": 612.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 26.5, "frames": {"chat": 196}, "mem_gb": 9.98}
66
+ {"step": 66, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.19568856958678613, "tokens": 120000, "cumulative_loss_tokens": 7920000, "grad_norm": 0.447265625, "lr": 3e-05, "finish_rate": 0.811, "comp_len": 566.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.8, "frames": {"chat": 212}, "mem_gb": 10.0}
67
+ {"step": 67, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2056147424393023, "tokens": 120000, "cumulative_loss_tokens": 8040000, "grad_norm": 0.447265625, "lr": 3e-05, "finish_rate": 0.877, "comp_len": 491.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.5, "frames": {"chat": 244}, "mem_gb": 9.91}
68
+ {"step": 68, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.19970010116541137, "tokens": 120000, "cumulative_loss_tokens": 8160000, "grad_norm": 0.435546875, "lr": 3e-05, "finish_rate": 0.838, "comp_len": 540.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.3, "frames": {"chat": 222}, "mem_gb": 9.95}
69
+ {"step": 69, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2406689625217269, "tokens": 120000, "cumulative_loss_tokens": 8280000, "grad_norm": 0.5, "lr": 3e-05, "finish_rate": 0.78, "comp_len": 550.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.1, "frames": {"chat": 218}, "mem_gb": 10.0}
70
+ {"step": 70, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2351806751595189, "tokens": 120000, "cumulative_loss_tokens": 8400000, "grad_norm": 0.494140625, "lr": 3e-05, "finish_rate": 0.913, "comp_len": 476.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 30.7, "frames": {"chat": 252}, "mem_gb": 9.88}
71
+ {"step": 71, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.23980661761003236, "tokens": 120000, "cumulative_loss_tokens": 8520000, "grad_norm": 0.486328125, "lr": 3e-05, "finish_rate": 0.728, "comp_len": 594.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.5, "frames": {"chat": 202}, "mem_gb": 10.05}
72
+ {"step": 72, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2756894440931578, "tokens": 120000, "cumulative_loss_tokens": 8640000, "grad_norm": 0.59765625, "lr": 3e-05, "finish_rate": 0.835, "comp_len": 506.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.9, "frames": {"chat": 237}, "mem_gb": 10.0}
73
+ {"step": 73, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.22606769043393432, "tokens": 120000, "cumulative_loss_tokens": 8760000, "grad_norm": 0.486328125, "lr": 3e-05, "finish_rate": 0.868, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.3, "frames": {"chat": 234}, "mem_gb": 9.99}
74
+ {"step": 74, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.17098218618429575, "tokens": 120000, "cumulative_loss_tokens": 8880000, "grad_norm": 0.4375, "lr": 3e-05, "finish_rate": 0.809, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.2, "frames": {"chat": 215}, "mem_gb": 10.0}
75
+ {"step": 75, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.18496631520005563, "tokens": 120000, "cumulative_loss_tokens": 9000000, "grad_norm": 0.4375, "lr": 3e-05, "finish_rate": 0.876, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.7, "frames": {"chat": 234}, "mem_gb": 9.93}
76
+ {"step": 76, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.16087721765795723, "tokens": 120000, "cumulative_loss_tokens": 9120000, "grad_norm": 0.404296875, "lr": 3e-05, "finish_rate": 0.801, "comp_len": 555.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.6, "frames": {"chat": 216}, "mem_gb": 9.99}
77
+ {"step": 77, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.1994732022792101, "tokens": 120000, "cumulative_loss_tokens": 9240000, "grad_norm": 0.4609375, "lr": 3e-05, "finish_rate": 0.805, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.5, "frames": {"chat": 210}, "mem_gb": 9.95}
78
+ {"step": 78, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.20232079331736702, "tokens": 120000, "cumulative_loss_tokens": 9360000, "grad_norm": 0.443359375, "lr": 3e-05, "finish_rate": 0.719, "comp_len": 603.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.2, "frames": {"chat": 199}, "mem_gb": 10.0}
79
+ {"step": 79, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.18569014142416418, "tokens": 120000, "cumulative_loss_tokens": 9480000, "grad_norm": 0.431640625, "lr": 3e-05, "finish_rate": 0.824, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.9, "frames": {"chat": 210}, "mem_gb": 10.01}
80
+ {"step": 80, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.20507156935961296, "tokens": 120000, "cumulative_loss_tokens": 9600000, "grad_norm": 0.5390625, "lr": 3e-05, "finish_rate": 0.902, "comp_len": 533.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.1, "frames": {"chat": 225}, "mem_gb": 9.95}
81
+ {"step": 81, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.20882732380144298, "tokens": 120000, "cumulative_loss_tokens": 9720000, "grad_norm": 0.486328125, "lr": 3e-05, "finish_rate": 0.913, "comp_len": 474.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.9, "frames": {"chat": 253}, "mem_gb": 9.85}
82
+ {"step": 82, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.22202617851061127, "tokens": 120000, "cumulative_loss_tokens": 9840000, "grad_norm": 0.458984375, "lr": 3e-05, "finish_rate": 0.903, "comp_len": 485.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 30.2, "frames": {"chat": 247}, "mem_gb": 9.98}
83
+ {"step": 83, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2211942642432948, "tokens": 120000, "cumulative_loss_tokens": 9960000, "grad_norm": 0.51171875, "lr": 3e-05, "finish_rate": 0.836, "comp_len": 504.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.1, "frames": {"chat": 238}, "mem_gb": 9.98}
84
+ {"step": 84, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.23084418179870894, "tokens": 120000, "cumulative_loss_tokens": 10080000, "grad_norm": 0.51171875, "lr": 3e-05, "finish_rate": 0.86, "comp_len": 510.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.6, "frames": {"chat": 235}, "mem_gb": 10.0}
85
+ {"step": 85, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2283516202347974, "tokens": 120000, "cumulative_loss_tokens": 10200000, "grad_norm": 0.498046875, "lr": 3e-05, "finish_rate": 0.805, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.5, "frames": {"chat": 215}, "mem_gb": 9.97}
86
+ {"step": 86, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.22033791828608762, "tokens": 120000, "cumulative_loss_tokens": 10320000, "grad_norm": 0.53515625, "lr": 3e-05, "finish_rate": 0.925, "comp_len": 447.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 30.0, "frames": {"chat": 268}, "mem_gb": 9.97}
87
+ {"step": 87, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.22125197298427424, "tokens": 120000, "cumulative_loss_tokens": 10440000, "grad_norm": 0.486328125, "lr": 3e-05, "finish_rate": 0.825, "comp_len": 526.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.0, "frames": {"chat": 228}, "mem_gb": 10.0}
88
+ {"step": 88, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.21613482079487295, "tokens": 120000, "cumulative_loss_tokens": 10560000, "grad_norm": 0.49609375, "lr": 3e-05, "finish_rate": 0.881, "comp_len": 476.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.8, "frames": {"chat": 252}, "mem_gb": 9.93}
89
+ {"step": 89, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.1895898001347358, "tokens": 120000, "cumulative_loss_tokens": 10680000, "grad_norm": 0.46484375, "lr": 3e-05, "finish_rate": 0.821, "comp_len": 538.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.6, "frames": {"chat": 223}, "mem_gb": 10.01}
90
+ {"step": 90, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.26797097403767206, "tokens": 120000, "cumulative_loss_tokens": 10800000, "grad_norm": 0.55078125, "lr": 3e-05, "finish_rate": 0.805, "comp_len": 531.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.2, "frames": {"chat": 226}, "mem_gb": 10.0}
91
+ {"step": 91, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.24367389039595921, "tokens": 120000, "cumulative_loss_tokens": 10920000, "grad_norm": 0.55078125, "lr": 3e-05, "finish_rate": 0.731, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 30.4, "frames": {"chat": 208}, "mem_gb": 10.05}
92
+ {"step": 92, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.17548983993784836, "tokens": 120000, "cumulative_loss_tokens": 11040000, "grad_norm": 0.4296875, "lr": 3e-05, "finish_rate": 0.883, "comp_len": 500.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.2, "frames": {"chat": 240}, "mem_gb": 9.93}
93
+ {"step": 93, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.21608840946884206, "tokens": 120000, "cumulative_loss_tokens": 11160000, "grad_norm": 0.47265625, "lr": 3e-05, "finish_rate": 0.842, "comp_len": 540.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.5, "frames": {"chat": 222}, "mem_gb": 9.93}
94
+ {"step": 94, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.1903983515452904, "tokens": 120000, "cumulative_loss_tokens": 11280000, "grad_norm": 0.455078125, "lr": 3e-05, "finish_rate": 0.881, "comp_len": 508.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.5, "frames": {"chat": 236}, "mem_gb": 10.0}
95
+ {"step": 95, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.17124010507706552, "tokens": 120000, "cumulative_loss_tokens": 11400000, "grad_norm": 0.427734375, "lr": 3e-05, "finish_rate": 0.834, "comp_len": 553.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.8, "frames": {"chat": 217}, "mem_gb": 9.97}
96
+ {"step": 96, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.21378996573444456, "tokens": 120000, "cumulative_loss_tokens": 11520000, "grad_norm": 0.59375, "lr": 3e-05, "finish_rate": 0.921, "comp_len": 476.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.6, "frames": {"chat": 252}, "mem_gb": 9.88}
97
+ {"step": 97, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.17583670812180888, "tokens": 120000, "cumulative_loss_tokens": 11640000, "grad_norm": 0.427734375, "lr": 3e-05, "finish_rate": 0.847, "comp_len": 540.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.1, "frames": {"chat": 222}, "mem_gb": 9.99}
98
+ {"step": 98, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2008310172341764, "tokens": 120000, "cumulative_loss_tokens": 11760000, "grad_norm": 0.470703125, "lr": 3e-05, "finish_rate": 0.901, "comp_len": 495.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.1, "frames": {"chat": 242}, "mem_gb": 9.87}
99
+ {"step": 99, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.2504596530464788, "tokens": 120000, "cumulative_loss_tokens": 11880000, "grad_norm": 0.53515625, "lr": 3e-05, "finish_rate": 0.804, "comp_len": 547.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.7, "frames": {"chat": 219}, "mem_gb": 9.93}
100
+ {"step": 100, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.20776792142900327, "tokens": 120000, "cumulative_loss_tokens": 12000000, "grad_norm": 0.44140625, "lr": 3e-05, "finish_rate": 0.879, "comp_len": 538.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.8, "frames": {"chat": 223}, "mem_gb": 9.94}
101
+ {"step": 101, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.21407153125032782, "tokens": 120000, "cumulative_loss_tokens": 12120000, "grad_norm": 0.47265625, "lr": 3e-05, "finish_rate": 0.858, "comp_len": 517.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.3, "frames": {"chat": 232}, "mem_gb": 9.95}
102
+ {"step": 102, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.22760161069128662, "tokens": 120000, "cumulative_loss_tokens": 12240000, "grad_norm": 0.53125, "lr": 3e-05, "finish_rate": 0.832, "comp_len": 545.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.2, "frames": {"chat": 220}, "mem_gb": 10.0}
103
+ {"step": 103, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.23536558933630586, "tokens": 120000, "cumulative_loss_tokens": 12360000, "grad_norm": 0.51171875, "lr": 3e-05, "finish_rate": 0.776, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.6, "frames": {"chat": 210}, "mem_gb": 10.04}
104
+ {"step": 104, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.20387262995596975, "tokens": 120000, "cumulative_loss_tokens": 12480000, "grad_norm": 0.482421875, "lr": 3e-05, "finish_rate": 0.81, "comp_len": 531.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.6, "frames": {"chat": 226}, "mem_gb": 9.97}
105
+ {"step": 105, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.1928112811392794, "tokens": 120000, "cumulative_loss_tokens": 12600000, "grad_norm": 0.466796875, "lr": 3e-05, "finish_rate": 0.741, "comp_len": 566.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.8, "frames": {"chat": 212}, "mem_gb": 10.0}
106
+ {"step": 106, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.18794621144427606, "tokens": 120000, "cumulative_loss_tokens": 12720000, "grad_norm": 0.48828125, "lr": 3e-05, "finish_rate": 0.839, "comp_len": 508.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.0, "frames": {"chat": 236}, "mem_gb": 10.01}
107
+ {"step": 107, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.16627150794851284, "tokens": 120000, "cumulative_loss_tokens": 12840000, "grad_norm": 0.56640625, "lr": 3e-05, "finish_rate": 0.928, "comp_len": 454.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.7, "frames": {"chat": 264}, "mem_gb": 9.88}
108
+ {"step": 108, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.21027346796846638, "tokens": 120000, "cumulative_loss_tokens": 12960000, "grad_norm": 0.48046875, "lr": 3e-05, "finish_rate": 0.834, "comp_len": 524.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.3, "frames": {"chat": 229}, "mem_gb": 9.98}
109
+ {"step": 109, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.16645300974740337, "tokens": 120000, "cumulative_loss_tokens": 13080000, "grad_norm": 0.5078125, "lr": 3e-05, "finish_rate": 0.903, "comp_len": 465.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.9, "frames": {"chat": 258}, "mem_gb": 9.85}
110
+ {"step": 110, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.21918750831211606, "tokens": 120000, "cumulative_loss_tokens": 13200000, "grad_norm": 0.55859375, "lr": 3e-05, "finish_rate": 0.755, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.4, "frames": {"chat": 208}, "mem_gb": 10.01}
111
+ {"step": 111, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.18296561656640842, "tokens": 120000, "cumulative_loss_tokens": 13320000, "grad_norm": 0.48046875, "lr": 3e-05, "finish_rate": 0.88, "comp_len": 481.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.4, "frames": {"chat": 249}, "mem_gb": 9.93}
112
+ {"step": 112, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.14481682858659575, "tokens": 120000, "cumulative_loss_tokens": 13440000, "grad_norm": 0.41015625, "lr": 3e-05, "finish_rate": 0.845, "comp_len": 545.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.5, "frames": {"chat": 220}, "mem_gb": 9.99}
113
+ {"step": 113, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1859534092683966, "tokens": 120000, "cumulative_loss_tokens": 13560000, "grad_norm": 0.56640625, "lr": 3e-05, "finish_rate": 0.834, "comp_len": 538.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.8, "frames": {"chat": 223}, "mem_gb": 9.99}
114
+ {"step": 114, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.13867307858555578, "tokens": 120000, "cumulative_loss_tokens": 13680000, "grad_norm": 0.458984375, "lr": 3e-05, "finish_rate": 0.833, "comp_len": 543.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.1, "frames": {"chat": 221}, "mem_gb": 10.0}
115
+ {"step": 115, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.15031419390533118, "tokens": 120000, "cumulative_loss_tokens": 13800000, "grad_norm": 0.427734375, "lr": 3e-05, "finish_rate": 0.9, "comp_len": 521.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.5, "frames": {"chat": 230}, "mem_gb": 9.9}
116
+ {"step": 116, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.17688426749395827, "tokens": 120000, "cumulative_loss_tokens": 13920000, "grad_norm": 0.458984375, "lr": 3e-05, "finish_rate": 0.776, "comp_len": 560.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.4, "frames": {"chat": 214}, "mem_gb": 9.97}
117
+ {"step": 117, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.2198834235218664, "tokens": 120000, "cumulative_loss_tokens": 14040000, "grad_norm": 0.51171875, "lr": 3e-05, "finish_rate": 0.766, "comp_len": 560.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.7, "frames": {"chat": 214}, "mem_gb": 9.99}
118
+ {"step": 118, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1655972873053203, "tokens": 120000, "cumulative_loss_tokens": 14160000, "grad_norm": 0.478515625, "lr": 3e-05, "finish_rate": 0.786, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.5, "frames": {"chat": 210}, "mem_gb": 10.04}
119
+ {"step": 119, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.18060686992689345, "tokens": 120000, "cumulative_loss_tokens": 14280000, "grad_norm": 0.45703125, "lr": 3e-05, "finish_rate": 0.776, "comp_len": 560.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.2, "frames": {"chat": 214}, "mem_gb": 10.0}
120
+ {"step": 120, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.18515299578898897, "tokens": 120000, "cumulative_loss_tokens": 14400000, "grad_norm": 0.451171875, "lr": 3e-05, "finish_rate": 0.791, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.9, "frames": {"chat": 215}, "mem_gb": 9.96}
121
+ {"step": 121, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.19961161344274878, "tokens": 120000, "cumulative_loss_tokens": 14520000, "grad_norm": 0.4921875, "lr": 3e-05, "finish_rate": 0.721, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.4, "frames": {"chat": 208}, "mem_gb": 9.99}
122
+ {"step": 122, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.150687341630583, "tokens": 120000, "cumulative_loss_tokens": 14640000, "grad_norm": 0.400390625, "lr": 3e-05, "finish_rate": 0.789, "comp_len": 550.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.4, "frames": {"chat": 218}, "mem_gb": 9.88}
123
+ {"step": 123, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1382546880559375, "tokens": 120000, "cumulative_loss_tokens": 14760000, "grad_norm": 0.388671875, "lr": 3e-05, "finish_rate": 0.876, "comp_len": 515.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.7, "frames": {"chat": 233}, "mem_gb": 9.9}
124
+ {"step": 124, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.13145577659346164, "tokens": 120000, "cumulative_loss_tokens": 14880000, "grad_norm": 0.388671875, "lr": 3e-05, "finish_rate": 0.861, "comp_len": 519.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.8, "frames": {"chat": 231}, "mem_gb": 9.94}
125
+ {"step": 125, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.16622866188141827, "tokens": 120000, "cumulative_loss_tokens": 15000000, "grad_norm": 0.435546875, "lr": 3e-05, "finish_rate": 0.868, "comp_len": 510.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.8, "frames": {"chat": 235}, "mem_gb": 10.13}
126
+ {"step": 126, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.17218660681688538, "tokens": 120000, "cumulative_loss_tokens": 15120000, "grad_norm": 0.419921875, "lr": 3e-05, "finish_rate": 0.843, "comp_len": 555.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.1, "frames": {"chat": 216}, "mem_gb": 9.99}
127
+ {"step": 127, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.16202461753959457, "tokens": 120000, "cumulative_loss_tokens": 15240000, "grad_norm": 0.427734375, "lr": 3e-05, "finish_rate": 0.831, "comp_len": 506.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.1, "frames": {"chat": 237}, "mem_gb": 10.01}
128
+ {"step": 128, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1946208799743404, "tokens": 120000, "cumulative_loss_tokens": 15360000, "grad_norm": 0.46484375, "lr": 3e-05, "finish_rate": 0.734, "comp_len": 591.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.9, "frames": {"chat": 203}, "mem_gb": 10.01}
129
+ {"step": 129, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.16476607479564845, "tokens": 120000, "cumulative_loss_tokens": 15480000, "grad_norm": 0.427734375, "lr": 3e-05, "finish_rate": 0.873, "comp_len": 508.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.5, "frames": {"chat": 236}, "mem_gb": 10.04}
130
+ {"step": 130, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.16959726118591303, "tokens": 120000, "cumulative_loss_tokens": 15600000, "grad_norm": 0.42578125, "lr": 3e-05, "finish_rate": 0.734, "comp_len": 560.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.0, "frames": {"chat": 214}, "mem_gb": 10.01}
131
+ {"step": 131, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.15791527282614262, "tokens": 120000, "cumulative_loss_tokens": 15720000, "grad_norm": 0.421875, "lr": 3e-05, "finish_rate": 0.78, "comp_len": 574.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.7, "frames": {"chat": 209}, "mem_gb": 10.0}
132
+ {"step": 132, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.16019434139728547, "tokens": 120000, "cumulative_loss_tokens": 15840000, "grad_norm": 0.443359375, "lr": 3e-05, "finish_rate": 0.906, "comp_len": 468.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 30.5, "frames": {"chat": 256}, "mem_gb": 10.0}
133
+ {"step": 133, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1639191758962348, "tokens": 120000, "cumulative_loss_tokens": 15960000, "grad_norm": 0.455078125, "lr": 3e-05, "finish_rate": 0.878, "comp_len": 521.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.0, "frames": {"chat": 230}, "mem_gb": 9.87}
134
+ {"step": 134, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.16772623434948425, "tokens": 120000, "cumulative_loss_tokens": 16080000, "grad_norm": 0.439453125, "lr": 3e-05, "finish_rate": 0.822, "comp_len": 521.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.3, "frames": {"chat": 230}, "mem_gb": 10.06}
135
+ {"step": 135, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1819166350507488, "tokens": 120000, "cumulative_loss_tokens": 16200000, "grad_norm": 0.44921875, "lr": 3e-05, "finish_rate": 0.881, "comp_len": 528.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.6, "frames": {"chat": 227}, "mem_gb": 9.95}
136
+ {"step": 136, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.1837459068759655, "tokens": 120000, "cumulative_loss_tokens": 16320000, "grad_norm": 0.455078125, "lr": 3e-05, "finish_rate": 0.755, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.7, "frames": {"chat": 208}, "mem_gb": 10.01}
137
+ {"step": 137, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.19119155149323244, "tokens": 120000, "cumulative_loss_tokens": 16440000, "grad_norm": 0.474609375, "lr": 3e-05, "finish_rate": 0.699, "comp_len": 582.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.9, "frames": {"chat": 206}, "mem_gb": 10.03}
138
+ {"step": 138, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.17735388877652586, "tokens": 120000, "cumulative_loss_tokens": 16560000, "grad_norm": 0.48046875, "lr": 3e-05, "finish_rate": 0.82, "comp_len": 526.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.5, "frames": {"chat": 228}, "mem_gb": 9.9}
139
+ {"step": 139, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.17459255363605916, "tokens": 120000, "cumulative_loss_tokens": 16680000, "grad_norm": 0.462890625, "lr": 3e-05, "finish_rate": 0.835, "comp_len": 535.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.5, "frames": {"chat": 224}, "mem_gb": 10.0}
140
+ {"step": 140, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.15427039775537948, "tokens": 120000, "cumulative_loss_tokens": 16800000, "grad_norm": 0.408203125, "lr": 3e-05, "finish_rate": 0.66, "comp_len": 600.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.1, "frames": {"chat": 200}, "mem_gb": 10.03}
141
+ {"step": 141, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.17032653220028926, "tokens": 120000, "cumulative_loss_tokens": 16920000, "grad_norm": 0.45703125, "lr": 3e-05, "finish_rate": 0.714, "comp_len": 612.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 26.8, "frames": {"chat": 196}, "mem_gb": 10.01}
142
+ {"step": 142, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.14531434423498188, "tokens": 120000, "cumulative_loss_tokens": 17040000, "grad_norm": 0.400390625, "lr": 3e-05, "finish_rate": 0.834, "comp_len": 538.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.3, "frames": {"chat": 223}, "mem_gb": 10.0}
143
+ {"step": 143, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.13336851223480578, "tokens": 120000, "cumulative_loss_tokens": 17160000, "grad_norm": 0.376953125, "lr": 3e-05, "finish_rate": 0.869, "comp_len": 563.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.4, "frames": {"chat": 213}, "mem_gb": 9.89}
144
+ {"step": 144, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.14154418901971852, "tokens": 120000, "cumulative_loss_tokens": 17280000, "grad_norm": 0.427734375, "lr": 3e-05, "finish_rate": 0.879, "comp_len": 517.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.9, "frames": {"chat": 232}, "mem_gb": 9.93}
145
+ {"step": 145, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.15019642852085333, "tokens": 120000, "cumulative_loss_tokens": 17400000, "grad_norm": 0.447265625, "lr": 3e-05, "finish_rate": 0.861, "comp_len": 538.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.3, "frames": {"chat": 223}, "mem_gb": 9.93}
146
+ {"step": 146, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.15932164447531105, "tokens": 120000, "cumulative_loss_tokens": 17520000, "grad_norm": 0.421875, "lr": 3e-05, "finish_rate": 0.85, "comp_len": 515.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.7, "frames": {"chat": 233}, "mem_gb": 10.02}
147
+ {"step": 147, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.16258757215073952, "tokens": 120000, "cumulative_loss_tokens": 17640000, "grad_norm": 0.431640625, "lr": 3e-05, "finish_rate": 0.816, "comp_len": 553.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 27.9, "frames": {"chat": 217}, "mem_gb": 10.01}
148
+ {"step": 148, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.21068223652498175, "tokens": 120000, "cumulative_loss_tokens": 17760000, "grad_norm": 0.515625, "lr": 3e-05, "finish_rate": 0.752, "comp_len": 594.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.3, "frames": {"chat": 202}, "mem_gb": 10.08}
149
+ {"step": 149, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.14553914751367023, "tokens": 120000, "cumulative_loss_tokens": 17880000, "grad_norm": 0.4453125, "lr": 3e-05, "finish_rate": 0.858, "comp_len": 474.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 29.9, "frames": {"chat": 253}, "mem_gb": 9.93}
150
+ {"step": 150, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.13818583363940318, "tokens": 120000, "cumulative_loss_tokens": 18000000, "grad_norm": 0.396484375, "lr": 3e-05, "finish_rate": 0.879, "comp_len": 519.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 28.6, "frames": {"chat": 231}, "mem_gb": 9.94}
healed/grid_math/uniform_keep75_s1226/args.json ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "student": "outputs/pruned/uniform-0125inst-math-keep75",
3
+ "teacher": "allenai/OLMoE-1B-7B-0125-Instruct",
4
+ "training_mode": "off-policy",
5
+ "kl_direction": "forward",
6
+ "dataset": "allenai/RLVR-MATH",
7
+ "dataset_sources": null,
8
+ "max_difficulty": null,
9
+ "trajectories": "outputs/teacher_trajectories/dolci_math_curated.jsonl",
10
+ "trajectory_dataset": "allenai/Dolci-Instruct-RL",
11
+ "off_policy_frames": "chat",
12
+ "off_policy_max_seq_len": 2048,
13
+ "topk_targets": "outputs/teacher_trajectories/dolci_math_curated_opd_top128",
14
+ "max_loss_tokens": null,
15
+ "loss_tokens_per_step": 120000,
16
+ "teacher_device": "cuda:0",
17
+ "student_device": "cuda:0",
18
+ "lr": 3e-05,
19
+ "optimizer": "adamw8bit",
20
+ "weight_decay": 0.1,
21
+ "epochs": 3,
22
+ "prompts_per_step": 256,
23
+ "group_size": 1,
24
+ "rollout_batch": 64,
25
+ "micro_batch": 3,
26
+ "max_new_tokens": 256,
27
+ "max_prompt_len": 1024,
28
+ "warmup_steps": 10,
29
+ "max_grad_norm": 1.0,
30
+ "eval_every": 10,
31
+ "gsm8k_every": 0,
32
+ "gsm8k_n": 256,
33
+ "gsm8k_batch": 16,
34
+ "gsm8k_max_new_tokens": 512,
35
+ "gsm8k_frames": "chat",
36
+ "save_every": 50,
37
+ "out_dir": "outputs/healed/grid_math/uniform_keep75_s1226",
38
+ "sweep": 150,
39
+ "wandb": true,
40
+ "wandb_project": "glean-grid",
41
+ "wandb_run_name": "uniform-math-keep75-s1226",
42
+ "wandb_run_id": null,
43
+ "wandb_resume": null,
44
+ "wandb_mode": "online",
45
+ "no_wandb_sync": false,
46
+ "debug": false,
47
+ "resume_from": null,
48
+ "start_step": 0,
49
+ "no_grad_checkpointing": false,
50
+ "seed": 1226,
51
+ "no_teacher_overlap": false,
52
+ "sync_checkpoints": false,
53
+ "rollout_engine": "hf",
54
+ "vllm_gpu": null,
55
+ "vllm_port": 8377,
56
+ "vllm_refresh_every": 5,
57
+ "vllm_serve_bin": "vllm-plugin/.venv/bin/python",
58
+ "vllm_gpu_mem_util": 0.85,
59
+ "vllm_refresh_mode": "reload",
60
+ "vllm_live_dir": null,
61
+ "resolved_kl_direction": "forward"
62
+ }
healed/grid_math/uniform_keep75_s1226/train_log.jsonl ADDED
@@ -0,0 +1,150 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"step": 1, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.11920493249129505, "tokens": 120000, "cumulative_loss_tokens": 120000, "grad_norm": 1.4453125, "lr": 6e-06, "finish_rate": 0.902, "comp_len": 472.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 54.8, "frames": {"chat": 254}, "mem_gb": 21.82}
2
+ {"step": 2, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.13180598530409238, "tokens": 120000, "cumulative_loss_tokens": 240000, "grad_norm": 1.640625, "lr": 9e-06, "finish_rate": 0.876, "comp_len": 497.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.6, "frames": {"chat": 241}, "mem_gb": 22.07}
3
+ {"step": 3, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.14182352375667542, "tokens": 120000, "cumulative_loss_tokens": 360000, "grad_norm": 1.40625, "lr": 1.2e-05, "finish_rate": 0.746, "comp_len": 563.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.6, "frames": {"chat": 213}, "mem_gb": 22.1}
4
+ {"step": 4, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.11608811415806412, "tokens": 120000, "cumulative_loss_tokens": 480000, "grad_norm": 0.94921875, "lr": 1.5e-05, "finish_rate": 0.864, "comp_len": 543.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.3, "frames": {"chat": 221}, "mem_gb": 22.15}
5
+ {"step": 5, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.12302177668878188, "tokens": 120000, "cumulative_loss_tokens": 600000, "grad_norm": 0.8515625, "lr": 1.8e-05, "finish_rate": 0.745, "comp_len": 612.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.1, "frames": {"chat": 196}, "mem_gb": 22.11}
6
+ {"step": 6, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.0868311198878102, "tokens": 120000, "cumulative_loss_tokens": 720000, "grad_norm": 0.75, "lr": 2.1e-05, "finish_rate": 0.926, "comp_len": 444.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 49.6, "frames": {"chat": 270}, "mem_gb": 21.91}
7
+ {"step": 7, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.09967558807097375, "tokens": 120000, "cumulative_loss_tokens": 840000, "grad_norm": 0.859375, "lr": 2.4e-05, "finish_rate": 0.815, "comp_len": 555.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.7, "frames": {"chat": 216}, "mem_gb": 22.09}
8
+ {"step": 8, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.09974337469547366, "tokens": 120000, "cumulative_loss_tokens": 960000, "grad_norm": 0.73828125, "lr": 2.7000000000000002e-05, "finish_rate": 0.775, "comp_len": 600.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.7, "frames": {"chat": 200}, "mem_gb": 22.06}
9
+ {"step": 9, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.08927157517972713, "tokens": 120000, "cumulative_loss_tokens": 1080000, "grad_norm": 0.64453125, "lr": 3e-05, "finish_rate": 0.767, "comp_len": 582.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.9, "frames": {"chat": 206}, "mem_gb": 22.01}
10
+ {"step": 10, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.0773138673888209, "tokens": 120000, "cumulative_loss_tokens": 1200000, "grad_norm": 0.54296875, "lr": 3e-05, "finish_rate": 0.902, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.4, "frames": {"chat": 234}, "mem_gb": 22.04}
11
+ {"step": 11, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.08695185078081365, "tokens": 120000, "cumulative_loss_tokens": 1320000, "grad_norm": 0.51953125, "lr": 3e-05, "finish_rate": 0.823, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.6, "frames": {"chat": 215}, "mem_gb": 22.05}
12
+ {"step": 12, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.07132623356043671, "tokens": 120000, "cumulative_loss_tokens": 1440000, "grad_norm": 0.455078125, "lr": 3e-05, "finish_rate": 0.922, "comp_len": 470.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.7, "frames": {"chat": 255}, "mem_gb": 22.04}
13
+ {"step": 13, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.0732923347460106, "tokens": 120000, "cumulative_loss_tokens": 1560000, "grad_norm": 0.453125, "lr": 3e-05, "finish_rate": 0.892, "comp_len": 480.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.8, "frames": {"chat": 250}, "mem_gb": 21.92}
14
+ {"step": 14, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.07680565097226451, "tokens": 120000, "cumulative_loss_tokens": 1680000, "grad_norm": 0.4296875, "lr": 3e-05, "finish_rate": 0.884, "comp_len": 495.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.1, "frames": {"chat": 242}, "mem_gb": 22.09}
15
+ {"step": 15, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.09355738684851676, "tokens": 120000, "cumulative_loss_tokens": 1800000, "grad_norm": 0.484375, "lr": 3e-05, "finish_rate": 0.729, "comp_len": 603.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.6, "frames": {"chat": 199}, "mem_gb": 22.1}
16
+ {"step": 16, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.10224050603043287, "tokens": 120000, "cumulative_loss_tokens": 1920000, "grad_norm": 0.57421875, "lr": 3e-05, "finish_rate": 0.784, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.7, "frames": {"chat": 208}, "mem_gb": 22.13}
17
+ {"step": 17, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.0901082269590348, "tokens": 120000, "cumulative_loss_tokens": 2040000, "grad_norm": 0.6328125, "lr": 3e-05, "finish_rate": 0.764, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.5, "frames": {"chat": 208}, "mem_gb": 22.07}
18
+ {"step": 18, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.09369800321385265, "tokens": 120000, "cumulative_loss_tokens": 2160000, "grad_norm": 0.46484375, "lr": 3e-05, "finish_rate": 0.732, "comp_len": 574.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.7, "frames": {"chat": 209}, "mem_gb": 22.22}
19
+ {"step": 19, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.06841314388358345, "tokens": 120000, "cumulative_loss_tokens": 2280000, "grad_norm": 0.380859375, "lr": 3e-05, "finish_rate": 0.855, "comp_len": 510.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.2, "frames": {"chat": 235}, "mem_gb": 22.05}
20
+ {"step": 20, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.07472310523393874, "tokens": 120000, "cumulative_loss_tokens": 2400000, "grad_norm": 0.396484375, "lr": 3e-05, "finish_rate": 0.74, "comp_len": 588.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.7, "frames": {"chat": 204}, "mem_gb": 22.04}
21
+ {"step": 21, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.09344625266479949, "tokens": 120000, "cumulative_loss_tokens": 2520000, "grad_norm": 0.55859375, "lr": 3e-05, "finish_rate": 0.745, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.6, "frames": {"chat": 208}, "mem_gb": 22.1}
22
+ {"step": 22, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.07106327358403554, "tokens": 120000, "cumulative_loss_tokens": 2640000, "grad_norm": 0.392578125, "lr": 3e-05, "finish_rate": 0.825, "comp_len": 500.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 47.1, "frames": {"chat": 240}, "mem_gb": 22.09}
23
+ {"step": 23, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.06766866139359773, "tokens": 120000, "cumulative_loss_tokens": 2760000, "grad_norm": 0.40234375, "lr": 3e-05, "finish_rate": 0.89, "comp_len": 487.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.0, "frames": {"chat": 246}, "mem_gb": 22.09}
24
+ {"step": 24, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.06455388144083941, "tokens": 120000, "cumulative_loss_tokens": 2880000, "grad_norm": 0.384765625, "lr": 3e-05, "finish_rate": 0.909, "comp_len": 493.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.5, "frames": {"chat": 243}, "mem_gb": 21.91}
25
+ {"step": 25, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.07318340876198684, "tokens": 120000, "cumulative_loss_tokens": 3000000, "grad_norm": 0.38671875, "lr": 3e-05, "finish_rate": 0.745, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.9, "frames": {"chat": 208}, "mem_gb": 22.1}
26
+ {"step": 26, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.07118238478129109, "tokens": 120000, "cumulative_loss_tokens": 3120000, "grad_norm": 0.36328125, "lr": 3e-05, "finish_rate": 0.817, "comp_len": 547.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.9, "frames": {"chat": 219}, "mem_gb": 22.1}
27
+ {"step": 27, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.0817199091798005, "tokens": 120000, "cumulative_loss_tokens": 3240000, "grad_norm": 0.431640625, "lr": 3e-05, "finish_rate": 0.782, "comp_len": 568.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.1, "frames": {"chat": 211}, "mem_gb": 22.11}
28
+ {"step": 28, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.06760180381714988, "tokens": 120000, "cumulative_loss_tokens": 3360000, "grad_norm": 0.3828125, "lr": 3e-05, "finish_rate": 0.862, "comp_len": 517.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.3, "frames": {"chat": 232}, "mem_gb": 22.07}
29
+ {"step": 29, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.0715414209818157, "tokens": 120000, "cumulative_loss_tokens": 3480000, "grad_norm": 0.365234375, "lr": 3e-05, "finish_rate": 0.804, "comp_len": 560.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.8, "frames": {"chat": 214}, "mem_gb": 22.1}
30
+ {"step": 30, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.06632242755762612, "tokens": 120000, "cumulative_loss_tokens": 3600000, "grad_norm": 0.3359375, "lr": 3e-05, "finish_rate": 0.819, "comp_len": 531.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.8, "frames": {"chat": 226}, "mem_gb": 21.99}
31
+ {"step": 31, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.06462202967116609, "tokens": 120000, "cumulative_loss_tokens": 3720000, "grad_norm": 0.33984375, "lr": 3e-05, "finish_rate": 0.819, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.8, "frames": {"chat": 210}, "mem_gb": 22.11}
32
+ {"step": 32, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.06298202505940571, "tokens": 120000, "cumulative_loss_tokens": 3840000, "grad_norm": 0.404296875, "lr": 3e-05, "finish_rate": 0.858, "comp_len": 550.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.0, "frames": {"chat": 218}, "mem_gb": 21.93}
33
+ {"step": 33, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.05829383838605136, "tokens": 120000, "cumulative_loss_tokens": 3960000, "grad_norm": 0.330078125, "lr": 3e-05, "finish_rate": 0.858, "comp_len": 515.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.0, "frames": {"chat": 233}, "mem_gb": 22.08}
34
+ {"step": 34, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.07446961007329325, "tokens": 120000, "cumulative_loss_tokens": 4080000, "grad_norm": 0.388671875, "lr": 3e-05, "finish_rate": 0.786, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.2, "frames": {"chat": 215}, "mem_gb": 22.1}
35
+ {"step": 35, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.06408025780933288, "tokens": 120000, "cumulative_loss_tokens": 4200000, "grad_norm": 0.341796875, "lr": 3e-05, "finish_rate": 0.845, "comp_len": 515.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.3, "frames": {"chat": 233}, "mem_gb": 22.08}
36
+ {"step": 36, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.06080306692530091, "tokens": 120000, "cumulative_loss_tokens": 4320000, "grad_norm": 0.353515625, "lr": 3e-05, "finish_rate": 0.766, "comp_len": 574.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.4, "frames": {"chat": 209}, "mem_gb": 22.04}
37
+ {"step": 37, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.05366904664818818, "tokens": 120000, "cumulative_loss_tokens": 4440000, "grad_norm": 0.310546875, "lr": 3e-05, "finish_rate": 0.908, "comp_len": 458.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 47.9, "frames": {"chat": 262}, "mem_gb": 21.97}
38
+ {"step": 38, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.05671166869318113, "tokens": 120000, "cumulative_loss_tokens": 4560000, "grad_norm": 0.326171875, "lr": 3e-05, "finish_rate": 0.9, "comp_len": 481.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.7, "frames": {"chat": 249}, "mem_gb": 22.06}
39
+ {"step": 39, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.07106226710767174, "tokens": 120000, "cumulative_loss_tokens": 4680000, "grad_norm": 0.3671875, "lr": 3e-05, "finish_rate": 0.819, "comp_len": 528.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.6, "frames": {"chat": 227}, "mem_gb": 22.09}
40
+ {"step": 40, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.05462301931890349, "tokens": 120000, "cumulative_loss_tokens": 4800000, "grad_norm": 0.33984375, "lr": 3e-05, "finish_rate": 0.814, "comp_len": 543.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.3, "frames": {"chat": 221}, "mem_gb": 22.09}
41
+ {"step": 41, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.0582238324320099, "tokens": 120000, "cumulative_loss_tokens": 4920000, "grad_norm": 0.333984375, "lr": 3e-05, "finish_rate": 0.859, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.4, "frames": {"chat": 234}, "mem_gb": 22.1}
42
+ {"step": 42, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.05399296633934913, "tokens": 120000, "cumulative_loss_tokens": 5040000, "grad_norm": 0.330078125, "lr": 3e-05, "finish_rate": 0.817, "comp_len": 563.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.6, "frames": {"chat": 213}, "mem_gb": 22.05}
43
+ {"step": 43, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.05386831222480784, "tokens": 120000, "cumulative_loss_tokens": 5160000, "grad_norm": 0.298828125, "lr": 3e-05, "finish_rate": 0.836, "comp_len": 563.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.8, "frames": {"chat": 213}, "mem_gb": 21.99}
44
+ {"step": 44, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.05173095208670323, "tokens": 120000, "cumulative_loss_tokens": 5280000, "grad_norm": 0.314453125, "lr": 3e-05, "finish_rate": 0.906, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.9, "frames": {"chat": 234}, "mem_gb": 22.02}
45
+ {"step": 45, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.05363421202797132, "tokens": 120000, "cumulative_loss_tokens": 5400000, "grad_norm": 0.322265625, "lr": 3e-05, "finish_rate": 0.793, "comp_len": 540.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.3, "frames": {"chat": 222}, "mem_gb": 22.09}
46
+ {"step": 46, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.0622261526023969, "tokens": 120000, "cumulative_loss_tokens": 5520000, "grad_norm": 0.361328125, "lr": 3e-05, "finish_rate": 0.806, "comp_len": 528.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 48.7, "frames": {"chat": 227}, "mem_gb": 22.1}
47
+ {"step": 47, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.05994350822979274, "tokens": 120000, "cumulative_loss_tokens": 5640000, "grad_norm": 0.32421875, "lr": 3e-05, "finish_rate": 0.835, "comp_len": 550.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.1, "frames": {"chat": 218}, "mem_gb": 22.14}
48
+ {"step": 48, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.06174757863963023, "tokens": 120000, "cumulative_loss_tokens": 5760000, "grad_norm": 0.341796875, "lr": 3e-05, "finish_rate": 0.798, "comp_len": 538.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.9, "frames": {"chat": 223}, "mem_gb": 22.1}
49
+ {"step": 49, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.06565356359497333, "tokens": 120000, "cumulative_loss_tokens": 5880000, "grad_norm": 0.359375, "lr": 3e-05, "finish_rate": 0.772, "comp_len": 582.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.4, "frames": {"chat": 206}, "mem_gb": 22.1}
50
+ {"step": 50, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.05670059742469651, "tokens": 120000, "cumulative_loss_tokens": 6000000, "grad_norm": 0.33984375, "lr": 3e-05, "finish_rate": 0.784, "comp_len": 563.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.7, "frames": {"chat": 213}, "mem_gb": 22.02}
51
+ {"step": 51, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.07032583207286273, "tokens": 120000, "cumulative_loss_tokens": 6120000, "grad_norm": 0.357421875, "lr": 3e-05, "finish_rate": 0.843, "comp_len": 538.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.2, "frames": {"chat": 223}, "mem_gb": 21.96}
52
+ {"step": 52, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.059393302978233746, "tokens": 120000, "cumulative_loss_tokens": 6240000, "grad_norm": 0.333984375, "lr": 3e-05, "finish_rate": 0.828, "comp_len": 528.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.3, "frames": {"chat": 227}, "mem_gb": 22.07}
53
+ {"step": 53, "epoch": 0, "training_mode": "off-policy", "forward_topk_kl": 0.05184000949463807, "tokens": 120000, "cumulative_loss_tokens": 6360000, "grad_norm": 0.32421875, "lr": 3e-05, "finish_rate": 0.889, "comp_len": 474.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 47.2, "frames": {"chat": 253}, "mem_gb": 22.09}
54
+ {"step": 54, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.038828376208990815, "tokens": 120000, "cumulative_loss_tokens": 6480000, "grad_norm": 0.259765625, "lr": 3e-05, "finish_rate": 0.792, "comp_len": 555.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.3, "frames": {"chat": 216}, "mem_gb": 22.1}
55
+ {"step": 55, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.039238105912537624, "tokens": 120000, "cumulative_loss_tokens": 6600000, "grad_norm": 0.255859375, "lr": 3e-05, "finish_rate": 0.766, "comp_len": 585.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.0, "frames": {"chat": 205}, "mem_gb": 22.07}
56
+ {"step": 56, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.047572107557269434, "tokens": 120000, "cumulative_loss_tokens": 6720000, "grad_norm": 0.2890625, "lr": 3e-05, "finish_rate": 0.729, "comp_len": 579.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.7, "frames": {"chat": 207}, "mem_gb": 22.16}
57
+ {"step": 57, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.04661659479457885, "tokens": 120000, "cumulative_loss_tokens": 6840000, "grad_norm": 0.28125, "lr": 3e-05, "finish_rate": 0.814, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.9, "frames": {"chat": 215}, "mem_gb": 22.08}
58
+ {"step": 58, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.034201694558095186, "tokens": 120000, "cumulative_loss_tokens": 6960000, "grad_norm": 0.27734375, "lr": 3e-05, "finish_rate": 0.86, "comp_len": 526.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.8, "frames": {"chat": 228}, "mem_gb": 22.1}
59
+ {"step": 59, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.04092416759391005, "tokens": 120000, "cumulative_loss_tokens": 7080000, "grad_norm": 0.267578125, "lr": 3e-05, "finish_rate": 0.747, "comp_len": 543.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.4, "frames": {"chat": 221}, "mem_gb": 22.14}
60
+ {"step": 60, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.029803095571851977, "tokens": 120000, "cumulative_loss_tokens": 7200000, "grad_norm": 0.2138671875, "lr": 3e-05, "finish_rate": 0.882, "comp_len": 472.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.7, "frames": {"chat": 254}, "mem_gb": 21.93}
61
+ {"step": 61, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03530521812000467, "tokens": 120000, "cumulative_loss_tokens": 7320000, "grad_norm": 0.28515625, "lr": 3e-05, "finish_rate": 0.843, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.3, "frames": {"chat": 210}, "mem_gb": 22.06}
62
+ {"step": 62, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.033094389010050025, "tokens": 120000, "cumulative_loss_tokens": 7440000, "grad_norm": 0.21484375, "lr": 3e-05, "finish_rate": 0.827, "comp_len": 531.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.1, "frames": {"chat": 226}, "mem_gb": 22.02}
63
+ {"step": 63, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03637227460097832, "tokens": 120000, "cumulative_loss_tokens": 7560000, "grad_norm": 0.26953125, "lr": 3e-05, "finish_rate": 0.802, "comp_len": 566.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.8, "frames": {"chat": 212}, "mem_gb": 22.09}
64
+ {"step": 64, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.041987014663917945, "tokens": 120000, "cumulative_loss_tokens": 7680000, "grad_norm": 0.265625, "lr": 3e-05, "finish_rate": 0.754, "comp_len": 568.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.6, "frames": {"chat": 211}, "mem_gb": 22.02}
65
+ {"step": 65, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03841204237483131, "tokens": 120000, "cumulative_loss_tokens": 7800000, "grad_norm": 0.26953125, "lr": 3e-05, "finish_rate": 0.776, "comp_len": 612.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 41.3, "frames": {"chat": 196}, "mem_gb": 22.07}
66
+ {"step": 66, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03594996488027585, "tokens": 120000, "cumulative_loss_tokens": 7920000, "grad_norm": 0.267578125, "lr": 3e-05, "finish_rate": 0.811, "comp_len": 566.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.4, "frames": {"chat": 212}, "mem_gb": 22.09}
67
+ {"step": 67, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.032274080974639706, "tokens": 120000, "cumulative_loss_tokens": 8040000, "grad_norm": 0.2177734375, "lr": 3e-05, "finish_rate": 0.877, "comp_len": 491.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.2, "frames": {"chat": 244}, "mem_gb": 22.0}
68
+ {"step": 68, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03448318504060929, "tokens": 120000, "cumulative_loss_tokens": 8160000, "grad_norm": 0.2392578125, "lr": 3e-05, "finish_rate": 0.838, "comp_len": 540.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.2, "frames": {"chat": 222}, "mem_gb": 22.05}
69
+ {"step": 69, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.035316669923362014, "tokens": 120000, "cumulative_loss_tokens": 8280000, "grad_norm": 0.232421875, "lr": 3e-05, "finish_rate": 0.78, "comp_len": 550.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.8, "frames": {"chat": 218}, "mem_gb": 22.09}
70
+ {"step": 70, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.0348655257951313, "tokens": 120000, "cumulative_loss_tokens": 8400000, "grad_norm": 0.2353515625, "lr": 3e-05, "finish_rate": 0.913, "comp_len": 476.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 48.0, "frames": {"chat": 252}, "mem_gb": 21.97}
71
+ {"step": 71, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03988810336939059, "tokens": 120000, "cumulative_loss_tokens": 8520000, "grad_norm": 0.26171875, "lr": 3e-05, "finish_rate": 0.728, "comp_len": 594.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.8, "frames": {"chat": 202}, "mem_gb": 22.14}
72
+ {"step": 72, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.04227977843079716, "tokens": 120000, "cumulative_loss_tokens": 8640000, "grad_norm": 0.29296875, "lr": 3e-05, "finish_rate": 0.835, "comp_len": 506.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.8, "frames": {"chat": 237}, "mem_gb": 22.1}
73
+ {"step": 73, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03656210743497747, "tokens": 120000, "cumulative_loss_tokens": 8760000, "grad_norm": 0.234375, "lr": 3e-05, "finish_rate": 0.868, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.7, "frames": {"chat": 234}, "mem_gb": 22.08}
74
+ {"step": 74, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03134735133058081, "tokens": 120000, "cumulative_loss_tokens": 8880000, "grad_norm": 0.25, "lr": 3e-05, "finish_rate": 0.809, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.1, "frames": {"chat": 215}, "mem_gb": 22.1}
75
+ {"step": 75, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03046899615646495, "tokens": 120000, "cumulative_loss_tokens": 9000000, "grad_norm": 0.216796875, "lr": 3e-05, "finish_rate": 0.876, "comp_len": 512.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.2, "frames": {"chat": 234}, "mem_gb": 22.03}
76
+ {"step": 76, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.029838484331507546, "tokens": 120000, "cumulative_loss_tokens": 9120000, "grad_norm": 0.2099609375, "lr": 3e-05, "finish_rate": 0.801, "comp_len": 555.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.1, "frames": {"chat": 216}, "mem_gb": 22.08}
77
+ {"step": 77, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03392047594260269, "tokens": 120000, "cumulative_loss_tokens": 9240000, "grad_norm": 0.248046875, "lr": 3e-05, "finish_rate": 0.805, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.6, "frames": {"chat": 210}, "mem_gb": 22.05}
78
+ {"step": 78, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03507764483672411, "tokens": 120000, "cumulative_loss_tokens": 9360000, "grad_norm": 0.25, "lr": 3e-05, "finish_rate": 0.719, "comp_len": 603.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.5, "frames": {"chat": 199}, "mem_gb": 22.1}
79
+ {"step": 79, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03239439557224202, "tokens": 120000, "cumulative_loss_tokens": 9480000, "grad_norm": 0.2216796875, "lr": 3e-05, "finish_rate": 0.824, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.4, "frames": {"chat": 210}, "mem_gb": 22.11}
80
+ {"step": 80, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03286070772272845, "tokens": 120000, "cumulative_loss_tokens": 9600000, "grad_norm": 0.265625, "lr": 3e-05, "finish_rate": 0.902, "comp_len": 533.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.0, "frames": {"chat": 225}, "mem_gb": 22.05}
81
+ {"step": 81, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.032128432613688834, "tokens": 120000, "cumulative_loss_tokens": 9720000, "grad_norm": 0.25, "lr": 3e-05, "finish_rate": 0.913, "comp_len": 474.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.7, "frames": {"chat": 253}, "mem_gb": 21.95}
82
+ {"step": 82, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.031100997417598652, "tokens": 120000, "cumulative_loss_tokens": 9840000, "grad_norm": 0.21875, "lr": 3e-05, "finish_rate": 0.903, "comp_len": 485.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 47.2, "frames": {"chat": 247}, "mem_gb": 22.07}
83
+ {"step": 83, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.0324900306135416, "tokens": 120000, "cumulative_loss_tokens": 9960000, "grad_norm": 0.2412109375, "lr": 3e-05, "finish_rate": 0.836, "comp_len": 504.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.5, "frames": {"chat": 238}, "mem_gb": 22.07}
84
+ {"step": 84, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03485131102278829, "tokens": 120000, "cumulative_loss_tokens": 10080000, "grad_norm": 0.2451171875, "lr": 3e-05, "finish_rate": 0.86, "comp_len": 510.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.7, "frames": {"chat": 235}, "mem_gb": 22.09}
85
+ {"step": 85, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03912405168330297, "tokens": 120000, "cumulative_loss_tokens": 10200000, "grad_norm": 0.265625, "lr": 3e-05, "finish_rate": 0.805, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.4, "frames": {"chat": 215}, "mem_gb": 22.06}
86
+ {"step": 86, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.033098834516915185, "tokens": 120000, "cumulative_loss_tokens": 10320000, "grad_norm": 0.28125, "lr": 3e-05, "finish_rate": 0.925, "comp_len": 447.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 47.1, "frames": {"chat": 268}, "mem_gb": 22.06}
87
+ {"step": 87, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.034645479888437934, "tokens": 120000, "cumulative_loss_tokens": 10440000, "grad_norm": 0.232421875, "lr": 3e-05, "finish_rate": 0.825, "comp_len": 526.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.3, "frames": {"chat": 228}, "mem_gb": 22.1}
88
+ {"step": 88, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03237436599015103, "tokens": 120000, "cumulative_loss_tokens": 10560000, "grad_norm": 0.240234375, "lr": 3e-05, "finish_rate": 0.881, "comp_len": 476.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.7, "frames": {"chat": 252}, "mem_gb": 22.03}
89
+ {"step": 89, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03435483168452047, "tokens": 120000, "cumulative_loss_tokens": 10680000, "grad_norm": 0.2734375, "lr": 3e-05, "finish_rate": 0.821, "comp_len": 538.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.3, "frames": {"chat": 223}, "mem_gb": 22.11}
90
+ {"step": 90, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.04248241032837735, "tokens": 120000, "cumulative_loss_tokens": 10800000, "grad_norm": 0.294921875, "lr": 3e-05, "finish_rate": 0.805, "comp_len": 531.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.6, "frames": {"chat": 226}, "mem_gb": 22.09}
91
+ {"step": 91, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.04142716025705449, "tokens": 120000, "cumulative_loss_tokens": 10920000, "grad_norm": 0.283203125, "lr": 3e-05, "finish_rate": 0.731, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 47.0, "frames": {"chat": 208}, "mem_gb": 22.14}
92
+ {"step": 92, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.031166394373402, "tokens": 120000, "cumulative_loss_tokens": 11040000, "grad_norm": 0.251953125, "lr": 3e-05, "finish_rate": 0.883, "comp_len": 500.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.8, "frames": {"chat": 240}, "mem_gb": 22.03}
93
+ {"step": 93, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03443516800537861, "tokens": 120000, "cumulative_loss_tokens": 11160000, "grad_norm": 0.23046875, "lr": 3e-05, "finish_rate": 0.842, "comp_len": 540.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.9, "frames": {"chat": 222}, "mem_gb": 22.02}
94
+ {"step": 94, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.030224586171346407, "tokens": 120000, "cumulative_loss_tokens": 11280000, "grad_norm": 0.2333984375, "lr": 3e-05, "finish_rate": 0.881, "comp_len": 508.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.7, "frames": {"chat": 236}, "mem_gb": 22.09}
95
+ {"step": 95, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03236451368537576, "tokens": 120000, "cumulative_loss_tokens": 11400000, "grad_norm": 0.2353515625, "lr": 3e-05, "finish_rate": 0.834, "comp_len": 553.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.5, "frames": {"chat": 217}, "mem_gb": 22.06}
96
+ {"step": 96, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.034898156173845445, "tokens": 120000, "cumulative_loss_tokens": 11520000, "grad_norm": 0.263671875, "lr": 3e-05, "finish_rate": 0.921, "comp_len": 476.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.3, "frames": {"chat": 252}, "mem_gb": 21.97}
97
+ {"step": 97, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.02913232143558562, "tokens": 120000, "cumulative_loss_tokens": 11640000, "grad_norm": 0.2255859375, "lr": 3e-05, "finish_rate": 0.847, "comp_len": 540.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.8, "frames": {"chat": 222}, "mem_gb": 22.08}
98
+ {"step": 98, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.029040355308757475, "tokens": 120000, "cumulative_loss_tokens": 11760000, "grad_norm": 0.21484375, "lr": 3e-05, "finish_rate": 0.901, "comp_len": 495.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.6, "frames": {"chat": 242}, "mem_gb": 21.97}
99
+ {"step": 99, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.041233757554925976, "tokens": 120000, "cumulative_loss_tokens": 11880000, "grad_norm": 0.2734375, "lr": 3e-05, "finish_rate": 0.804, "comp_len": 547.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.1, "frames": {"chat": 219}, "mem_gb": 22.03}
100
+ {"step": 100, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.031569313709802614, "tokens": 120000, "cumulative_loss_tokens": 12000000, "grad_norm": 0.2138671875, "lr": 3e-05, "finish_rate": 0.879, "comp_len": 538.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.6, "frames": {"chat": 223}, "mem_gb": 22.04}
101
+ {"step": 101, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03175070390204589, "tokens": 120000, "cumulative_loss_tokens": 12120000, "grad_norm": 0.228515625, "lr": 3e-05, "finish_rate": 0.858, "comp_len": 517.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.5, "frames": {"chat": 232}, "mem_gb": 22.04}
102
+ {"step": 102, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03924315899686578, "tokens": 120000, "cumulative_loss_tokens": 12240000, "grad_norm": 0.28125, "lr": 3e-05, "finish_rate": 0.832, "comp_len": 545.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.3, "frames": {"chat": 220}, "mem_gb": 22.09}
103
+ {"step": 103, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03732912681673964, "tokens": 120000, "cumulative_loss_tokens": 12360000, "grad_norm": 0.236328125, "lr": 3e-05, "finish_rate": 0.776, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.5, "frames": {"chat": 210}, "mem_gb": 22.14}
104
+ {"step": 104, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.03158854953125119, "tokens": 120000, "cumulative_loss_tokens": 12480000, "grad_norm": 0.205078125, "lr": 3e-05, "finish_rate": 0.81, "comp_len": 531.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.0, "frames": {"chat": 226}, "mem_gb": 22.06}
105
+ {"step": 105, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.033342002084447694, "tokens": 120000, "cumulative_loss_tokens": 12600000, "grad_norm": 0.259765625, "lr": 3e-05, "finish_rate": 0.741, "comp_len": 566.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.2, "frames": {"chat": 212}, "mem_gb": 22.09}
106
+ {"step": 106, "epoch": 1, "training_mode": "off-policy", "forward_topk_kl": 0.030606075126398354, "tokens": 120000, "cumulative_loss_tokens": 12720000, "grad_norm": 0.232421875, "lr": 3e-05, "finish_rate": 0.839, "comp_len": 508.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.3, "frames": {"chat": 236}, "mem_gb": 22.1}
107
+ {"step": 107, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.021876435576969135, "tokens": 120000, "cumulative_loss_tokens": 12840000, "grad_norm": 0.2001953125, "lr": 3e-05, "finish_rate": 0.928, "comp_len": 454.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.7, "frames": {"chat": 264}, "mem_gb": 21.97}
108
+ {"step": 108, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.031198739765095525, "tokens": 120000, "cumulative_loss_tokens": 12960000, "grad_norm": 0.21875, "lr": 3e-05, "finish_rate": 0.834, "comp_len": 524.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.5, "frames": {"chat": 229}, "mem_gb": 22.07}
109
+ {"step": 109, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.02165067300312221, "tokens": 120000, "cumulative_loss_tokens": 13080000, "grad_norm": 0.1806640625, "lr": 3e-05, "finish_rate": 0.903, "comp_len": 465.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.9, "frames": {"chat": 258}, "mem_gb": 21.95}
110
+ {"step": 110, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.028640677657909692, "tokens": 120000, "cumulative_loss_tokens": 13200000, "grad_norm": 0.193359375, "lr": 3e-05, "finish_rate": 0.755, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.1, "frames": {"chat": 208}, "mem_gb": 22.11}
111
+ {"step": 111, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.022872452891978902, "tokens": 120000, "cumulative_loss_tokens": 13320000, "grad_norm": 0.18359375, "lr": 3e-05, "finish_rate": 0.88, "comp_len": 481.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.0, "frames": {"chat": 249}, "mem_gb": 22.02}
112
+ {"step": 112, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.02100934052404482, "tokens": 120000, "cumulative_loss_tokens": 13440000, "grad_norm": 0.19140625, "lr": 3e-05, "finish_rate": 0.845, "comp_len": 545.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 42.8, "frames": {"chat": 220}, "mem_gb": 22.09}
113
+ {"step": 113, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.024147335090193275, "tokens": 120000, "cumulative_loss_tokens": 13560000, "grad_norm": 0.1806640625, "lr": 3e-05, "finish_rate": 0.834, "comp_len": 538.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.2, "frames": {"chat": 223}, "mem_gb": 22.08}
114
+ {"step": 114, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.024052943600167055, "tokens": 120000, "cumulative_loss_tokens": 13680000, "grad_norm": 0.201171875, "lr": 3e-05, "finish_rate": 0.833, "comp_len": 543.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 43.7, "frames": {"chat": 221}, "mem_gb": 22.09}
115
+ {"step": 115, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.02166470361133882, "tokens": 120000, "cumulative_loss_tokens": 13800000, "grad_norm": 0.162109375, "lr": 3e-05, "finish_rate": 0.9, "comp_len": 521.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.5, "frames": {"chat": 230}, "mem_gb": 22.0}
116
+ {"step": 116, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.02461200471681077, "tokens": 120000, "cumulative_loss_tokens": 13920000, "grad_norm": 0.171875, "lr": 3e-05, "finish_rate": 0.776, "comp_len": 560.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.2, "frames": {"chat": 214}, "mem_gb": 22.07}
117
+ {"step": 117, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.03192179344429945, "tokens": 120000, "cumulative_loss_tokens": 14040000, "grad_norm": 0.224609375, "lr": 3e-05, "finish_rate": 0.766, "comp_len": 560.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 46.2, "frames": {"chat": 214}, "mem_gb": 22.08}
118
+ {"step": 118, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.027875469126327275, "tokens": 120000, "cumulative_loss_tokens": 14160000, "grad_norm": 0.2041015625, "lr": 3e-05, "finish_rate": 0.786, "comp_len": 571.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 44.4, "frames": {"chat": 210}, "mem_gb": 22.13}
119
+ {"step": 119, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.028064899455732668, "tokens": 120000, "cumulative_loss_tokens": 14280000, "grad_norm": 0.1845703125, "lr": 3e-05, "finish_rate": 0.776, "comp_len": 560.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 45.7, "frames": {"chat": 214}, "mem_gb": 22.09}
120
+ {"step": 120, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.02598108633084533, "tokens": 120000, "cumulative_loss_tokens": 14400000, "grad_norm": 0.185546875, "lr": 3e-05, "finish_rate": 0.791, "comp_len": 558.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 51.0, "frames": {"chat": 215}, "mem_gb": 22.05}
121
+ {"step": 121, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.029542491081776097, "tokens": 120000, "cumulative_loss_tokens": 14520000, "grad_norm": 0.2060546875, "lr": 3e-05, "finish_rate": 0.721, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 49.9, "frames": {"chat": 208}, "mem_gb": 22.09}
122
+ {"step": 122, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.023529452410464485, "tokens": 120000, "cumulative_loss_tokens": 14640000, "grad_norm": 0.201171875, "lr": 3e-05, "finish_rate": 0.789, "comp_len": 550.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 49.8, "frames": {"chat": 218}, "mem_gb": 21.97}
123
+ {"step": 123, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.021247040836008577, "tokens": 120000, "cumulative_loss_tokens": 14760000, "grad_norm": 0.16015625, "lr": 3e-05, "finish_rate": 0.876, "comp_len": 515.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 49.4, "frames": {"chat": 233}, "mem_gb": 21.99}
124
+ {"step": 124, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.02059493543021381, "tokens": 120000, "cumulative_loss_tokens": 14880000, "grad_norm": 0.173828125, "lr": 3e-05, "finish_rate": 0.861, "comp_len": 519.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 50.3, "frames": {"chat": 231}, "mem_gb": 22.03}
125
+ {"step": 125, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.024373427555907982, "tokens": 120000, "cumulative_loss_tokens": 15000000, "grad_norm": 0.1845703125, "lr": 3e-05, "finish_rate": 0.868, "comp_len": 510.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 52.0, "frames": {"chat": 235}, "mem_gb": 22.22}
126
+ {"step": 126, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.02307887886830916, "tokens": 120000, "cumulative_loss_tokens": 15120000, "grad_norm": 0.1650390625, "lr": 3e-05, "finish_rate": 0.843, "comp_len": 555.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 48.3, "frames": {"chat": 216}, "mem_gb": 22.08}
127
+ {"step": 127, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.02126138604179335, "tokens": 120000, "cumulative_loss_tokens": 15240000, "grad_norm": 0.162109375, "lr": 3e-05, "finish_rate": 0.831, "comp_len": 506.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 50.3, "frames": {"chat": 237}, "mem_gb": 22.1}
128
+ {"step": 128, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.029153408732265233, "tokens": 120000, "cumulative_loss_tokens": 15360000, "grad_norm": 0.2001953125, "lr": 3e-05, "finish_rate": 0.734, "comp_len": 591.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 49.1, "frames": {"chat": 203}, "mem_gb": 22.11}
129
+ {"step": 129, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.02260789417422687, "tokens": 120000, "cumulative_loss_tokens": 15480000, "grad_norm": 0.17578125, "lr": 3e-05, "finish_rate": 0.873, "comp_len": 508.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 50.6, "frames": {"chat": 236}, "mem_gb": 22.13}
130
+ {"step": 130, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.022835769553498055, "tokens": 120000, "cumulative_loss_tokens": 15600000, "grad_norm": 0.1728515625, "lr": 3e-05, "finish_rate": 0.734, "comp_len": 560.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 48.0, "frames": {"chat": 214}, "mem_gb": 22.1}
131
+ {"step": 131, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.026547468842783323, "tokens": 120000, "cumulative_loss_tokens": 15720000, "grad_norm": 0.189453125, "lr": 3e-05, "finish_rate": 0.78, "comp_len": 574.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 49.3, "frames": {"chat": 209}, "mem_gb": 22.09}
132
+ {"step": 132, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.022498934920489166, "tokens": 120000, "cumulative_loss_tokens": 15840000, "grad_norm": 0.1826171875, "lr": 3e-05, "finish_rate": 0.906, "comp_len": 468.8, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 52.3, "frames": {"chat": 256}, "mem_gb": 22.1}
133
+ {"step": 133, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.021946324599829193, "tokens": 120000, "cumulative_loss_tokens": 15960000, "grad_norm": 0.1787109375, "lr": 3e-05, "finish_rate": 0.878, "comp_len": 521.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 49.6, "frames": {"chat": 230}, "mem_gb": 21.96}
134
+ {"step": 134, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.024039489145313078, "tokens": 120000, "cumulative_loss_tokens": 16080000, "grad_norm": 0.18359375, "lr": 3e-05, "finish_rate": 0.822, "comp_len": 521.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 49.9, "frames": {"chat": 230}, "mem_gb": 22.15}
135
+ {"step": 135, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.02589866460277699, "tokens": 120000, "cumulative_loss_tokens": 16200000, "grad_norm": 0.1796875, "lr": 3e-05, "finish_rate": 0.881, "comp_len": 528.6, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 48.8, "frames": {"chat": 227}, "mem_gb": 22.05}
136
+ {"step": 136, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.023664263719754913, "tokens": 120000, "cumulative_loss_tokens": 16320000, "grad_norm": 0.166015625, "lr": 3e-05, "finish_rate": 0.755, "comp_len": 576.9, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 47.4, "frames": {"chat": 208}, "mem_gb": 22.11}
137
+ {"step": 137, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.027114476467870796, "tokens": 120000, "cumulative_loss_tokens": 16440000, "grad_norm": 0.1796875, "lr": 3e-05, "finish_rate": 0.699, "comp_len": 582.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 49.0, "frames": {"chat": 206}, "mem_gb": 22.12}
138
+ {"step": 138, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.02399630657126351, "tokens": 120000, "cumulative_loss_tokens": 16560000, "grad_norm": 0.1767578125, "lr": 3e-05, "finish_rate": 0.82, "comp_len": 526.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 48.7, "frames": {"chat": 228}, "mem_gb": 22.0}
139
+ {"step": 139, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.02343961157395194, "tokens": 120000, "cumulative_loss_tokens": 16680000, "grad_norm": 0.181640625, "lr": 3e-05, "finish_rate": 0.835, "comp_len": 535.7, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 48.9, "frames": {"chat": 224}, "mem_gb": 22.09}
140
+ {"step": 140, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.024890781568270178, "tokens": 120000, "cumulative_loss_tokens": 16800000, "grad_norm": 0.1962890625, "lr": 3e-05, "finish_rate": 0.66, "comp_len": 600.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 48.1, "frames": {"chat": 200}, "mem_gb": 22.13}
141
+ {"step": 141, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.02478467862996428, "tokens": 120000, "cumulative_loss_tokens": 16920000, "grad_norm": 0.1904296875, "lr": 3e-05, "finish_rate": 0.714, "comp_len": 612.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 47.1, "frames": {"chat": 196}, "mem_gb": 22.11}
142
+ {"step": 142, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.022272795487005108, "tokens": 120000, "cumulative_loss_tokens": 17040000, "grad_norm": 0.169921875, "lr": 3e-05, "finish_rate": 0.834, "comp_len": 538.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 49.4, "frames": {"chat": 223}, "mem_gb": 22.09}
143
+ {"step": 143, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.022915525561606045, "tokens": 120000, "cumulative_loss_tokens": 17160000, "grad_norm": 0.18359375, "lr": 3e-05, "finish_rate": 0.869, "comp_len": 563.4, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 47.0, "frames": {"chat": 213}, "mem_gb": 21.98}
144
+ {"step": 144, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.0212597823954653, "tokens": 120000, "cumulative_loss_tokens": 17280000, "grad_norm": 0.1953125, "lr": 3e-05, "finish_rate": 0.879, "comp_len": 517.2, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 49.4, "frames": {"chat": 232}, "mem_gb": 22.02}
145
+ {"step": 145, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.020988663152477237, "tokens": 120000, "cumulative_loss_tokens": 17400000, "grad_norm": 0.1796875, "lr": 3e-05, "finish_rate": 0.861, "comp_len": 538.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 50.2, "frames": {"chat": 223}, "mem_gb": 22.02}
146
+ {"step": 146, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.021486785067555806, "tokens": 120000, "cumulative_loss_tokens": 17520000, "grad_norm": 0.1630859375, "lr": 3e-05, "finish_rate": 0.85, "comp_len": 515.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 49.0, "frames": {"chat": 233}, "mem_gb": 22.11}
147
+ {"step": 147, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.02550031017503546, "tokens": 120000, "cumulative_loss_tokens": 17640000, "grad_norm": 0.2021484375, "lr": 3e-05, "finish_rate": 0.816, "comp_len": 553.0, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 47.8, "frames": {"chat": 217}, "mem_gb": 22.11}
148
+ {"step": 148, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.030582861898102175, "tokens": 120000, "cumulative_loss_tokens": 17760000, "grad_norm": 0.1962890625, "lr": 3e-05, "finish_rate": 0.752, "comp_len": 594.1, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 48.1, "frames": {"chat": 202}, "mem_gb": 22.17}
149
+ {"step": 149, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.021075663964031262, "tokens": 120000, "cumulative_loss_tokens": 17880000, "grad_norm": 0.1640625, "lr": 3e-05, "finish_rate": 0.858, "comp_len": 474.3, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 51.1, "frames": {"chat": 253}, "mem_gb": 22.03}
150
+ {"step": 150, "epoch": 2, "training_mode": "off-policy", "forward_topk_kl": 0.02149440993195555, "tokens": 120000, "cumulative_loss_tokens": 18000000, "grad_norm": 0.30859375, "lr": 3e-05, "finish_rate": 0.879, "comp_len": 519.5, "t_data_s": 0.0, "t_rollout_s": 0.0, "t_step_s": 49.8, "frames": {"chat": 231}, "mem_gb": 22.03}
healed/liger_mb8/step0004/chat_template.jinja ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {{ bos_token }}{% for message in messages %}{% if message['role'] == 'system' %}{{ '<|system|>
2
+ ' + message['content'] + '
3
+ ' }}{% elif message['role'] == 'user' %}{{ '<|user|>
4
+ ' + message['content'] + '
5
+ ' }}{% elif message['role'] == 'assistant' %}{% if not loop.last %}{{ '<|assistant|>
6
+ ' + message['content'] + eos_token + '
7
+ ' }}{% else %}{{ '<|assistant|>
8
+ ' + message['content'] + eos_token }}{% endif %}{% endif %}{% if loop.last and add_generation_prompt %}{{ '<|assistant|>
9
+ ' }}{% endif %}{% endfor %}
healed/liger_mb8/step0004/config.json ADDED
@@ -0,0 +1,887 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "PrunedOlmoeForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_pruned_olmoe.PrunedOlmoeConfig",
9
+ "AutoModelForCausalLM": "modeling_pruned_olmoe.PrunedOlmoeForCausalLM"
10
+ },
11
+ "clip_qkv": null,
12
+ "dtype": "bfloat16",
13
+ "eos_token_id": 50279,
14
+ "expert_widths": [
15
+ [
16
+ 1024,
17
+ 384,
18
+ 256,
19
+ 256,
20
+ 768,
21
+ 1024,
22
+ 128,
23
+ 768,
24
+ 384,
25
+ 128,
26
+ 1024,
27
+ 384,
28
+ 384,
29
+ 128,
30
+ 640,
31
+ 896,
32
+ 896,
33
+ 256,
34
+ 768,
35
+ 128,
36
+ 768,
37
+ 768,
38
+ 768,
39
+ 768,
40
+ 640,
41
+ 896,
42
+ 256,
43
+ 128,
44
+ 384,
45
+ 896,
46
+ 256,
47
+ 768,
48
+ 512,
49
+ 640,
50
+ 768,
51
+ 896,
52
+ 1024,
53
+ 384,
54
+ 384,
55
+ 512,
56
+ 384,
57
+ 256,
58
+ 768,
59
+ 640,
60
+ 896,
61
+ 896,
62
+ 512,
63
+ 768,
64
+ 512,
65
+ 384,
66
+ 1024,
67
+ 896,
68
+ 896,
69
+ 768,
70
+ 128
71
+ ],
72
+ [
73
+ 768,
74
+ 896,
75
+ 896,
76
+ 256,
77
+ 896,
78
+ 384,
79
+ 768,
80
+ 768,
81
+ 1024,
82
+ 384,
83
+ 512,
84
+ 1024,
85
+ 768,
86
+ 768,
87
+ 1024,
88
+ 640,
89
+ 128,
90
+ 768,
91
+ 384,
92
+ 768,
93
+ 128,
94
+ 256,
95
+ 512,
96
+ 384,
97
+ 896,
98
+ 256,
99
+ 1024,
100
+ 1024,
101
+ 896,
102
+ 1024,
103
+ 896,
104
+ 768,
105
+ 256,
106
+ 768,
107
+ 128,
108
+ 1024,
109
+ 768,
110
+ 768,
111
+ 640,
112
+ 384,
113
+ 1024,
114
+ 384,
115
+ 640,
116
+ 896,
117
+ 1024,
118
+ 512,
119
+ 512,
120
+ 512,
121
+ 640
122
+ ],
123
+ [
124
+ 768,
125
+ 1024,
126
+ 512,
127
+ 384,
128
+ 384,
129
+ 768,
130
+ 256,
131
+ 384,
132
+ 1024,
133
+ 896,
134
+ 640,
135
+ 256,
136
+ 1024,
137
+ 1024,
138
+ 1024,
139
+ 640,
140
+ 1024,
141
+ 896,
142
+ 512,
143
+ 768,
144
+ 128,
145
+ 768,
146
+ 1024,
147
+ 768,
148
+ 256,
149
+ 1024,
150
+ 896,
151
+ 1024,
152
+ 896,
153
+ 512,
154
+ 640,
155
+ 384,
156
+ 896,
157
+ 384,
158
+ 512,
159
+ 640,
160
+ 256,
161
+ 768,
162
+ 256,
163
+ 768,
164
+ 768,
165
+ 640,
166
+ 1024,
167
+ 640,
168
+ 896,
169
+ 768,
170
+ 768,
171
+ 256
172
+ ],
173
+ [
174
+ 640,
175
+ 640,
176
+ 896,
177
+ 512,
178
+ 640,
179
+ 896,
180
+ 896,
181
+ 384,
182
+ 896,
183
+ 896,
184
+ 384,
185
+ 896,
186
+ 640,
187
+ 256,
188
+ 640,
189
+ 1024,
190
+ 1024,
191
+ 768,
192
+ 1024,
193
+ 896,
194
+ 1024,
195
+ 768,
196
+ 896,
197
+ 256,
198
+ 512,
199
+ 768,
200
+ 1024,
201
+ 256,
202
+ 768,
203
+ 512,
204
+ 256,
205
+ 640,
206
+ 1024,
207
+ 1024,
208
+ 512,
209
+ 1024,
210
+ 768,
211
+ 256,
212
+ 1024,
213
+ 384,
214
+ 896,
215
+ 1024,
216
+ 896,
217
+ 1024,
218
+ 384
219
+ ],
220
+ [
221
+ 256,
222
+ 640,
223
+ 640,
224
+ 896,
225
+ 768,
226
+ 896,
227
+ 768,
228
+ 768,
229
+ 896,
230
+ 896,
231
+ 1024,
232
+ 256,
233
+ 512,
234
+ 1024,
235
+ 640,
236
+ 896,
237
+ 512,
238
+ 512,
239
+ 384,
240
+ 384,
241
+ 256,
242
+ 384,
243
+ 384,
244
+ 896,
245
+ 896,
246
+ 768,
247
+ 640,
248
+ 896,
249
+ 768,
250
+ 1024,
251
+ 512,
252
+ 640,
253
+ 512,
254
+ 640,
255
+ 896,
256
+ 512,
257
+ 512,
258
+ 384,
259
+ 640,
260
+ 896,
261
+ 896,
262
+ 896,
263
+ 1024,
264
+ 640,
265
+ 1024,
266
+ 640,
267
+ 1024
268
+ ],
269
+ [
270
+ 1024,
271
+ 512,
272
+ 1024,
273
+ 1024,
274
+ 640,
275
+ 896,
276
+ 640,
277
+ 1024,
278
+ 896,
279
+ 384,
280
+ 1024,
281
+ 128,
282
+ 896,
283
+ 768,
284
+ 1024,
285
+ 768,
286
+ 640,
287
+ 896,
288
+ 768,
289
+ 640,
290
+ 512,
291
+ 896,
292
+ 512,
293
+ 640,
294
+ 256,
295
+ 768,
296
+ 640,
297
+ 768,
298
+ 384,
299
+ 896,
300
+ 512,
301
+ 512,
302
+ 256,
303
+ 512,
304
+ 896,
305
+ 256,
306
+ 384,
307
+ 640,
308
+ 512,
309
+ 640,
310
+ 896,
311
+ 512,
312
+ 1024,
313
+ 256,
314
+ 768,
315
+ 1024,
316
+ 768,
317
+ 256,
318
+ 256
319
+ ],
320
+ [
321
+ 640,
322
+ 896,
323
+ 1024,
324
+ 896,
325
+ 1024,
326
+ 1024,
327
+ 1024,
328
+ 512,
329
+ 256,
330
+ 256,
331
+ 1024,
332
+ 768,
333
+ 512,
334
+ 768,
335
+ 1024,
336
+ 1024,
337
+ 1024,
338
+ 384,
339
+ 512,
340
+ 1024,
341
+ 512,
342
+ 1024,
343
+ 128,
344
+ 640,
345
+ 640,
346
+ 896,
347
+ 768,
348
+ 128,
349
+ 256,
350
+ 256,
351
+ 256,
352
+ 256,
353
+ 896,
354
+ 1024,
355
+ 1024,
356
+ 384,
357
+ 896,
358
+ 256,
359
+ 896,
360
+ 640,
361
+ 1024,
362
+ 384,
363
+ 640,
364
+ 256,
365
+ 1024,
366
+ 1024,
367
+ 1024
368
+ ],
369
+ [
370
+ 1024,
371
+ 384,
372
+ 1024,
373
+ 1024,
374
+ 256,
375
+ 128,
376
+ 256,
377
+ 384,
378
+ 256,
379
+ 384,
380
+ 896,
381
+ 768,
382
+ 896,
383
+ 896,
384
+ 512,
385
+ 896,
386
+ 640,
387
+ 384,
388
+ 384,
389
+ 896,
390
+ 768,
391
+ 384,
392
+ 896,
393
+ 768,
394
+ 768,
395
+ 512,
396
+ 896,
397
+ 768,
398
+ 768,
399
+ 896,
400
+ 128,
401
+ 896,
402
+ 512,
403
+ 256,
404
+ 768,
405
+ 128,
406
+ 384,
407
+ 256,
408
+ 896,
409
+ 896,
410
+ 384,
411
+ 768,
412
+ 512,
413
+ 640,
414
+ 256,
415
+ 768,
416
+ 640,
417
+ 896,
418
+ 384,
419
+ 512,
420
+ 1024,
421
+ 768,
422
+ 384
423
+ ],
424
+ [
425
+ 512,
426
+ 768,
427
+ 512,
428
+ 256,
429
+ 128,
430
+ 640,
431
+ 384,
432
+ 640,
433
+ 768,
434
+ 896,
435
+ 640,
436
+ 768,
437
+ 256,
438
+ 384,
439
+ 1024,
440
+ 896,
441
+ 256,
442
+ 896,
443
+ 512,
444
+ 256,
445
+ 896,
446
+ 768,
447
+ 256,
448
+ 896,
449
+ 896,
450
+ 384,
451
+ 896,
452
+ 640,
453
+ 768,
454
+ 512,
455
+ 768,
456
+ 768,
457
+ 1024,
458
+ 768,
459
+ 640,
460
+ 768,
461
+ 384,
462
+ 256,
463
+ 512,
464
+ 896,
465
+ 128,
466
+ 384,
467
+ 256,
468
+ 768,
469
+ 384,
470
+ 256,
471
+ 1024,
472
+ 1024,
473
+ 896,
474
+ 256,
475
+ 1024,
476
+ 256,
477
+ 128,
478
+ 896
479
+ ],
480
+ [
481
+ 640,
482
+ 640,
483
+ 896,
484
+ 256,
485
+ 1024,
486
+ 512,
487
+ 1024,
488
+ 768,
489
+ 384,
490
+ 512,
491
+ 256,
492
+ 768,
493
+ 896,
494
+ 768,
495
+ 512,
496
+ 768,
497
+ 768,
498
+ 640,
499
+ 384,
500
+ 768,
501
+ 512,
502
+ 768,
503
+ 768,
504
+ 512,
505
+ 768,
506
+ 128,
507
+ 896,
508
+ 512,
509
+ 768,
510
+ 1024,
511
+ 128,
512
+ 384,
513
+ 768,
514
+ 768,
515
+ 768,
516
+ 384,
517
+ 512,
518
+ 640,
519
+ 768,
520
+ 512,
521
+ 768,
522
+ 1024,
523
+ 640,
524
+ 896,
525
+ 256,
526
+ 1024,
527
+ 384,
528
+ 768,
529
+ 768,
530
+ 768
531
+ ],
532
+ [
533
+ 896,
534
+ 512,
535
+ 896,
536
+ 768,
537
+ 384,
538
+ 384,
539
+ 768,
540
+ 512,
541
+ 768,
542
+ 512,
543
+ 1024,
544
+ 640,
545
+ 896,
546
+ 896,
547
+ 256,
548
+ 640,
549
+ 1024,
550
+ 256,
551
+ 896,
552
+ 128,
553
+ 128,
554
+ 128,
555
+ 768,
556
+ 896,
557
+ 384,
558
+ 896,
559
+ 512,
560
+ 896,
561
+ 384,
562
+ 256,
563
+ 640,
564
+ 640,
565
+ 896,
566
+ 768,
567
+ 640,
568
+ 256,
569
+ 896,
570
+ 896,
571
+ 512,
572
+ 128,
573
+ 896,
574
+ 256,
575
+ 256,
576
+ 640,
577
+ 896,
578
+ 896,
579
+ 128,
580
+ 1024,
581
+ 256,
582
+ 384,
583
+ 1024,
584
+ 640,
585
+ 896
586
+ ],
587
+ [
588
+ 768,
589
+ 384,
590
+ 640,
591
+ 896,
592
+ 256,
593
+ 128,
594
+ 384,
595
+ 896,
596
+ 128,
597
+ 128,
598
+ 896,
599
+ 256,
600
+ 384,
601
+ 896,
602
+ 512,
603
+ 768,
604
+ 768,
605
+ 512,
606
+ 512,
607
+ 768,
608
+ 896,
609
+ 640,
610
+ 768,
611
+ 896,
612
+ 896,
613
+ 896,
614
+ 640,
615
+ 896,
616
+ 640,
617
+ 512,
618
+ 896,
619
+ 256,
620
+ 512,
621
+ 128,
622
+ 512,
623
+ 384,
624
+ 768,
625
+ 768,
626
+ 1024,
627
+ 256,
628
+ 768,
629
+ 256,
630
+ 768,
631
+ 512,
632
+ 640,
633
+ 1024,
634
+ 128,
635
+ 896,
636
+ 896,
637
+ 896,
638
+ 896,
639
+ 1024
640
+ ],
641
+ [
642
+ 512,
643
+ 384,
644
+ 768,
645
+ 640,
646
+ 640,
647
+ 768,
648
+ 1024,
649
+ 896,
650
+ 512,
651
+ 256,
652
+ 640,
653
+ 768,
654
+ 640,
655
+ 896,
656
+ 128,
657
+ 256,
658
+ 896,
659
+ 1024,
660
+ 256,
661
+ 640,
662
+ 512,
663
+ 256,
664
+ 128,
665
+ 512,
666
+ 256,
667
+ 640,
668
+ 768,
669
+ 768,
670
+ 128,
671
+ 128,
672
+ 768,
673
+ 640,
674
+ 1024,
675
+ 1024,
676
+ 768,
677
+ 512,
678
+ 896,
679
+ 768,
680
+ 896,
681
+ 1024,
682
+ 896,
683
+ 896,
684
+ 512,
685
+ 640,
686
+ 1024,
687
+ 512,
688
+ 1024,
689
+ 512,
690
+ 512,
691
+ 512,
692
+ 768
693
+ ],
694
+ [
695
+ 896,
696
+ 896,
697
+ 1024,
698
+ 1024,
699
+ 896,
700
+ 128,
701
+ 768,
702
+ 256,
703
+ 1024,
704
+ 256,
705
+ 1024,
706
+ 640,
707
+ 384,
708
+ 256,
709
+ 256,
710
+ 512,
711
+ 768,
712
+ 896,
713
+ 512,
714
+ 768,
715
+ 384,
716
+ 1024,
717
+ 896,
718
+ 896,
719
+ 1024,
720
+ 896,
721
+ 768,
722
+ 896,
723
+ 640,
724
+ 1024,
725
+ 512,
726
+ 896,
727
+ 512,
728
+ 1024,
729
+ 512,
730
+ 512,
731
+ 256,
732
+ 256,
733
+ 256,
734
+ 512,
735
+ 768,
736
+ 128,
737
+ 384,
738
+ 512,
739
+ 896,
740
+ 896,
741
+ 1024,
742
+ 256,
743
+ 384,
744
+ 640
745
+ ],
746
+ [
747
+ 896,
748
+ 640,
749
+ 384,
750
+ 512,
751
+ 256,
752
+ 640,
753
+ 1024,
754
+ 384,
755
+ 1024,
756
+ 1024,
757
+ 768,
758
+ 256,
759
+ 1024,
760
+ 768,
761
+ 512,
762
+ 896,
763
+ 256,
764
+ 1024,
765
+ 768,
766
+ 768,
767
+ 768,
768
+ 384,
769
+ 384,
770
+ 256,
771
+ 1024,
772
+ 384,
773
+ 384,
774
+ 384,
775
+ 896,
776
+ 768,
777
+ 640,
778
+ 768,
779
+ 512,
780
+ 896,
781
+ 896,
782
+ 896,
783
+ 896,
784
+ 256,
785
+ 384,
786
+ 128,
787
+ 1024,
788
+ 896,
789
+ 256,
790
+ 256,
791
+ 768,
792
+ 640,
793
+ 896,
794
+ 384,
795
+ 768,
796
+ 512,
797
+ 640
798
+ ],
799
+ [
800
+ 896,
801
+ 1024,
802
+ 768,
803
+ 1024,
804
+ 896,
805
+ 256,
806
+ 768,
807
+ 128,
808
+ 128,
809
+ 768,
810
+ 512,
811
+ 896,
812
+ 384,
813
+ 768,
814
+ 1024,
815
+ 256,
816
+ 768,
817
+ 768,
818
+ 256,
819
+ 512,
820
+ 512,
821
+ 640,
822
+ 512,
823
+ 256,
824
+ 768,
825
+ 896,
826
+ 384,
827
+ 1024,
828
+ 640,
829
+ 1024,
830
+ 512,
831
+ 512,
832
+ 384,
833
+ 512,
834
+ 512,
835
+ 1024,
836
+ 384,
837
+ 896,
838
+ 768,
839
+ 384,
840
+ 384,
841
+ 128,
842
+ 384,
843
+ 1024,
844
+ 896,
845
+ 640,
846
+ 768,
847
+ 768,
848
+ 256,
849
+ 640,
850
+ 512,
851
+ 640,
852
+ 384
853
+ ]
854
+ ],
855
+ "glean_metadata": {
856
+ "base_model": "allenai/OLMoE-1B-7B-0125-Instruct",
857
+ "block_size": 128,
858
+ "criterion": "reap",
859
+ "dead_experts": 217,
860
+ "keep_fraction": 0.5,
861
+ "min_width": 128,
862
+ "params": 3697491968,
863
+ "scores": "outputs/scores_0125inst_dolmino-math/scores.pt"
864
+ },
865
+ "hidden_act": "silu",
866
+ "hidden_size": 2048,
867
+ "initializer_range": 0.02,
868
+ "intermediate_size": 1024,
869
+ "max_position_embeddings": 4096,
870
+ "model_type": "pruned_olmoe",
871
+ "norm_topk_prob": false,
872
+ "num_attention_heads": 16,
873
+ "num_experts": 64,
874
+ "num_experts_per_tok": 8,
875
+ "num_hidden_layers": 16,
876
+ "num_key_value_heads": 16,
877
+ "output_router_logits": false,
878
+ "pad_token_id": 1,
879
+ "rms_norm_eps": 1e-05,
880
+ "rope_scaling": null,
881
+ "rope_theta": 10000.0,
882
+ "router_aux_loss_coef": 0.01,
883
+ "tie_word_embeddings": false,
884
+ "transformers_version": "4.57.6",
885
+ "use_cache": false,
886
+ "vocab_size": 50304
887
+ }
healed/liger_mb8/step0004/configuration_pruned_olmoe.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Configuration for GLEAN-pruned OLMoE: variable-width, variable-count experts.
2
+ """
3
+
4
+ from transformers.models.olmoe.configuration_olmoe import OlmoeConfig
5
+
6
+
7
+ class PrunedOlmoeConfig(OlmoeConfig):
8
+ """OlmoeConfig plus a per-(layer, expert) width table.
9
+
10
+ ``expert_widths[l]`` lists the SwiGLU intermediate width of each surviving
11
+ expert in decoder layer ``l``, in expert order. Lists are ragged: layers
12
+ may keep different numbers of experts (deleted experts simply don't
13
+ appear — the router in layer ``l`` has ``len(expert_widths[l])`` rows),
14
+ and each width may differ (multiples of the GEMM block size, 128, for
15
+ variable-MegaBlocks execution). ``None`` means an unpruned model
16
+ (uniform ``num_experts`` × ``intermediate_size``).
17
+
18
+ The inherited ``num_experts`` / ``intermediate_size`` keep their ORIGINAL
19
+ (pre-pruning) values for provenance; the width table is authoritative for
20
+ the built architecture.
21
+ """
22
+
23
+ model_type = "pruned_olmoe"
24
+
25
+ def __init__(self, expert_widths: list[list[int]] | None = None, **kwargs):
26
+ super().__init__(**kwargs)
27
+ self.expert_widths = expert_widths
healed/liger_mb8/step0004/generation_config.json ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "eos_token_id": 50279,
4
+ "pad_token_id": 1,
5
+ "transformers_version": "4.57.6"
6
+ }
healed/liger_mb8/step0004/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
healed/liger_mb8/step0004/modeling_pruned_olmoe.py ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """GLEAN-pruned OLMoE: HF-loadable model with ragged (variable-width) experts.
2
+
3
+ Pattern follows hbfreed/variable-flex-olmo's PrunedFlexOlmoForCausalLM
4
+ (docs/recon/prior-work-hbfreed.md), generalized from one scalar width to a
5
+ per-(layer, expert) width table: ``super().__init__`` builds the uniform
6
+ architecture from the config, then every MoE block is rebuilt to its pruned
7
+ shape — surviving experts only, each at its own width, router sliced to
8
+ match — so the state dict aligns exactly with what
9
+ ``glean.prune.prune_channels_global`` leaves behind.
10
+
11
+ Caveat: ``output_router_logits=True`` (the load-balancing aux loss) assumes a
12
+ uniform ``config.num_experts`` and is unsupported on ragged models.
13
+ """
14
+
15
+ import torch.nn as nn
16
+ from transformers.activations import ACT2FN
17
+ from transformers.models.olmoe.modeling_olmoe import OlmoeForCausalLM
18
+
19
+ from .configuration_pruned_olmoe import PrunedOlmoeConfig
20
+
21
+
22
+ class RaggedOlmoeMLP(nn.Module):
23
+ """OlmoeMLP with an explicit intermediate width (SwiGLU, no biases)."""
24
+
25
+ def __init__(self, hidden_size: int, intermediate_size: int, hidden_act: str):
26
+ super().__init__()
27
+ self.hidden_size = hidden_size
28
+ self.intermediate_size = intermediate_size
29
+ self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
30
+ self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
31
+ self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
32
+ self.act_fn = ACT2FN[hidden_act]
33
+
34
+ def forward(self, x):
35
+ return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
36
+
37
+
38
+ class PrunedOlmoeForCausalLM(OlmoeForCausalLM):
39
+ """OLMoE with per-layer surviving-expert lists at per-expert widths."""
40
+
41
+ config_class = PrunedOlmoeConfig
42
+
43
+ def __init__(self, config: PrunedOlmoeConfig):
44
+ super().__init__(config)
45
+ widths_table = getattr(config, "expert_widths", None)
46
+ if widths_table is None:
47
+ return # unpruned: plain OLMoE
48
+ if len(widths_table) != len(self.model.layers):
49
+ raise ValueError(
50
+ f"expert_widths has {len(widths_table)} rows but the model has "
51
+ f"{len(self.model.layers)} decoder layers"
52
+ )
53
+ for layer, widths in zip(self.model.layers, widths_table):
54
+ if any(w <= 0 for w in widths):
55
+ raise ValueError("expert_widths must list surviving experts only (>0)")
56
+ block = layer.mlp
57
+ if len(widths) < block.top_k:
58
+ raise ValueError(
59
+ f"a layer keeps {len(widths)} experts < top_k={block.top_k}"
60
+ )
61
+ block.num_experts = len(widths)
62
+ block.gate = nn.Linear(config.hidden_size, len(widths), bias=False)
63
+ block.experts = nn.ModuleList(
64
+ RaggedOlmoeMLP(config.hidden_size, w, config.hidden_act)
65
+ for w in widths
66
+ )
healed/liger_mb8/step0004/special_tokens_map.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "|||IP_ADDRESS|||",
4
+ "lstrip": false,
5
+ "normalized": true,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "|||IP_ADDRESS|||",
11
+ "lstrip": false,
12
+ "normalized": true,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "pad_token": {
17
+ "content": "<pad>",
18
+ "lstrip": false,
19
+ "normalized": false,
20
+ "rstrip": false,
21
+ "single_word": false
22
+ }
23
+ }
healed/liger_mb8/step0004/tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
healed/liger_mb8/step0004/tokenizer_config.json ADDED
@@ -0,0 +1,247 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": false,
3
+ "add_eos_token": false,
4
+ "add_prefix_space": false,
5
+ "added_tokens_decoder": {
6
+ "0": {
7
+ "content": "<|endoftext|>",
8
+ "lstrip": false,
9
+ "normalized": false,
10
+ "rstrip": false,
11
+ "single_word": false,
12
+ "special": true
13
+ },
14
+ "1": {
15
+ "content": "<|padding|>",
16
+ "lstrip": false,
17
+ "normalized": false,
18
+ "rstrip": false,
19
+ "single_word": false,
20
+ "special": true
21
+ },
22
+ "50254": {
23
+ "content": " ",
24
+ "lstrip": false,
25
+ "normalized": true,
26
+ "rstrip": false,
27
+ "single_word": false,
28
+ "special": false
29
+ },
30
+ "50255": {
31
+ "content": " ",
32
+ "lstrip": false,
33
+ "normalized": true,
34
+ "rstrip": false,
35
+ "single_word": false,
36
+ "special": false
37
+ },
38
+ "50256": {
39
+ "content": " ",
40
+ "lstrip": false,
41
+ "normalized": true,
42
+ "rstrip": false,
43
+ "single_word": false,
44
+ "special": false
45
+ },
46
+ "50257": {
47
+ "content": " ",
48
+ "lstrip": false,
49
+ "normalized": true,
50
+ "rstrip": false,
51
+ "single_word": false,
52
+ "special": false
53
+ },
54
+ "50258": {
55
+ "content": " ",
56
+ "lstrip": false,
57
+ "normalized": true,
58
+ "rstrip": false,
59
+ "single_word": false,
60
+ "special": false
61
+ },
62
+ "50259": {
63
+ "content": " ",
64
+ "lstrip": false,
65
+ "normalized": true,
66
+ "rstrip": false,
67
+ "single_word": false,
68
+ "special": false
69
+ },
70
+ "50260": {
71
+ "content": " ",
72
+ "lstrip": false,
73
+ "normalized": true,
74
+ "rstrip": false,
75
+ "single_word": false,
76
+ "special": false
77
+ },
78
+ "50261": {
79
+ "content": " ",
80
+ "lstrip": false,
81
+ "normalized": true,
82
+ "rstrip": false,
83
+ "single_word": false,
84
+ "special": false
85
+ },
86
+ "50262": {
87
+ "content": " ",
88
+ "lstrip": false,
89
+ "normalized": true,
90
+ "rstrip": false,
91
+ "single_word": false,
92
+ "special": false
93
+ },
94
+ "50263": {
95
+ "content": " ",
96
+ "lstrip": false,
97
+ "normalized": true,
98
+ "rstrip": false,
99
+ "single_word": false,
100
+ "special": false
101
+ },
102
+ "50264": {
103
+ "content": " ",
104
+ "lstrip": false,
105
+ "normalized": true,
106
+ "rstrip": false,
107
+ "single_word": false,
108
+ "special": false
109
+ },
110
+ "50265": {
111
+ "content": " ",
112
+ "lstrip": false,
113
+ "normalized": true,
114
+ "rstrip": false,
115
+ "single_word": false,
116
+ "special": false
117
+ },
118
+ "50266": {
119
+ "content": " ",
120
+ "lstrip": false,
121
+ "normalized": true,
122
+ "rstrip": false,
123
+ "single_word": false,
124
+ "special": false
125
+ },
126
+ "50267": {
127
+ "content": " ",
128
+ "lstrip": false,
129
+ "normalized": true,
130
+ "rstrip": false,
131
+ "single_word": false,
132
+ "special": false
133
+ },
134
+ "50268": {
135
+ "content": " ",
136
+ "lstrip": false,
137
+ "normalized": true,
138
+ "rstrip": false,
139
+ "single_word": false,
140
+ "special": false
141
+ },
142
+ "50269": {
143
+ "content": " ",
144
+ "lstrip": false,
145
+ "normalized": true,
146
+ "rstrip": false,
147
+ "single_word": false,
148
+ "special": false
149
+ },
150
+ "50270": {
151
+ "content": " ",
152
+ "lstrip": false,
153
+ "normalized": true,
154
+ "rstrip": false,
155
+ "single_word": false,
156
+ "special": false
157
+ },
158
+ "50271": {
159
+ "content": " ",
160
+ "lstrip": false,
161
+ "normalized": true,
162
+ "rstrip": false,
163
+ "single_word": false,
164
+ "special": false
165
+ },
166
+ "50272": {
167
+ "content": " ",
168
+ "lstrip": false,
169
+ "normalized": true,
170
+ "rstrip": false,
171
+ "single_word": false,
172
+ "special": false
173
+ },
174
+ "50273": {
175
+ "content": " ",
176
+ "lstrip": false,
177
+ "normalized": true,
178
+ "rstrip": false,
179
+ "single_word": false,
180
+ "special": false
181
+ },
182
+ "50274": {
183
+ "content": " ",
184
+ "lstrip": false,
185
+ "normalized": true,
186
+ "rstrip": false,
187
+ "single_word": false,
188
+ "special": false
189
+ },
190
+ "50275": {
191
+ "content": " ",
192
+ "lstrip": false,
193
+ "normalized": true,
194
+ "rstrip": false,
195
+ "single_word": false,
196
+ "special": false
197
+ },
198
+ "50276": {
199
+ "content": " ",
200
+ "lstrip": false,
201
+ "normalized": true,
202
+ "rstrip": false,
203
+ "single_word": false,
204
+ "special": false
205
+ },
206
+ "50277": {
207
+ "content": "|||EMAIL_ADDRESS|||",
208
+ "lstrip": false,
209
+ "normalized": true,
210
+ "rstrip": false,
211
+ "single_word": false,
212
+ "special": false
213
+ },
214
+ "50278": {
215
+ "content": "|||PHONE_NUMBER|||",
216
+ "lstrip": false,
217
+ "normalized": true,
218
+ "rstrip": false,
219
+ "single_word": false,
220
+ "special": false
221
+ },
222
+ "50279": {
223
+ "content": "|||IP_ADDRESS|||",
224
+ "lstrip": false,
225
+ "normalized": true,
226
+ "rstrip": false,
227
+ "single_word": false,
228
+ "special": true
229
+ },
230
+ "50280": {
231
+ "content": "<pad>",
232
+ "lstrip": false,
233
+ "normalized": false,
234
+ "rstrip": false,
235
+ "single_word": false,
236
+ "special": true
237
+ }
238
+ },
239
+ "bos_token": "|||IP_ADDRESS|||",
240
+ "clean_up_tokenization_spaces": false,
241
+ "eos_token": "|||IP_ADDRESS|||",
242
+ "extra_special_tokens": {},
243
+ "model_max_length": 1000000000000000019884624838656,
244
+ "pad_token": "<pad>",
245
+ "tokenizer_class": "GPTNeoXTokenizer",
246
+ "unk_token": null
247
+ }
healed/liger_smoke/step0006/chat_template.jinja ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {{ bos_token }}{% for message in messages %}{% if message['role'] == 'system' %}{{ '<|system|>
2
+ ' + message['content'] + '
3
+ ' }}{% elif message['role'] == 'user' %}{{ '<|user|>
4
+ ' + message['content'] + '
5
+ ' }}{% elif message['role'] == 'assistant' %}{% if not loop.last %}{{ '<|assistant|>
6
+ ' + message['content'] + eos_token + '
7
+ ' }}{% else %}{{ '<|assistant|>
8
+ ' + message['content'] + eos_token }}{% endif %}{% endif %}{% if loop.last and add_generation_prompt %}{{ '<|assistant|>
9
+ ' }}{% endif %}{% endfor %}
healed/liger_smoke/step0006/config.json ADDED
@@ -0,0 +1,887 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "PrunedOlmoeForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_pruned_olmoe.PrunedOlmoeConfig",
9
+ "AutoModelForCausalLM": "modeling_pruned_olmoe.PrunedOlmoeForCausalLM"
10
+ },
11
+ "clip_qkv": null,
12
+ "dtype": "bfloat16",
13
+ "eos_token_id": 50279,
14
+ "expert_widths": [
15
+ [
16
+ 1024,
17
+ 384,
18
+ 256,
19
+ 256,
20
+ 768,
21
+ 1024,
22
+ 128,
23
+ 768,
24
+ 384,
25
+ 128,
26
+ 1024,
27
+ 384,
28
+ 384,
29
+ 128,
30
+ 640,
31
+ 896,
32
+ 896,
33
+ 256,
34
+ 768,
35
+ 128,
36
+ 768,
37
+ 768,
38
+ 768,
39
+ 768,
40
+ 640,
41
+ 896,
42
+ 256,
43
+ 128,
44
+ 384,
45
+ 896,
46
+ 256,
47
+ 768,
48
+ 512,
49
+ 640,
50
+ 768,
51
+ 896,
52
+ 1024,
53
+ 384,
54
+ 384,
55
+ 512,
56
+ 384,
57
+ 256,
58
+ 768,
59
+ 640,
60
+ 896,
61
+ 896,
62
+ 512,
63
+ 768,
64
+ 512,
65
+ 384,
66
+ 1024,
67
+ 896,
68
+ 896,
69
+ 768,
70
+ 128
71
+ ],
72
+ [
73
+ 768,
74
+ 896,
75
+ 896,
76
+ 256,
77
+ 896,
78
+ 384,
79
+ 768,
80
+ 768,
81
+ 1024,
82
+ 384,
83
+ 512,
84
+ 1024,
85
+ 768,
86
+ 768,
87
+ 1024,
88
+ 640,
89
+ 128,
90
+ 768,
91
+ 384,
92
+ 768,
93
+ 128,
94
+ 256,
95
+ 512,
96
+ 384,
97
+ 896,
98
+ 256,
99
+ 1024,
100
+ 1024,
101
+ 896,
102
+ 1024,
103
+ 896,
104
+ 768,
105
+ 256,
106
+ 768,
107
+ 128,
108
+ 1024,
109
+ 768,
110
+ 768,
111
+ 640,
112
+ 384,
113
+ 1024,
114
+ 384,
115
+ 640,
116
+ 896,
117
+ 1024,
118
+ 512,
119
+ 512,
120
+ 512,
121
+ 640
122
+ ],
123
+ [
124
+ 768,
125
+ 1024,
126
+ 512,
127
+ 384,
128
+ 384,
129
+ 768,
130
+ 256,
131
+ 384,
132
+ 1024,
133
+ 896,
134
+ 640,
135
+ 256,
136
+ 1024,
137
+ 1024,
138
+ 1024,
139
+ 640,
140
+ 1024,
141
+ 896,
142
+ 512,
143
+ 768,
144
+ 128,
145
+ 768,
146
+ 1024,
147
+ 768,
148
+ 256,
149
+ 1024,
150
+ 896,
151
+ 1024,
152
+ 896,
153
+ 512,
154
+ 640,
155
+ 384,
156
+ 896,
157
+ 384,
158
+ 512,
159
+ 640,
160
+ 256,
161
+ 768,
162
+ 256,
163
+ 768,
164
+ 768,
165
+ 640,
166
+ 1024,
167
+ 640,
168
+ 896,
169
+ 768,
170
+ 768,
171
+ 256
172
+ ],
173
+ [
174
+ 640,
175
+ 640,
176
+ 896,
177
+ 512,
178
+ 640,
179
+ 896,
180
+ 896,
181
+ 384,
182
+ 896,
183
+ 896,
184
+ 384,
185
+ 896,
186
+ 640,
187
+ 256,
188
+ 640,
189
+ 1024,
190
+ 1024,
191
+ 768,
192
+ 1024,
193
+ 896,
194
+ 1024,
195
+ 768,
196
+ 896,
197
+ 256,
198
+ 512,
199
+ 768,
200
+ 1024,
201
+ 256,
202
+ 768,
203
+ 512,
204
+ 256,
205
+ 640,
206
+ 1024,
207
+ 1024,
208
+ 512,
209
+ 1024,
210
+ 768,
211
+ 256,
212
+ 1024,
213
+ 384,
214
+ 896,
215
+ 1024,
216
+ 896,
217
+ 1024,
218
+ 384
219
+ ],
220
+ [
221
+ 256,
222
+ 640,
223
+ 640,
224
+ 896,
225
+ 768,
226
+ 896,
227
+ 768,
228
+ 768,
229
+ 896,
230
+ 896,
231
+ 1024,
232
+ 256,
233
+ 512,
234
+ 1024,
235
+ 640,
236
+ 896,
237
+ 512,
238
+ 512,
239
+ 384,
240
+ 384,
241
+ 256,
242
+ 384,
243
+ 384,
244
+ 896,
245
+ 896,
246
+ 768,
247
+ 640,
248
+ 896,
249
+ 768,
250
+ 1024,
251
+ 512,
252
+ 640,
253
+ 512,
254
+ 640,
255
+ 896,
256
+ 512,
257
+ 512,
258
+ 384,
259
+ 640,
260
+ 896,
261
+ 896,
262
+ 896,
263
+ 1024,
264
+ 640,
265
+ 1024,
266
+ 640,
267
+ 1024
268
+ ],
269
+ [
270
+ 1024,
271
+ 512,
272
+ 1024,
273
+ 1024,
274
+ 640,
275
+ 896,
276
+ 640,
277
+ 1024,
278
+ 896,
279
+ 384,
280
+ 1024,
281
+ 128,
282
+ 896,
283
+ 768,
284
+ 1024,
285
+ 768,
286
+ 640,
287
+ 896,
288
+ 768,
289
+ 640,
290
+ 512,
291
+ 896,
292
+ 512,
293
+ 640,
294
+ 256,
295
+ 768,
296
+ 640,
297
+ 768,
298
+ 384,
299
+ 896,
300
+ 512,
301
+ 512,
302
+ 256,
303
+ 512,
304
+ 896,
305
+ 256,
306
+ 384,
307
+ 640,
308
+ 512,
309
+ 640,
310
+ 896,
311
+ 512,
312
+ 1024,
313
+ 256,
314
+ 768,
315
+ 1024,
316
+ 768,
317
+ 256,
318
+ 256
319
+ ],
320
+ [
321
+ 640,
322
+ 896,
323
+ 1024,
324
+ 896,
325
+ 1024,
326
+ 1024,
327
+ 1024,
328
+ 512,
329
+ 256,
330
+ 256,
331
+ 1024,
332
+ 768,
333
+ 512,
334
+ 768,
335
+ 1024,
336
+ 1024,
337
+ 1024,
338
+ 384,
339
+ 512,
340
+ 1024,
341
+ 512,
342
+ 1024,
343
+ 128,
344
+ 640,
345
+ 640,
346
+ 896,
347
+ 768,
348
+ 128,
349
+ 256,
350
+ 256,
351
+ 256,
352
+ 256,
353
+ 896,
354
+ 1024,
355
+ 1024,
356
+ 384,
357
+ 896,
358
+ 256,
359
+ 896,
360
+ 640,
361
+ 1024,
362
+ 384,
363
+ 640,
364
+ 256,
365
+ 1024,
366
+ 1024,
367
+ 1024
368
+ ],
369
+ [
370
+ 1024,
371
+ 384,
372
+ 1024,
373
+ 1024,
374
+ 256,
375
+ 128,
376
+ 256,
377
+ 384,
378
+ 256,
379
+ 384,
380
+ 896,
381
+ 768,
382
+ 896,
383
+ 896,
384
+ 512,
385
+ 896,
386
+ 640,
387
+ 384,
388
+ 384,
389
+ 896,
390
+ 768,
391
+ 384,
392
+ 896,
393
+ 768,
394
+ 768,
395
+ 512,
396
+ 896,
397
+ 768,
398
+ 768,
399
+ 896,
400
+ 128,
401
+ 896,
402
+ 512,
403
+ 256,
404
+ 768,
405
+ 128,
406
+ 384,
407
+ 256,
408
+ 896,
409
+ 896,
410
+ 384,
411
+ 768,
412
+ 512,
413
+ 640,
414
+ 256,
415
+ 768,
416
+ 640,
417
+ 896,
418
+ 384,
419
+ 512,
420
+ 1024,
421
+ 768,
422
+ 384
423
+ ],
424
+ [
425
+ 512,
426
+ 768,
427
+ 512,
428
+ 256,
429
+ 128,
430
+ 640,
431
+ 384,
432
+ 640,
433
+ 768,
434
+ 896,
435
+ 640,
436
+ 768,
437
+ 256,
438
+ 384,
439
+ 1024,
440
+ 896,
441
+ 256,
442
+ 896,
443
+ 512,
444
+ 256,
445
+ 896,
446
+ 768,
447
+ 256,
448
+ 896,
449
+ 896,
450
+ 384,
451
+ 896,
452
+ 640,
453
+ 768,
454
+ 512,
455
+ 768,
456
+ 768,
457
+ 1024,
458
+ 768,
459
+ 640,
460
+ 768,
461
+ 384,
462
+ 256,
463
+ 512,
464
+ 896,
465
+ 128,
466
+ 384,
467
+ 256,
468
+ 768,
469
+ 384,
470
+ 256,
471
+ 1024,
472
+ 1024,
473
+ 896,
474
+ 256,
475
+ 1024,
476
+ 256,
477
+ 128,
478
+ 896
479
+ ],
480
+ [
481
+ 640,
482
+ 640,
483
+ 896,
484
+ 256,
485
+ 1024,
486
+ 512,
487
+ 1024,
488
+ 768,
489
+ 384,
490
+ 512,
491
+ 256,
492
+ 768,
493
+ 896,
494
+ 768,
495
+ 512,
496
+ 768,
497
+ 768,
498
+ 640,
499
+ 384,
500
+ 768,
501
+ 512,
502
+ 768,
503
+ 768,
504
+ 512,
505
+ 768,
506
+ 128,
507
+ 896,
508
+ 512,
509
+ 768,
510
+ 1024,
511
+ 128,
512
+ 384,
513
+ 768,
514
+ 768,
515
+ 768,
516
+ 384,
517
+ 512,
518
+ 640,
519
+ 768,
520
+ 512,
521
+ 768,
522
+ 1024,
523
+ 640,
524
+ 896,
525
+ 256,
526
+ 1024,
527
+ 384,
528
+ 768,
529
+ 768,
530
+ 768
531
+ ],
532
+ [
533
+ 896,
534
+ 512,
535
+ 896,
536
+ 768,
537
+ 384,
538
+ 384,
539
+ 768,
540
+ 512,
541
+ 768,
542
+ 512,
543
+ 1024,
544
+ 640,
545
+ 896,
546
+ 896,
547
+ 256,
548
+ 640,
549
+ 1024,
550
+ 256,
551
+ 896,
552
+ 128,
553
+ 128,
554
+ 128,
555
+ 768,
556
+ 896,
557
+ 384,
558
+ 896,
559
+ 512,
560
+ 896,
561
+ 384,
562
+ 256,
563
+ 640,
564
+ 640,
565
+ 896,
566
+ 768,
567
+ 640,
568
+ 256,
569
+ 896,
570
+ 896,
571
+ 512,
572
+ 128,
573
+ 896,
574
+ 256,
575
+ 256,
576
+ 640,
577
+ 896,
578
+ 896,
579
+ 128,
580
+ 1024,
581
+ 256,
582
+ 384,
583
+ 1024,
584
+ 640,
585
+ 896
586
+ ],
587
+ [
588
+ 768,
589
+ 384,
590
+ 640,
591
+ 896,
592
+ 256,
593
+ 128,
594
+ 384,
595
+ 896,
596
+ 128,
597
+ 128,
598
+ 896,
599
+ 256,
600
+ 384,
601
+ 896,
602
+ 512,
603
+ 768,
604
+ 768,
605
+ 512,
606
+ 512,
607
+ 768,
608
+ 896,
609
+ 640,
610
+ 768,
611
+ 896,
612
+ 896,
613
+ 896,
614
+ 640,
615
+ 896,
616
+ 640,
617
+ 512,
618
+ 896,
619
+ 256,
620
+ 512,
621
+ 128,
622
+ 512,
623
+ 384,
624
+ 768,
625
+ 768,
626
+ 1024,
627
+ 256,
628
+ 768,
629
+ 256,
630
+ 768,
631
+ 512,
632
+ 640,
633
+ 1024,
634
+ 128,
635
+ 896,
636
+ 896,
637
+ 896,
638
+ 896,
639
+ 1024
640
+ ],
641
+ [
642
+ 512,
643
+ 384,
644
+ 768,
645
+ 640,
646
+ 640,
647
+ 768,
648
+ 1024,
649
+ 896,
650
+ 512,
651
+ 256,
652
+ 640,
653
+ 768,
654
+ 640,
655
+ 896,
656
+ 128,
657
+ 256,
658
+ 896,
659
+ 1024,
660
+ 256,
661
+ 640,
662
+ 512,
663
+ 256,
664
+ 128,
665
+ 512,
666
+ 256,
667
+ 640,
668
+ 768,
669
+ 768,
670
+ 128,
671
+ 128,
672
+ 768,
673
+ 640,
674
+ 1024,
675
+ 1024,
676
+ 768,
677
+ 512,
678
+ 896,
679
+ 768,
680
+ 896,
681
+ 1024,
682
+ 896,
683
+ 896,
684
+ 512,
685
+ 640,
686
+ 1024,
687
+ 512,
688
+ 1024,
689
+ 512,
690
+ 512,
691
+ 512,
692
+ 768
693
+ ],
694
+ [
695
+ 896,
696
+ 896,
697
+ 1024,
698
+ 1024,
699
+ 896,
700
+ 128,
701
+ 768,
702
+ 256,
703
+ 1024,
704
+ 256,
705
+ 1024,
706
+ 640,
707
+ 384,
708
+ 256,
709
+ 256,
710
+ 512,
711
+ 768,
712
+ 896,
713
+ 512,
714
+ 768,
715
+ 384,
716
+ 1024,
717
+ 896,
718
+ 896,
719
+ 1024,
720
+ 896,
721
+ 768,
722
+ 896,
723
+ 640,
724
+ 1024,
725
+ 512,
726
+ 896,
727
+ 512,
728
+ 1024,
729
+ 512,
730
+ 512,
731
+ 256,
732
+ 256,
733
+ 256,
734
+ 512,
735
+ 768,
736
+ 128,
737
+ 384,
738
+ 512,
739
+ 896,
740
+ 896,
741
+ 1024,
742
+ 256,
743
+ 384,
744
+ 640
745
+ ],
746
+ [
747
+ 896,
748
+ 640,
749
+ 384,
750
+ 512,
751
+ 256,
752
+ 640,
753
+ 1024,
754
+ 384,
755
+ 1024,
756
+ 1024,
757
+ 768,
758
+ 256,
759
+ 1024,
760
+ 768,
761
+ 512,
762
+ 896,
763
+ 256,
764
+ 1024,
765
+ 768,
766
+ 768,
767
+ 768,
768
+ 384,
769
+ 384,
770
+ 256,
771
+ 1024,
772
+ 384,
773
+ 384,
774
+ 384,
775
+ 896,
776
+ 768,
777
+ 640,
778
+ 768,
779
+ 512,
780
+ 896,
781
+ 896,
782
+ 896,
783
+ 896,
784
+ 256,
785
+ 384,
786
+ 128,
787
+ 1024,
788
+ 896,
789
+ 256,
790
+ 256,
791
+ 768,
792
+ 640,
793
+ 896,
794
+ 384,
795
+ 768,
796
+ 512,
797
+ 640
798
+ ],
799
+ [
800
+ 896,
801
+ 1024,
802
+ 768,
803
+ 1024,
804
+ 896,
805
+ 256,
806
+ 768,
807
+ 128,
808
+ 128,
809
+ 768,
810
+ 512,
811
+ 896,
812
+ 384,
813
+ 768,
814
+ 1024,
815
+ 256,
816
+ 768,
817
+ 768,
818
+ 256,
819
+ 512,
820
+ 512,
821
+ 640,
822
+ 512,
823
+ 256,
824
+ 768,
825
+ 896,
826
+ 384,
827
+ 1024,
828
+ 640,
829
+ 1024,
830
+ 512,
831
+ 512,
832
+ 384,
833
+ 512,
834
+ 512,
835
+ 1024,
836
+ 384,
837
+ 896,
838
+ 768,
839
+ 384,
840
+ 384,
841
+ 128,
842
+ 384,
843
+ 1024,
844
+ 896,
845
+ 640,
846
+ 768,
847
+ 768,
848
+ 256,
849
+ 640,
850
+ 512,
851
+ 640,
852
+ 384
853
+ ]
854
+ ],
855
+ "glean_metadata": {
856
+ "base_model": "allenai/OLMoE-1B-7B-0125-Instruct",
857
+ "block_size": 128,
858
+ "criterion": "reap",
859
+ "dead_experts": 217,
860
+ "keep_fraction": 0.5,
861
+ "min_width": 128,
862
+ "params": 3697491968,
863
+ "scores": "outputs/scores_0125inst_dolmino-math/scores.pt"
864
+ },
865
+ "hidden_act": "silu",
866
+ "hidden_size": 2048,
867
+ "initializer_range": 0.02,
868
+ "intermediate_size": 1024,
869
+ "max_position_embeddings": 4096,
870
+ "model_type": "pruned_olmoe",
871
+ "norm_topk_prob": false,
872
+ "num_attention_heads": 16,
873
+ "num_experts": 64,
874
+ "num_experts_per_tok": 8,
875
+ "num_hidden_layers": 16,
876
+ "num_key_value_heads": 16,
877
+ "output_router_logits": false,
878
+ "pad_token_id": 1,
879
+ "rms_norm_eps": 1e-05,
880
+ "rope_scaling": null,
881
+ "rope_theta": 10000.0,
882
+ "router_aux_loss_coef": 0.01,
883
+ "tie_word_embeddings": false,
884
+ "transformers_version": "4.57.6",
885
+ "use_cache": false,
886
+ "vocab_size": 50304
887
+ }
healed/liger_smoke/step0006/configuration_pruned_olmoe.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Configuration for GLEAN-pruned OLMoE: variable-width, variable-count experts.
2
+ """
3
+
4
+ from transformers.models.olmoe.configuration_olmoe import OlmoeConfig
5
+
6
+
7
+ class PrunedOlmoeConfig(OlmoeConfig):
8
+ """OlmoeConfig plus a per-(layer, expert) width table.
9
+
10
+ ``expert_widths[l]`` lists the SwiGLU intermediate width of each surviving
11
+ expert in decoder layer ``l``, in expert order. Lists are ragged: layers
12
+ may keep different numbers of experts (deleted experts simply don't
13
+ appear — the router in layer ``l`` has ``len(expert_widths[l])`` rows),
14
+ and each width may differ (multiples of the GEMM block size, 128, for
15
+ variable-MegaBlocks execution). ``None`` means an unpruned model
16
+ (uniform ``num_experts`` × ``intermediate_size``).
17
+
18
+ The inherited ``num_experts`` / ``intermediate_size`` keep their ORIGINAL
19
+ (pre-pruning) values for provenance; the width table is authoritative for
20
+ the built architecture.
21
+ """
22
+
23
+ model_type = "pruned_olmoe"
24
+
25
+ def __init__(self, expert_widths: list[list[int]] | None = None, **kwargs):
26
+ super().__init__(**kwargs)
27
+ self.expert_widths = expert_widths
healed/liger_smoke/step0006/generation_config.json ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "eos_token_id": 50279,
4
+ "pad_token_id": 1,
5
+ "transformers_version": "4.57.6"
6
+ }
healed/liger_smoke/step0006/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
healed/liger_smoke/step0006/modeling_pruned_olmoe.py ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """GLEAN-pruned OLMoE: HF-loadable model with ragged (variable-width) experts.
2
+
3
+ Pattern follows hbfreed/variable-flex-olmo's PrunedFlexOlmoForCausalLM
4
+ (docs/recon/prior-work-hbfreed.md), generalized from one scalar width to a
5
+ per-(layer, expert) width table: ``super().__init__`` builds the uniform
6
+ architecture from the config, then every MoE block is rebuilt to its pruned
7
+ shape — surviving experts only, each at its own width, router sliced to
8
+ match — so the state dict aligns exactly with what
9
+ ``glean.prune.prune_channels_global`` leaves behind.
10
+
11
+ Caveat: ``output_router_logits=True`` (the load-balancing aux loss) assumes a
12
+ uniform ``config.num_experts`` and is unsupported on ragged models.
13
+ """
14
+
15
+ import torch.nn as nn
16
+ from transformers.activations import ACT2FN
17
+ from transformers.models.olmoe.modeling_olmoe import OlmoeForCausalLM
18
+
19
+ from .configuration_pruned_olmoe import PrunedOlmoeConfig
20
+
21
+
22
+ class RaggedOlmoeMLP(nn.Module):
23
+ """OlmoeMLP with an explicit intermediate width (SwiGLU, no biases)."""
24
+
25
+ def __init__(self, hidden_size: int, intermediate_size: int, hidden_act: str):
26
+ super().__init__()
27
+ self.hidden_size = hidden_size
28
+ self.intermediate_size = intermediate_size
29
+ self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
30
+ self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
31
+ self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
32
+ self.act_fn = ACT2FN[hidden_act]
33
+
34
+ def forward(self, x):
35
+ return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
36
+
37
+
38
+ class PrunedOlmoeForCausalLM(OlmoeForCausalLM):
39
+ """OLMoE with per-layer surviving-expert lists at per-expert widths."""
40
+
41
+ config_class = PrunedOlmoeConfig
42
+
43
+ def __init__(self, config: PrunedOlmoeConfig):
44
+ super().__init__(config)
45
+ widths_table = getattr(config, "expert_widths", None)
46
+ if widths_table is None:
47
+ return # unpruned: plain OLMoE
48
+ if len(widths_table) != len(self.model.layers):
49
+ raise ValueError(
50
+ f"expert_widths has {len(widths_table)} rows but the model has "
51
+ f"{len(self.model.layers)} decoder layers"
52
+ )
53
+ for layer, widths in zip(self.model.layers, widths_table):
54
+ if any(w <= 0 for w in widths):
55
+ raise ValueError("expert_widths must list surviving experts only (>0)")
56
+ block = layer.mlp
57
+ if len(widths) < block.top_k:
58
+ raise ValueError(
59
+ f"a layer keeps {len(widths)} experts < top_k={block.top_k}"
60
+ )
61
+ block.num_experts = len(widths)
62
+ block.gate = nn.Linear(config.hidden_size, len(widths), bias=False)
63
+ block.experts = nn.ModuleList(
64
+ RaggedOlmoeMLP(config.hidden_size, w, config.hidden_act)
65
+ for w in widths
66
+ )
healed/liger_smoke/step0006/special_tokens_map.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "|||IP_ADDRESS|||",
4
+ "lstrip": false,
5
+ "normalized": true,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "|||IP_ADDRESS|||",
11
+ "lstrip": false,
12
+ "normalized": true,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "pad_token": {
17
+ "content": "<pad>",
18
+ "lstrip": false,
19
+ "normalized": false,
20
+ "rstrip": false,
21
+ "single_word": false
22
+ }
23
+ }
healed/liger_smoke/step0006/tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
healed/liger_smoke/step0006/tokenizer_config.json ADDED
@@ -0,0 +1,247 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": false,
3
+ "add_eos_token": false,
4
+ "add_prefix_space": false,
5
+ "added_tokens_decoder": {
6
+ "0": {
7
+ "content": "<|endoftext|>",
8
+ "lstrip": false,
9
+ "normalized": false,
10
+ "rstrip": false,
11
+ "single_word": false,
12
+ "special": true
13
+ },
14
+ "1": {
15
+ "content": "<|padding|>",
16
+ "lstrip": false,
17
+ "normalized": false,
18
+ "rstrip": false,
19
+ "single_word": false,
20
+ "special": true
21
+ },
22
+ "50254": {
23
+ "content": " ",
24
+ "lstrip": false,
25
+ "normalized": true,
26
+ "rstrip": false,
27
+ "single_word": false,
28
+ "special": false
29
+ },
30
+ "50255": {
31
+ "content": " ",
32
+ "lstrip": false,
33
+ "normalized": true,
34
+ "rstrip": false,
35
+ "single_word": false,
36
+ "special": false
37
+ },
38
+ "50256": {
39
+ "content": " ",
40
+ "lstrip": false,
41
+ "normalized": true,
42
+ "rstrip": false,
43
+ "single_word": false,
44
+ "special": false
45
+ },
46
+ "50257": {
47
+ "content": " ",
48
+ "lstrip": false,
49
+ "normalized": true,
50
+ "rstrip": false,
51
+ "single_word": false,
52
+ "special": false
53
+ },
54
+ "50258": {
55
+ "content": " ",
56
+ "lstrip": false,
57
+ "normalized": true,
58
+ "rstrip": false,
59
+ "single_word": false,
60
+ "special": false
61
+ },
62
+ "50259": {
63
+ "content": " ",
64
+ "lstrip": false,
65
+ "normalized": true,
66
+ "rstrip": false,
67
+ "single_word": false,
68
+ "special": false
69
+ },
70
+ "50260": {
71
+ "content": " ",
72
+ "lstrip": false,
73
+ "normalized": true,
74
+ "rstrip": false,
75
+ "single_word": false,
76
+ "special": false
77
+ },
78
+ "50261": {
79
+ "content": " ",
80
+ "lstrip": false,
81
+ "normalized": true,
82
+ "rstrip": false,
83
+ "single_word": false,
84
+ "special": false
85
+ },
86
+ "50262": {
87
+ "content": " ",
88
+ "lstrip": false,
89
+ "normalized": true,
90
+ "rstrip": false,
91
+ "single_word": false,
92
+ "special": false
93
+ },
94
+ "50263": {
95
+ "content": " ",
96
+ "lstrip": false,
97
+ "normalized": true,
98
+ "rstrip": false,
99
+ "single_word": false,
100
+ "special": false
101
+ },
102
+ "50264": {
103
+ "content": " ",
104
+ "lstrip": false,
105
+ "normalized": true,
106
+ "rstrip": false,
107
+ "single_word": false,
108
+ "special": false
109
+ },
110
+ "50265": {
111
+ "content": " ",
112
+ "lstrip": false,
113
+ "normalized": true,
114
+ "rstrip": false,
115
+ "single_word": false,
116
+ "special": false
117
+ },
118
+ "50266": {
119
+ "content": " ",
120
+ "lstrip": false,
121
+ "normalized": true,
122
+ "rstrip": false,
123
+ "single_word": false,
124
+ "special": false
125
+ },
126
+ "50267": {
127
+ "content": " ",
128
+ "lstrip": false,
129
+ "normalized": true,
130
+ "rstrip": false,
131
+ "single_word": false,
132
+ "special": false
133
+ },
134
+ "50268": {
135
+ "content": " ",
136
+ "lstrip": false,
137
+ "normalized": true,
138
+ "rstrip": false,
139
+ "single_word": false,
140
+ "special": false
141
+ },
142
+ "50269": {
143
+ "content": " ",
144
+ "lstrip": false,
145
+ "normalized": true,
146
+ "rstrip": false,
147
+ "single_word": false,
148
+ "special": false
149
+ },
150
+ "50270": {
151
+ "content": " ",
152
+ "lstrip": false,
153
+ "normalized": true,
154
+ "rstrip": false,
155
+ "single_word": false,
156
+ "special": false
157
+ },
158
+ "50271": {
159
+ "content": " ",
160
+ "lstrip": false,
161
+ "normalized": true,
162
+ "rstrip": false,
163
+ "single_word": false,
164
+ "special": false
165
+ },
166
+ "50272": {
167
+ "content": " ",
168
+ "lstrip": false,
169
+ "normalized": true,
170
+ "rstrip": false,
171
+ "single_word": false,
172
+ "special": false
173
+ },
174
+ "50273": {
175
+ "content": " ",
176
+ "lstrip": false,
177
+ "normalized": true,
178
+ "rstrip": false,
179
+ "single_word": false,
180
+ "special": false
181
+ },
182
+ "50274": {
183
+ "content": " ",
184
+ "lstrip": false,
185
+ "normalized": true,
186
+ "rstrip": false,
187
+ "single_word": false,
188
+ "special": false
189
+ },
190
+ "50275": {
191
+ "content": " ",
192
+ "lstrip": false,
193
+ "normalized": true,
194
+ "rstrip": false,
195
+ "single_word": false,
196
+ "special": false
197
+ },
198
+ "50276": {
199
+ "content": " ",
200
+ "lstrip": false,
201
+ "normalized": true,
202
+ "rstrip": false,
203
+ "single_word": false,
204
+ "special": false
205
+ },
206
+ "50277": {
207
+ "content": "|||EMAIL_ADDRESS|||",
208
+ "lstrip": false,
209
+ "normalized": true,
210
+ "rstrip": false,
211
+ "single_word": false,
212
+ "special": false
213
+ },
214
+ "50278": {
215
+ "content": "|||PHONE_NUMBER|||",
216
+ "lstrip": false,
217
+ "normalized": true,
218
+ "rstrip": false,
219
+ "single_word": false,
220
+ "special": false
221
+ },
222
+ "50279": {
223
+ "content": "|||IP_ADDRESS|||",
224
+ "lstrip": false,
225
+ "normalized": true,
226
+ "rstrip": false,
227
+ "single_word": false,
228
+ "special": true
229
+ },
230
+ "50280": {
231
+ "content": "<pad>",
232
+ "lstrip": false,
233
+ "normalized": false,
234
+ "rstrip": false,
235
+ "single_word": false,
236
+ "special": true
237
+ }
238
+ },
239
+ "bos_token": "|||IP_ADDRESS|||",
240
+ "clean_up_tokenization_spaces": false,
241
+ "eos_token": "|||IP_ADDRESS|||",
242
+ "extra_special_tokens": {},
243
+ "model_max_length": 1000000000000000019884624838656,
244
+ "pad_token": "<pad>",
245
+ "tokenizer_class": "GPTNeoXTokenizer",
246
+ "unk_token": null
247
+ }
healed/mixonly_keep50/vllm_live/chat_template.jinja ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {{ bos_token }}{% for message in messages %}{% if message['role'] == 'system' %}{{ '<|system|>
2
+ ' + message['content'] + '
3
+ ' }}{% elif message['role'] == 'user' %}{{ '<|user|>
4
+ ' + message['content'] + '
5
+ ' }}{% elif message['role'] == 'assistant' %}{% if not loop.last %}{{ '<|assistant|>
6
+ ' + message['content'] + eos_token + '
7
+ ' }}{% else %}{{ '<|assistant|>
8
+ ' + message['content'] + eos_token }}{% endif %}{% endif %}{% if loop.last and add_generation_prompt %}{{ '<|assistant|>
9
+ ' }}{% endif %}{% endfor %}
healed/mixonly_keep50/vllm_live/config.json ADDED
@@ -0,0 +1,887 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "PrunedOlmoeForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_pruned_olmoe.PrunedOlmoeConfig",
9
+ "AutoModelForCausalLM": "modeling_pruned_olmoe.PrunedOlmoeForCausalLM"
10
+ },
11
+ "clip_qkv": null,
12
+ "dtype": "bfloat16",
13
+ "eos_token_id": 50279,
14
+ "expert_widths": [
15
+ [
16
+ 1024,
17
+ 384,
18
+ 256,
19
+ 256,
20
+ 768,
21
+ 1024,
22
+ 128,
23
+ 768,
24
+ 384,
25
+ 128,
26
+ 1024,
27
+ 384,
28
+ 384,
29
+ 128,
30
+ 640,
31
+ 896,
32
+ 896,
33
+ 256,
34
+ 768,
35
+ 128,
36
+ 768,
37
+ 768,
38
+ 768,
39
+ 768,
40
+ 640,
41
+ 896,
42
+ 256,
43
+ 128,
44
+ 384,
45
+ 896,
46
+ 256,
47
+ 768,
48
+ 512,
49
+ 640,
50
+ 768,
51
+ 896,
52
+ 1024,
53
+ 384,
54
+ 384,
55
+ 512,
56
+ 384,
57
+ 256,
58
+ 768,
59
+ 640,
60
+ 896,
61
+ 896,
62
+ 512,
63
+ 768,
64
+ 512,
65
+ 384,
66
+ 1024,
67
+ 896,
68
+ 896,
69
+ 768,
70
+ 128
71
+ ],
72
+ [
73
+ 768,
74
+ 896,
75
+ 896,
76
+ 256,
77
+ 896,
78
+ 384,
79
+ 768,
80
+ 768,
81
+ 1024,
82
+ 384,
83
+ 512,
84
+ 1024,
85
+ 768,
86
+ 768,
87
+ 1024,
88
+ 640,
89
+ 128,
90
+ 768,
91
+ 384,
92
+ 768,
93
+ 128,
94
+ 256,
95
+ 512,
96
+ 384,
97
+ 896,
98
+ 256,
99
+ 1024,
100
+ 1024,
101
+ 896,
102
+ 1024,
103
+ 896,
104
+ 768,
105
+ 256,
106
+ 768,
107
+ 128,
108
+ 1024,
109
+ 768,
110
+ 768,
111
+ 640,
112
+ 384,
113
+ 1024,
114
+ 384,
115
+ 640,
116
+ 896,
117
+ 1024,
118
+ 512,
119
+ 512,
120
+ 512,
121
+ 640
122
+ ],
123
+ [
124
+ 768,
125
+ 1024,
126
+ 512,
127
+ 384,
128
+ 384,
129
+ 768,
130
+ 256,
131
+ 384,
132
+ 1024,
133
+ 896,
134
+ 640,
135
+ 256,
136
+ 1024,
137
+ 1024,
138
+ 1024,
139
+ 640,
140
+ 1024,
141
+ 896,
142
+ 512,
143
+ 768,
144
+ 128,
145
+ 768,
146
+ 1024,
147
+ 768,
148
+ 256,
149
+ 1024,
150
+ 896,
151
+ 1024,
152
+ 896,
153
+ 512,
154
+ 640,
155
+ 384,
156
+ 896,
157
+ 384,
158
+ 512,
159
+ 640,
160
+ 256,
161
+ 768,
162
+ 256,
163
+ 768,
164
+ 768,
165
+ 640,
166
+ 1024,
167
+ 640,
168
+ 896,
169
+ 768,
170
+ 768,
171
+ 256
172
+ ],
173
+ [
174
+ 640,
175
+ 640,
176
+ 896,
177
+ 512,
178
+ 640,
179
+ 896,
180
+ 896,
181
+ 384,
182
+ 896,
183
+ 896,
184
+ 384,
185
+ 896,
186
+ 640,
187
+ 256,
188
+ 640,
189
+ 1024,
190
+ 1024,
191
+ 768,
192
+ 1024,
193
+ 896,
194
+ 1024,
195
+ 768,
196
+ 896,
197
+ 256,
198
+ 512,
199
+ 768,
200
+ 1024,
201
+ 256,
202
+ 768,
203
+ 512,
204
+ 256,
205
+ 640,
206
+ 1024,
207
+ 1024,
208
+ 512,
209
+ 1024,
210
+ 768,
211
+ 256,
212
+ 1024,
213
+ 384,
214
+ 896,
215
+ 1024,
216
+ 896,
217
+ 1024,
218
+ 384
219
+ ],
220
+ [
221
+ 256,
222
+ 640,
223
+ 640,
224
+ 896,
225
+ 768,
226
+ 896,
227
+ 768,
228
+ 768,
229
+ 896,
230
+ 896,
231
+ 1024,
232
+ 256,
233
+ 512,
234
+ 1024,
235
+ 640,
236
+ 896,
237
+ 512,
238
+ 512,
239
+ 384,
240
+ 384,
241
+ 256,
242
+ 384,
243
+ 384,
244
+ 896,
245
+ 896,
246
+ 768,
247
+ 640,
248
+ 896,
249
+ 768,
250
+ 1024,
251
+ 512,
252
+ 640,
253
+ 512,
254
+ 640,
255
+ 896,
256
+ 512,
257
+ 512,
258
+ 384,
259
+ 640,
260
+ 896,
261
+ 896,
262
+ 896,
263
+ 1024,
264
+ 640,
265
+ 1024,
266
+ 640,
267
+ 1024
268
+ ],
269
+ [
270
+ 1024,
271
+ 512,
272
+ 1024,
273
+ 1024,
274
+ 640,
275
+ 896,
276
+ 640,
277
+ 1024,
278
+ 896,
279
+ 384,
280
+ 1024,
281
+ 128,
282
+ 896,
283
+ 768,
284
+ 1024,
285
+ 768,
286
+ 640,
287
+ 896,
288
+ 768,
289
+ 640,
290
+ 512,
291
+ 896,
292
+ 512,
293
+ 640,
294
+ 256,
295
+ 768,
296
+ 640,
297
+ 768,
298
+ 384,
299
+ 896,
300
+ 512,
301
+ 512,
302
+ 256,
303
+ 512,
304
+ 896,
305
+ 256,
306
+ 384,
307
+ 640,
308
+ 512,
309
+ 640,
310
+ 896,
311
+ 512,
312
+ 1024,
313
+ 256,
314
+ 768,
315
+ 1024,
316
+ 768,
317
+ 256,
318
+ 256
319
+ ],
320
+ [
321
+ 640,
322
+ 896,
323
+ 1024,
324
+ 896,
325
+ 1024,
326
+ 1024,
327
+ 1024,
328
+ 512,
329
+ 256,
330
+ 256,
331
+ 1024,
332
+ 768,
333
+ 512,
334
+ 768,
335
+ 1024,
336
+ 1024,
337
+ 1024,
338
+ 384,
339
+ 512,
340
+ 1024,
341
+ 512,
342
+ 1024,
343
+ 128,
344
+ 640,
345
+ 640,
346
+ 896,
347
+ 768,
348
+ 128,
349
+ 256,
350
+ 256,
351
+ 256,
352
+ 256,
353
+ 896,
354
+ 1024,
355
+ 1024,
356
+ 384,
357
+ 896,
358
+ 256,
359
+ 896,
360
+ 640,
361
+ 1024,
362
+ 384,
363
+ 640,
364
+ 256,
365
+ 1024,
366
+ 1024,
367
+ 1024
368
+ ],
369
+ [
370
+ 1024,
371
+ 384,
372
+ 1024,
373
+ 1024,
374
+ 256,
375
+ 128,
376
+ 256,
377
+ 384,
378
+ 256,
379
+ 384,
380
+ 896,
381
+ 768,
382
+ 896,
383
+ 896,
384
+ 512,
385
+ 896,
386
+ 640,
387
+ 384,
388
+ 384,
389
+ 896,
390
+ 768,
391
+ 384,
392
+ 896,
393
+ 768,
394
+ 768,
395
+ 512,
396
+ 896,
397
+ 768,
398
+ 768,
399
+ 896,
400
+ 128,
401
+ 896,
402
+ 512,
403
+ 256,
404
+ 768,
405
+ 128,
406
+ 384,
407
+ 256,
408
+ 896,
409
+ 896,
410
+ 384,
411
+ 768,
412
+ 512,
413
+ 640,
414
+ 256,
415
+ 768,
416
+ 640,
417
+ 896,
418
+ 384,
419
+ 512,
420
+ 1024,
421
+ 768,
422
+ 384
423
+ ],
424
+ [
425
+ 512,
426
+ 768,
427
+ 512,
428
+ 256,
429
+ 128,
430
+ 640,
431
+ 384,
432
+ 640,
433
+ 768,
434
+ 896,
435
+ 640,
436
+ 768,
437
+ 256,
438
+ 384,
439
+ 1024,
440
+ 896,
441
+ 256,
442
+ 896,
443
+ 512,
444
+ 256,
445
+ 896,
446
+ 768,
447
+ 256,
448
+ 896,
449
+ 896,
450
+ 384,
451
+ 896,
452
+ 640,
453
+ 768,
454
+ 512,
455
+ 768,
456
+ 768,
457
+ 1024,
458
+ 768,
459
+ 640,
460
+ 768,
461
+ 384,
462
+ 256,
463
+ 512,
464
+ 896,
465
+ 128,
466
+ 384,
467
+ 256,
468
+ 768,
469
+ 384,
470
+ 256,
471
+ 1024,
472
+ 1024,
473
+ 896,
474
+ 256,
475
+ 1024,
476
+ 256,
477
+ 128,
478
+ 896
479
+ ],
480
+ [
481
+ 640,
482
+ 640,
483
+ 896,
484
+ 256,
485
+ 1024,
486
+ 512,
487
+ 1024,
488
+ 768,
489
+ 384,
490
+ 512,
491
+ 256,
492
+ 768,
493
+ 896,
494
+ 768,
495
+ 512,
496
+ 768,
497
+ 768,
498
+ 640,
499
+ 384,
500
+ 768,
501
+ 512,
502
+ 768,
503
+ 768,
504
+ 512,
505
+ 768,
506
+ 128,
507
+ 896,
508
+ 512,
509
+ 768,
510
+ 1024,
511
+ 128,
512
+ 384,
513
+ 768,
514
+ 768,
515
+ 768,
516
+ 384,
517
+ 512,
518
+ 640,
519
+ 768,
520
+ 512,
521
+ 768,
522
+ 1024,
523
+ 640,
524
+ 896,
525
+ 256,
526
+ 1024,
527
+ 384,
528
+ 768,
529
+ 768,
530
+ 768
531
+ ],
532
+ [
533
+ 896,
534
+ 512,
535
+ 896,
536
+ 768,
537
+ 384,
538
+ 384,
539
+ 768,
540
+ 512,
541
+ 768,
542
+ 512,
543
+ 1024,
544
+ 640,
545
+ 896,
546
+ 896,
547
+ 256,
548
+ 640,
549
+ 1024,
550
+ 256,
551
+ 896,
552
+ 128,
553
+ 128,
554
+ 128,
555
+ 768,
556
+ 896,
557
+ 384,
558
+ 896,
559
+ 512,
560
+ 896,
561
+ 384,
562
+ 256,
563
+ 640,
564
+ 640,
565
+ 896,
566
+ 768,
567
+ 640,
568
+ 256,
569
+ 896,
570
+ 896,
571
+ 512,
572
+ 128,
573
+ 896,
574
+ 256,
575
+ 256,
576
+ 640,
577
+ 896,
578
+ 896,
579
+ 128,
580
+ 1024,
581
+ 256,
582
+ 384,
583
+ 1024,
584
+ 640,
585
+ 896
586
+ ],
587
+ [
588
+ 768,
589
+ 384,
590
+ 640,
591
+ 896,
592
+ 256,
593
+ 128,
594
+ 384,
595
+ 896,
596
+ 128,
597
+ 128,
598
+ 896,
599
+ 256,
600
+ 384,
601
+ 896,
602
+ 512,
603
+ 768,
604
+ 768,
605
+ 512,
606
+ 512,
607
+ 768,
608
+ 896,
609
+ 640,
610
+ 768,
611
+ 896,
612
+ 896,
613
+ 896,
614
+ 640,
615
+ 896,
616
+ 640,
617
+ 512,
618
+ 896,
619
+ 256,
620
+ 512,
621
+ 128,
622
+ 512,
623
+ 384,
624
+ 768,
625
+ 768,
626
+ 1024,
627
+ 256,
628
+ 768,
629
+ 256,
630
+ 768,
631
+ 512,
632
+ 640,
633
+ 1024,
634
+ 128,
635
+ 896,
636
+ 896,
637
+ 896,
638
+ 896,
639
+ 1024
640
+ ],
641
+ [
642
+ 512,
643
+ 384,
644
+ 768,
645
+ 640,
646
+ 640,
647
+ 768,
648
+ 1024,
649
+ 896,
650
+ 512,
651
+ 256,
652
+ 640,
653
+ 768,
654
+ 640,
655
+ 896,
656
+ 128,
657
+ 256,
658
+ 896,
659
+ 1024,
660
+ 256,
661
+ 640,
662
+ 512,
663
+ 256,
664
+ 128,
665
+ 512,
666
+ 256,
667
+ 640,
668
+ 768,
669
+ 768,
670
+ 128,
671
+ 128,
672
+ 768,
673
+ 640,
674
+ 1024,
675
+ 1024,
676
+ 768,
677
+ 512,
678
+ 896,
679
+ 768,
680
+ 896,
681
+ 1024,
682
+ 896,
683
+ 896,
684
+ 512,
685
+ 640,
686
+ 1024,
687
+ 512,
688
+ 1024,
689
+ 512,
690
+ 512,
691
+ 512,
692
+ 768
693
+ ],
694
+ [
695
+ 896,
696
+ 896,
697
+ 1024,
698
+ 1024,
699
+ 896,
700
+ 128,
701
+ 768,
702
+ 256,
703
+ 1024,
704
+ 256,
705
+ 1024,
706
+ 640,
707
+ 384,
708
+ 256,
709
+ 256,
710
+ 512,
711
+ 768,
712
+ 896,
713
+ 512,
714
+ 768,
715
+ 384,
716
+ 1024,
717
+ 896,
718
+ 896,
719
+ 1024,
720
+ 896,
721
+ 768,
722
+ 896,
723
+ 640,
724
+ 1024,
725
+ 512,
726
+ 896,
727
+ 512,
728
+ 1024,
729
+ 512,
730
+ 512,
731
+ 256,
732
+ 256,
733
+ 256,
734
+ 512,
735
+ 768,
736
+ 128,
737
+ 384,
738
+ 512,
739
+ 896,
740
+ 896,
741
+ 1024,
742
+ 256,
743
+ 384,
744
+ 640
745
+ ],
746
+ [
747
+ 896,
748
+ 640,
749
+ 384,
750
+ 512,
751
+ 256,
752
+ 640,
753
+ 1024,
754
+ 384,
755
+ 1024,
756
+ 1024,
757
+ 768,
758
+ 256,
759
+ 1024,
760
+ 768,
761
+ 512,
762
+ 896,
763
+ 256,
764
+ 1024,
765
+ 768,
766
+ 768,
767
+ 768,
768
+ 384,
769
+ 384,
770
+ 256,
771
+ 1024,
772
+ 384,
773
+ 384,
774
+ 384,
775
+ 896,
776
+ 768,
777
+ 640,
778
+ 768,
779
+ 512,
780
+ 896,
781
+ 896,
782
+ 896,
783
+ 896,
784
+ 256,
785
+ 384,
786
+ 128,
787
+ 1024,
788
+ 896,
789
+ 256,
790
+ 256,
791
+ 768,
792
+ 640,
793
+ 896,
794
+ 384,
795
+ 768,
796
+ 512,
797
+ 640
798
+ ],
799
+ [
800
+ 896,
801
+ 1024,
802
+ 768,
803
+ 1024,
804
+ 896,
805
+ 256,
806
+ 768,
807
+ 128,
808
+ 128,
809
+ 768,
810
+ 512,
811
+ 896,
812
+ 384,
813
+ 768,
814
+ 1024,
815
+ 256,
816
+ 768,
817
+ 768,
818
+ 256,
819
+ 512,
820
+ 512,
821
+ 640,
822
+ 512,
823
+ 256,
824
+ 768,
825
+ 896,
826
+ 384,
827
+ 1024,
828
+ 640,
829
+ 1024,
830
+ 512,
831
+ 512,
832
+ 384,
833
+ 512,
834
+ 512,
835
+ 1024,
836
+ 384,
837
+ 896,
838
+ 768,
839
+ 384,
840
+ 384,
841
+ 128,
842
+ 384,
843
+ 1024,
844
+ 896,
845
+ 640,
846
+ 768,
847
+ 768,
848
+ 256,
849
+ 640,
850
+ 512,
851
+ 640,
852
+ 384
853
+ ]
854
+ ],
855
+ "glean_metadata": {
856
+ "base_model": "allenai/OLMoE-1B-7B-0125-Instruct",
857
+ "block_size": 128,
858
+ "criterion": "reap",
859
+ "dead_experts": 217,
860
+ "keep_fraction": 0.5,
861
+ "min_width": 128,
862
+ "params": 3697491968,
863
+ "scores": "outputs/scores_0125inst_dolmino-math/scores.pt"
864
+ },
865
+ "hidden_act": "silu",
866
+ "hidden_size": 2048,
867
+ "initializer_range": 0.02,
868
+ "intermediate_size": 1024,
869
+ "max_position_embeddings": 4096,
870
+ "model_type": "pruned_olmoe",
871
+ "norm_topk_prob": false,
872
+ "num_attention_heads": 16,
873
+ "num_experts": 64,
874
+ "num_experts_per_tok": 8,
875
+ "num_hidden_layers": 16,
876
+ "num_key_value_heads": 16,
877
+ "output_router_logits": false,
878
+ "pad_token_id": 1,
879
+ "rms_norm_eps": 1e-05,
880
+ "rope_scaling": null,
881
+ "rope_theta": 10000.0,
882
+ "router_aux_loss_coef": 0.01,
883
+ "tie_word_embeddings": false,
884
+ "transformers_version": "4.57.6",
885
+ "use_cache": false,
886
+ "vocab_size": 50304
887
+ }
healed/mixonly_keep50/vllm_live/configuration_pruned_olmoe.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Configuration for GLEAN-pruned OLMoE: variable-width, variable-count experts.
2
+ """
3
+
4
+ from transformers.models.olmoe.configuration_olmoe import OlmoeConfig
5
+
6
+
7
+ class PrunedOlmoeConfig(OlmoeConfig):
8
+ """OlmoeConfig plus a per-(layer, expert) width table.
9
+
10
+ ``expert_widths[l]`` lists the SwiGLU intermediate width of each surviving
11
+ expert in decoder layer ``l``, in expert order. Lists are ragged: layers
12
+ may keep different numbers of experts (deleted experts simply don't
13
+ appear — the router in layer ``l`` has ``len(expert_widths[l])`` rows),
14
+ and each width may differ (multiples of the GEMM block size, 128, for
15
+ variable-MegaBlocks execution). ``None`` means an unpruned model
16
+ (uniform ``num_experts`` × ``intermediate_size``).
17
+
18
+ The inherited ``num_experts`` / ``intermediate_size`` keep their ORIGINAL
19
+ (pre-pruning) values for provenance; the width table is authoritative for
20
+ the built architecture.
21
+ """
22
+
23
+ model_type = "pruned_olmoe"
24
+
25
+ def __init__(self, expert_widths: list[list[int]] | None = None, **kwargs):
26
+ super().__init__(**kwargs)
27
+ self.expert_widths = expert_widths
healed/mixonly_keep50/vllm_live/generation_config.json ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "eos_token_id": 50279,
4
+ "pad_token_id": 1,
5
+ "transformers_version": "4.57.6"
6
+ }
healed/mixonly_keep50/vllm_live/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
healed/mixonly_keep50/vllm_live/modeling_pruned_olmoe.py ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """GLEAN-pruned OLMoE: HF-loadable model with ragged (variable-width) experts.
2
+
3
+ Pattern follows hbfreed/variable-flex-olmo's PrunedFlexOlmoForCausalLM
4
+ (docs/recon/prior-work-hbfreed.md), generalized from one scalar width to a
5
+ per-(layer, expert) width table: ``super().__init__`` builds the uniform
6
+ architecture from the config, then every MoE block is rebuilt to its pruned
7
+ shape — surviving experts only, each at its own width, router sliced to
8
+ match — so the state dict aligns exactly with what
9
+ ``glean.prune.prune_channels_global`` leaves behind.
10
+
11
+ Caveat: ``output_router_logits=True`` (the load-balancing aux loss) assumes a
12
+ uniform ``config.num_experts`` and is unsupported on ragged models.
13
+ """
14
+
15
+ import torch.nn as nn
16
+ from transformers.activations import ACT2FN
17
+ from transformers.models.olmoe.modeling_olmoe import OlmoeForCausalLM
18
+
19
+ from .configuration_pruned_olmoe import PrunedOlmoeConfig
20
+
21
+
22
+ class RaggedOlmoeMLP(nn.Module):
23
+ """OlmoeMLP with an explicit intermediate width (SwiGLU, no biases)."""
24
+
25
+ def __init__(self, hidden_size: int, intermediate_size: int, hidden_act: str):
26
+ super().__init__()
27
+ self.hidden_size = hidden_size
28
+ self.intermediate_size = intermediate_size
29
+ self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
30
+ self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
31
+ self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
32
+ self.act_fn = ACT2FN[hidden_act]
33
+
34
+ def forward(self, x):
35
+ return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
36
+
37
+
38
+ class PrunedOlmoeForCausalLM(OlmoeForCausalLM):
39
+ """OLMoE with per-layer surviving-expert lists at per-expert widths."""
40
+
41
+ config_class = PrunedOlmoeConfig
42
+
43
+ def __init__(self, config: PrunedOlmoeConfig):
44
+ super().__init__(config)
45
+ widths_table = getattr(config, "expert_widths", None)
46
+ if widths_table is None:
47
+ return # unpruned: plain OLMoE
48
+ if len(widths_table) != len(self.model.layers):
49
+ raise ValueError(
50
+ f"expert_widths has {len(widths_table)} rows but the model has "
51
+ f"{len(self.model.layers)} decoder layers"
52
+ )
53
+ for layer, widths in zip(self.model.layers, widths_table):
54
+ if any(w <= 0 for w in widths):
55
+ raise ValueError("expert_widths must list surviving experts only (>0)")
56
+ block = layer.mlp
57
+ if len(widths) < block.top_k:
58
+ raise ValueError(
59
+ f"a layer keeps {len(widths)} experts < top_k={block.top_k}"
60
+ )
61
+ block.num_experts = len(widths)
62
+ block.gate = nn.Linear(config.hidden_size, len(widths), bias=False)
63
+ block.experts = nn.ModuleList(
64
+ RaggedOlmoeMLP(config.hidden_size, w, config.hidden_act)
65
+ for w in widths
66
+ )
healed/mixonly_keep50/vllm_live/special_tokens_map.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "|||IP_ADDRESS|||",
4
+ "lstrip": false,
5
+ "normalized": true,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "|||IP_ADDRESS|||",
11
+ "lstrip": false,
12
+ "normalized": true,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "pad_token": {
17
+ "content": "<pad>",
18
+ "lstrip": false,
19
+ "normalized": false,
20
+ "rstrip": false,
21
+ "single_word": false
22
+ }
23
+ }
healed/mixonly_keep50/vllm_live/tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
healed/mixonly_keep50/vllm_live/tokenizer_config.json ADDED
@@ -0,0 +1,247 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": false,
3
+ "add_eos_token": false,
4
+ "add_prefix_space": false,
5
+ "added_tokens_decoder": {
6
+ "0": {
7
+ "content": "<|endoftext|>",
8
+ "lstrip": false,
9
+ "normalized": false,
10
+ "rstrip": false,
11
+ "single_word": false,
12
+ "special": true
13
+ },
14
+ "1": {
15
+ "content": "<|padding|>",
16
+ "lstrip": false,
17
+ "normalized": false,
18
+ "rstrip": false,
19
+ "single_word": false,
20
+ "special": true
21
+ },
22
+ "50254": {
23
+ "content": " ",
24
+ "lstrip": false,
25
+ "normalized": true,
26
+ "rstrip": false,
27
+ "single_word": false,
28
+ "special": false
29
+ },
30
+ "50255": {
31
+ "content": " ",
32
+ "lstrip": false,
33
+ "normalized": true,
34
+ "rstrip": false,
35
+ "single_word": false,
36
+ "special": false
37
+ },
38
+ "50256": {
39
+ "content": " ",
40
+ "lstrip": false,
41
+ "normalized": true,
42
+ "rstrip": false,
43
+ "single_word": false,
44
+ "special": false
45
+ },
46
+ "50257": {
47
+ "content": " ",
48
+ "lstrip": false,
49
+ "normalized": true,
50
+ "rstrip": false,
51
+ "single_word": false,
52
+ "special": false
53
+ },
54
+ "50258": {
55
+ "content": " ",
56
+ "lstrip": false,
57
+ "normalized": true,
58
+ "rstrip": false,
59
+ "single_word": false,
60
+ "special": false
61
+ },
62
+ "50259": {
63
+ "content": " ",
64
+ "lstrip": false,
65
+ "normalized": true,
66
+ "rstrip": false,
67
+ "single_word": false,
68
+ "special": false
69
+ },
70
+ "50260": {
71
+ "content": " ",
72
+ "lstrip": false,
73
+ "normalized": true,
74
+ "rstrip": false,
75
+ "single_word": false,
76
+ "special": false
77
+ },
78
+ "50261": {
79
+ "content": " ",
80
+ "lstrip": false,
81
+ "normalized": true,
82
+ "rstrip": false,
83
+ "single_word": false,
84
+ "special": false
85
+ },
86
+ "50262": {
87
+ "content": " ",
88
+ "lstrip": false,
89
+ "normalized": true,
90
+ "rstrip": false,
91
+ "single_word": false,
92
+ "special": false
93
+ },
94
+ "50263": {
95
+ "content": " ",
96
+ "lstrip": false,
97
+ "normalized": true,
98
+ "rstrip": false,
99
+ "single_word": false,
100
+ "special": false
101
+ },
102
+ "50264": {
103
+ "content": " ",
104
+ "lstrip": false,
105
+ "normalized": true,
106
+ "rstrip": false,
107
+ "single_word": false,
108
+ "special": false
109
+ },
110
+ "50265": {
111
+ "content": " ",
112
+ "lstrip": false,
113
+ "normalized": true,
114
+ "rstrip": false,
115
+ "single_word": false,
116
+ "special": false
117
+ },
118
+ "50266": {
119
+ "content": " ",
120
+ "lstrip": false,
121
+ "normalized": true,
122
+ "rstrip": false,
123
+ "single_word": false,
124
+ "special": false
125
+ },
126
+ "50267": {
127
+ "content": " ",
128
+ "lstrip": false,
129
+ "normalized": true,
130
+ "rstrip": false,
131
+ "single_word": false,
132
+ "special": false
133
+ },
134
+ "50268": {
135
+ "content": " ",
136
+ "lstrip": false,
137
+ "normalized": true,
138
+ "rstrip": false,
139
+ "single_word": false,
140
+ "special": false
141
+ },
142
+ "50269": {
143
+ "content": " ",
144
+ "lstrip": false,
145
+ "normalized": true,
146
+ "rstrip": false,
147
+ "single_word": false,
148
+ "special": false
149
+ },
150
+ "50270": {
151
+ "content": " ",
152
+ "lstrip": false,
153
+ "normalized": true,
154
+ "rstrip": false,
155
+ "single_word": false,
156
+ "special": false
157
+ },
158
+ "50271": {
159
+ "content": " ",
160
+ "lstrip": false,
161
+ "normalized": true,
162
+ "rstrip": false,
163
+ "single_word": false,
164
+ "special": false
165
+ },
166
+ "50272": {
167
+ "content": " ",
168
+ "lstrip": false,
169
+ "normalized": true,
170
+ "rstrip": false,
171
+ "single_word": false,
172
+ "special": false
173
+ },
174
+ "50273": {
175
+ "content": " ",
176
+ "lstrip": false,
177
+ "normalized": true,
178
+ "rstrip": false,
179
+ "single_word": false,
180
+ "special": false
181
+ },
182
+ "50274": {
183
+ "content": " ",
184
+ "lstrip": false,
185
+ "normalized": true,
186
+ "rstrip": false,
187
+ "single_word": false,
188
+ "special": false
189
+ },
190
+ "50275": {
191
+ "content": " ",
192
+ "lstrip": false,
193
+ "normalized": true,
194
+ "rstrip": false,
195
+ "single_word": false,
196
+ "special": false
197
+ },
198
+ "50276": {
199
+ "content": " ",
200
+ "lstrip": false,
201
+ "normalized": true,
202
+ "rstrip": false,
203
+ "single_word": false,
204
+ "special": false
205
+ },
206
+ "50277": {
207
+ "content": "|||EMAIL_ADDRESS|||",
208
+ "lstrip": false,
209
+ "normalized": true,
210
+ "rstrip": false,
211
+ "single_word": false,
212
+ "special": false
213
+ },
214
+ "50278": {
215
+ "content": "|||PHONE_NUMBER|||",
216
+ "lstrip": false,
217
+ "normalized": true,
218
+ "rstrip": false,
219
+ "single_word": false,
220
+ "special": false
221
+ },
222
+ "50279": {
223
+ "content": "|||IP_ADDRESS|||",
224
+ "lstrip": false,
225
+ "normalized": true,
226
+ "rstrip": false,
227
+ "single_word": false,
228
+ "special": true
229
+ },
230
+ "50280": {
231
+ "content": "<pad>",
232
+ "lstrip": false,
233
+ "normalized": false,
234
+ "rstrip": false,
235
+ "single_word": false,
236
+ "special": true
237
+ }
238
+ },
239
+ "bos_token": "|||IP_ADDRESS|||",
240
+ "clean_up_tokenization_spaces": false,
241
+ "eos_token": "|||IP_ADDRESS|||",
242
+ "extra_special_tokens": {},
243
+ "model_max_length": 1000000000000000019884624838656,
244
+ "pad_token": "<pad>",
245
+ "tokenizer_class": "GPTNeoXTokenizer",
246
+ "unk_token": null
247
+ }
healed/mixonly_keep50/wandb/debug-internal.log ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"time":"2026-07-31T01:15:14.143006306-07:00","level":"INFO","msg":"wandb-core"}
2
+ {"time":"2026-07-31T01:15:14.143036247-07:00","level":"INFO","msg":"stream: starting","core version":"0.28.0"}
3
+ {"time":"2026-07-31T01:15:14.255553325-07:00","level":"WARN","msg":"featurechecker: GraphQL client is nil, skipping feature loading"}
4
+ {"time":"2026-07-31T01:15:14.255573436-07:00","level":"WARN","msg":"featurechecker: GraphQL client is nil, skipping feature loading"}
5
+ {"time":"2026-07-31T01:15:14.255616137-07:00","level":"INFO","msg":"stream: created new stream","id":"gexxktae"}
6
+ {"time":"2026-07-31T01:15:14.255765612-07:00","level":"INFO","msg":"stream: started"}
7
+ {"time":"2026-07-31T01:15:14.255812423-07:00","level":"INFO","msg":"writer: started","stream_id":"gexxktae"}
8
+ {"time":"2026-07-31T01:15:14.255852534-07:00","level":"INFO","msg":"handler: started"}
9
+ {"time":"2026-07-31T01:15:14.255895105-07:00","level":"INFO","msg":"sender: started"}
10
+ {"time":"2026-07-31T01:15:14.271152628-07:00","level":"WARN","msg":"featurechecker: GraphQL client is nil, skipping feature loading"}
11
+ {"time":"2026-07-31T01:15:14.271190299-07:00","level":"WARN","msg":"runupserter: server does not expand metric globs but the x_server_side_expand_glob_metrics setting is set; ignoring"}
healed/mixonly_keep50/wandb/debug.log ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
2
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_setup.py:_flush():81] Configure stats pid to 2043723
3
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_setup.py:_flush():81] Loading settings from environment variables
4
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_init.py:setup_run_log_directory():725] Logging user logs to outputs/healed/mixonly_keep50/wandb/offline-run-20260731_011513-gexxktae/logs/debug.log
5
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to outputs/healed/mixonly_keep50/wandb/offline-run-20260731_011513-gexxktae/logs/debug-internal.log
6
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_init.py:init():768] calling init triggers
7
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
8
+ config: {'student': 'outputs/pruned/glean-0125inst-math-keep50', 'teacher': 'allenai/OLMoE-1B-7B-0125-Instruct', 'training_mode': 'on-policy', 'kl_direction': 'reverse', 'dataset': 'allenai/Dolci-Instruct-RL', 'dataset_sources': None, 'max_difficulty': None, 'trajectories': 'outputs/teacher_trajectories/dolci_math_curated.jsonl', 'trajectory_dataset': 'allenai/Dolci-Instruct-RL', 'off_policy_frames': 'chat', 'off_policy_max_seq_len': 2048, 'topk_targets': None, 'max_loss_tokens': None, 'loss_tokens_per_step': None, 'teacher_device': 'cuda:0', 'student_device': 'cuda:1', 'lr': 3e-05, 'optimizer': 'adamw8bit', 'weight_decay': 0.1, 'epochs': 2, 'prompts_per_step': 256, 'group_size': 4, 'rollout_batch': 64, 'micro_batch': 4, 'max_new_tokens': 2048, 'max_prompt_len': 1024, 'warmup_steps': 10, 'max_grad_norm': 1.0, 'eval_every': 10, 'gsm8k_every': 20, 'gsm8k_n': 256, 'gsm8k_batch': 16, 'gsm8k_max_new_tokens': 1024, 'gsm8k_frames': 'chat', 'save_every': 1000, 'out_dir': 'outputs/healed/mixonly_keep50', 'sweep': 60, 'wandb': True, 'wandb_project': 'glean-heal', 'wandb_run_name': 'mixonly_keep50-s1223', 'wandb_run_id': None, 'wandb_resume': None, 'wandb_mode': 'offline', 'no_wandb_sync': False, 'debug': False, 'resume_from': None, 'start_step': 0, 'no_grad_checkpointing': False, 'seed': 1223, 'no_teacher_overlap': False, 'sync_checkpoints': False, 'rollout_engine': 'vllm', 'vllm_gpu': '2', 'vllm_port': 8377, 'vllm_refresh_every': 1, 'vllm_serve_bin': 'vllm-plugin/.venv/bin/python', 'vllm_gpu_mem_util': 0.85, 'liger_loss': True, 'gold_mix_lambda': 0.5, 'gold_topk_targets': 'outputs/teacher_trajectories/dolci_combined_top128', 'gold_mix_decay': 0.0, 'fast_teacher': True, 'reference_kl_beta': 0.0, 'drop_truncated_rollouts': False, 'vllm_max_model_len': None, 'vllm_refresh_mode': 'reload', 'vllm_live_dir': None, 'resolved_kl_direction': 'reverse', '_wandb': {}}
9
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_init.py:init():816] starting backend
10
+ 2026-07-31 01:15:14,134 INFO MainThread:2043723 [wandb_init.py:init():831] sending inform_init request
11
+ 2026-07-31 01:15:14,256 INFO MainThread:2043723 [wandb_init.py:init():836] backend started and connected
12
+ 2026-07-31 01:15:14,258 INFO MainThread:2043723 [wandb_init.py:init():906] updated telemetry
13
+ 2026-07-31 01:15:14,265 INFO MainThread:2043723 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
14
+ 2026-07-31 01:15:14,273 INFO MainThread:2043723 [wandb_init.py:init():974] starting run threads in backend
15
+ 2026-07-31 01:15:14,385 INFO MainThread:2043723 [wandb_run.py:_console_start():2523] atexit reg
16
+ 2026-07-31 01:15:14,386 INFO MainThread:2043723 [wandb_run.py:_redirect():2373] redirect: wrap_raw
17
+ 2026-07-31 01:15:14,386 INFO MainThread:2043723 [wandb_run.py:_redirect():2442] Wrapping output streams.
18
+ 2026-07-31 01:15:14,386 INFO MainThread:2043723 [wandb_run.py:_redirect():2465] Redirects installed.
19
+ 2026-07-31 01:15:14,387 INFO MainThread:2043723 [wandb_init.py:init():1012] run started, returning control to user process
healed/mixonly_keep50/wandb/offline-run-20260731_011513-gexxktae/files/requirements.txt ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ certifi==2026.6.17
2
+ fonttools==4.63.0
3
+ urllib3==2.7.0
4
+ requests==2.34.2
5
+ nvidia-cublas-cu12==12.6.4.1
6
+ packaging==26.2
7
+ regex==2026.6.28
8
+ portalocker==3.2.0
9
+ safetensors==0.8.0
10
+ datasets==5.0.0
11
+ sqlitedict==2.1.0
12
+ narwhals==2.23.0
13
+ latex2sympy2_extended==1.11.0
14
+ smmap==5.0.3
15
+ colorama==0.4.6
16
+ anyio==4.14.1
17
+ cycler==0.12.1
18
+ pytablewriter==1.2.1
19
+ pydantic_core==2.46.4
20
+ huggingface_hub==0.36.2
21
+ charset-normalizer==3.4.8
22
+ nvidia-cusolver-cu12==11.7.1.2
23
+ joblib==1.5.3
24
+ typing-inspection==0.4.2
25
+ wandb==0.28.0
26
+ bitsandbytes==0.49.2
27
+ kiwisolver==1.5.0
28
+ cuda-toolkit==12.6.3
29
+ propcache==0.5.2
30
+ Jinja2==3.1.6
31
+ cuda-bindings==12.9.7
32
+ pluggy==1.6.0
33
+ pytest==9.1.1
34
+ fsspec==2026.4.0
35
+ tqdm==4.68.3
36
+ psutil==7.2.2
37
+ pydantic==2.13.4
38
+ typepy==1.3.5
39
+ xxhash==3.8.1
40
+ DataProperty==1.1.1
41
+ aiosignal==1.4.0
42
+ annotated-types==0.7.0
43
+ Pygments==2.20.0
44
+ zstandard==0.25.0
45
+ threadpoolctl==3.6.0
46
+ liger_kernel==0.8.1
47
+ nvidia-cudnn-cu12==9.10.2.21
48
+ evaluate==0.4.6
49
+ pillow==12.3.0
50
+ GitPython==3.1.50
51
+ httpx==0.28.1
52
+ lm_eval==0.4.12
53
+ pytz==2026.2
54
+ mbstrdecoder==1.1.5
55
+ immutabledict==4.3.1
56
+ dill==0.4.1
57
+ mpmath==1.3.0
58
+ yarl==1.24.2
59
+ word2number==1.1
60
+ networkx==3.6.1
61
+ einops==0.8.2
62
+ tabulate==0.10.0
63
+ pathvalidate==3.3.1
64
+ tenacity==9.1.4
65
+ sentry-sdk==2.64.0
66
+ nvidia-nvshmem-cu12==3.4.5
67
+ hf-xet==1.5.1
68
+ six==1.17.0
69
+ pyarrow==24.0.0
70
+ nvidia-curand-cu12==10.3.7.77
71
+ cuda-pathfinder==1.5.6
72
+ sacrebleu==2.6.0
73
+ triton==3.7.1
74
+ frozenlist==1.8.0
75
+ antlr4-python3-runtime==4.11.0
76
+ tabledata==1.3.5
77
+ h11==0.16.0
78
+ nvidia-cusparselt-cu12==0.7.1
79
+ nvidia-nvjitlink-cu12==12.6.85
80
+ scipy==1.18.0
81
+ rouge_score==0.1.2
82
+ transformers==4.57.6
83
+ multiprocess==0.70.19
84
+ absl-py==2.5.0
85
+ accelerate==1.14.0
86
+ click==8.4.2
87
+ lxml==6.1.1
88
+ defusedxml==0.7.1
89
+ setuptools==81.0.0
90
+ nvidia-cuda-cupti-cu12==12.6.80
91
+ httpcore==1.0.9
92
+ langdetect==1.0.9
93
+ nltk==3.10.0
94
+ iniconfig==2.3.0
95
+ nvidia-cuda-runtime-cu12==12.6.77
96
+ nvidia-cusparse-cu12==12.5.4.2
97
+ matplotlib==3.11.0
98
+ protobuf==7.35.1
99
+ nvidia-cufft-cu12==11.3.0.4
100
+ pyparsing==3.3.2
101
+ chardet==6.0.0.post1
102
+ typing_extensions==4.16.0
103
+ stanford-stk==0.7.1
104
+ aiohappyeyeballs==2.7.1
105
+ MarkupSafe==3.0.3
106
+ nvidia-cufile-cu12==1.11.1.6
107
+ tcolorpy==0.1.7
108
+ scikit-learn==1.9.0
109
+ tokenizers==0.22.2
110
+ attrs==26.1.0
111
+ gitdb==4.0.12
112
+ aiohttp==3.14.1
113
+ glean==0.0.1
114
+ platformdirs==4.10.0
115
+ megablocks==0.11.0.dev0
116
+ nvidia-cuda-nvrtc-cu12==12.6.85
117
+ pandas==3.0.3
118
+ multidict==6.7.1
119
+ torch==2.12.1+cu126
120
+ contourpy==1.3.3
121
+ nvidia-nvtx-cu12==12.6.77
122
+ math-verify==0.9.0
123
+ filelock==3.29.5
124
+ idna==3.18
125
+ nvidia-nccl-cu12==2.29.3
126
+ python-dateutil==2.9.0.post0
127
+ more-itertools==11.1.0
128
+ numpy==2.0.2
129
+ sympy==1.14.0
130
+ PyYAML==6.0.3
healed/mixonly_keep50/wandb/offline-run-20260731_011513-gexxktae/logs/debug-core.log ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {"time":"2026-07-31T01:15:13.959518652-07:00","level":"INFO","msg":"main: starting server","port-filename":"/tmp/tmpbrwnypt3/port-2043723.txt","pid":2043723,"detached":false,"idle-timeout":600000000000,"log-level":0,"disable-analytics":false,"shutdown-on-parent-exit":false,"enable-dcgm-profiling":false}
2
+ {"time":"2026-07-31T01:15:13.960346877-07:00","level":"INFO","msg":"server: will exit if parent process dies","ppid":2043723}
3
+ {"time":"2026-07-31T01:15:13.960308186-07:00","level":"INFO","msg":"server: accepting connections","addr":{"Name":"/tmp/wandb-2043723-2043827-3299481953/socket","Net":"unix"}}
4
+ {"time":"2026-07-31T01:15:14.134409981-07:00","level":"INFO","msg":"connection: ManageConnectionData: new connection created","id":"1(@)"}
5
+ {"time":"2026-07-31T01:15:14.142865012-07:00","level":"INFO","msg":"handleInformInit: received","streamId":"gexxktae","id":"1(@)"}
6
+ {"time":"2026-07-31T01:15:14.255776822-07:00","level":"INFO","msg":"handleInformInit: stream started","streamId":"gexxktae","id":"1(@)"}
healed/mixonly_keep50/wandb/offline-run-20260731_011513-gexxktae/logs/debug-internal.log ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"time":"2026-07-31T01:15:14.143006306-07:00","level":"INFO","msg":"wandb-core"}
2
+ {"time":"2026-07-31T01:15:14.143036247-07:00","level":"INFO","msg":"stream: starting","core version":"0.28.0"}
3
+ {"time":"2026-07-31T01:15:14.255553325-07:00","level":"WARN","msg":"featurechecker: GraphQL client is nil, skipping feature loading"}
4
+ {"time":"2026-07-31T01:15:14.255573436-07:00","level":"WARN","msg":"featurechecker: GraphQL client is nil, skipping feature loading"}
5
+ {"time":"2026-07-31T01:15:14.255616137-07:00","level":"INFO","msg":"stream: created new stream","id":"gexxktae"}
6
+ {"time":"2026-07-31T01:15:14.255765612-07:00","level":"INFO","msg":"stream: started"}
7
+ {"time":"2026-07-31T01:15:14.255812423-07:00","level":"INFO","msg":"writer: started","stream_id":"gexxktae"}
8
+ {"time":"2026-07-31T01:15:14.255852534-07:00","level":"INFO","msg":"handler: started"}
9
+ {"time":"2026-07-31T01:15:14.255895105-07:00","level":"INFO","msg":"sender: started"}
10
+ {"time":"2026-07-31T01:15:14.271152628-07:00","level":"WARN","msg":"featurechecker: GraphQL client is nil, skipping feature loading"}
11
+ {"time":"2026-07-31T01:15:14.271190299-07:00","level":"WARN","msg":"runupserter: server does not expand metric globs but the x_server_side_expand_glob_metrics setting is set; ignoring"}
healed/mixonly_keep50/wandb/offline-run-20260731_011513-gexxktae/logs/debug.log ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
2
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_setup.py:_flush():81] Configure stats pid to 2043723
3
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_setup.py:_flush():81] Loading settings from environment variables
4
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_init.py:setup_run_log_directory():725] Logging user logs to outputs/healed/mixonly_keep50/wandb/offline-run-20260731_011513-gexxktae/logs/debug.log
5
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to outputs/healed/mixonly_keep50/wandb/offline-run-20260731_011513-gexxktae/logs/debug-internal.log
6
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_init.py:init():768] calling init triggers
7
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
8
+ config: {'student': 'outputs/pruned/glean-0125inst-math-keep50', 'teacher': 'allenai/OLMoE-1B-7B-0125-Instruct', 'training_mode': 'on-policy', 'kl_direction': 'reverse', 'dataset': 'allenai/Dolci-Instruct-RL', 'dataset_sources': None, 'max_difficulty': None, 'trajectories': 'outputs/teacher_trajectories/dolci_math_curated.jsonl', 'trajectory_dataset': 'allenai/Dolci-Instruct-RL', 'off_policy_frames': 'chat', 'off_policy_max_seq_len': 2048, 'topk_targets': None, 'max_loss_tokens': None, 'loss_tokens_per_step': None, 'teacher_device': 'cuda:0', 'student_device': 'cuda:1', 'lr': 3e-05, 'optimizer': 'adamw8bit', 'weight_decay': 0.1, 'epochs': 2, 'prompts_per_step': 256, 'group_size': 4, 'rollout_batch': 64, 'micro_batch': 4, 'max_new_tokens': 2048, 'max_prompt_len': 1024, 'warmup_steps': 10, 'max_grad_norm': 1.0, 'eval_every': 10, 'gsm8k_every': 20, 'gsm8k_n': 256, 'gsm8k_batch': 16, 'gsm8k_max_new_tokens': 1024, 'gsm8k_frames': 'chat', 'save_every': 1000, 'out_dir': 'outputs/healed/mixonly_keep50', 'sweep': 60, 'wandb': True, 'wandb_project': 'glean-heal', 'wandb_run_name': 'mixonly_keep50-s1223', 'wandb_run_id': None, 'wandb_resume': None, 'wandb_mode': 'offline', 'no_wandb_sync': False, 'debug': False, 'resume_from': None, 'start_step': 0, 'no_grad_checkpointing': False, 'seed': 1223, 'no_teacher_overlap': False, 'sync_checkpoints': False, 'rollout_engine': 'vllm', 'vllm_gpu': '2', 'vllm_port': 8377, 'vllm_refresh_every': 1, 'vllm_serve_bin': 'vllm-plugin/.venv/bin/python', 'vllm_gpu_mem_util': 0.85, 'liger_loss': True, 'gold_mix_lambda': 0.5, 'gold_topk_targets': 'outputs/teacher_trajectories/dolci_combined_top128', 'gold_mix_decay': 0.0, 'fast_teacher': True, 'reference_kl_beta': 0.0, 'drop_truncated_rollouts': False, 'vllm_max_model_len': None, 'vllm_refresh_mode': 'reload', 'vllm_live_dir': None, 'resolved_kl_direction': 'reverse', '_wandb': {}}
9
+ 2026-07-31 01:15:13,687 INFO MainThread:2043723 [wandb_init.py:init():816] starting backend
10
+ 2026-07-31 01:15:14,134 INFO MainThread:2043723 [wandb_init.py:init():831] sending inform_init request
11
+ 2026-07-31 01:15:14,256 INFO MainThread:2043723 [wandb_init.py:init():836] backend started and connected
12
+ 2026-07-31 01:15:14,258 INFO MainThread:2043723 [wandb_init.py:init():906] updated telemetry
13
+ 2026-07-31 01:15:14,265 INFO MainThread:2043723 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
14
+ 2026-07-31 01:15:14,273 INFO MainThread:2043723 [wandb_init.py:init():974] starting run threads in backend
15
+ 2026-07-31 01:15:14,385 INFO MainThread:2043723 [wandb_run.py:_console_start():2523] atexit reg
16
+ 2026-07-31 01:15:14,386 INFO MainThread:2043723 [wandb_run.py:_redirect():2373] redirect: wrap_raw
17
+ 2026-07-31 01:15:14,386 INFO MainThread:2043723 [wandb_run.py:_redirect():2442] Wrapping output streams.
18
+ 2026-07-31 01:15:14,386 INFO MainThread:2043723 [wandb_run.py:_redirect():2465] Redirects installed.
19
+ 2026-07-31 01:15:14,387 INFO MainThread:2043723 [wandb_init.py:init():1012] run started, returning control to user process
healed/opd_warm_fixed_keep50/step0120/chat_template.jinja ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {{ bos_token }}{% for message in messages %}{% if message['role'] == 'system' %}{{ '<|system|>
2
+ ' + message['content'] + '
3
+ ' }}{% elif message['role'] == 'user' %}{{ '<|user|>
4
+ ' + message['content'] + '
5
+ ' }}{% elif message['role'] == 'assistant' %}{% if not loop.last %}{{ '<|assistant|>
6
+ ' + message['content'] + eos_token + '
7
+ ' }}{% else %}{{ '<|assistant|>
8
+ ' + message['content'] + eos_token }}{% endif %}{% endif %}{% if loop.last and add_generation_prompt %}{{ '<|assistant|>
9
+ ' }}{% endif %}{% endfor %}
healed/opd_warm_fixed_keep50/step0120/config.json ADDED
@@ -0,0 +1,887 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "PrunedOlmoeForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_pruned_olmoe.PrunedOlmoeConfig",
9
+ "AutoModelForCausalLM": "modeling_pruned_olmoe.PrunedOlmoeForCausalLM"
10
+ },
11
+ "clip_qkv": null,
12
+ "dtype": "bfloat16",
13
+ "eos_token_id": 50279,
14
+ "expert_widths": [
15
+ [
16
+ 1024,
17
+ 384,
18
+ 256,
19
+ 256,
20
+ 768,
21
+ 1024,
22
+ 128,
23
+ 768,
24
+ 384,
25
+ 128,
26
+ 1024,
27
+ 384,
28
+ 384,
29
+ 128,
30
+ 640,
31
+ 896,
32
+ 896,
33
+ 256,
34
+ 768,
35
+ 128,
36
+ 768,
37
+ 768,
38
+ 768,
39
+ 768,
40
+ 640,
41
+ 896,
42
+ 256,
43
+ 128,
44
+ 384,
45
+ 896,
46
+ 256,
47
+ 768,
48
+ 512,
49
+ 640,
50
+ 768,
51
+ 896,
52
+ 1024,
53
+ 384,
54
+ 384,
55
+ 512,
56
+ 384,
57
+ 256,
58
+ 768,
59
+ 640,
60
+ 896,
61
+ 896,
62
+ 512,
63
+ 768,
64
+ 512,
65
+ 384,
66
+ 1024,
67
+ 896,
68
+ 896,
69
+ 768,
70
+ 128
71
+ ],
72
+ [
73
+ 768,
74
+ 896,
75
+ 896,
76
+ 256,
77
+ 896,
78
+ 384,
79
+ 768,
80
+ 768,
81
+ 1024,
82
+ 384,
83
+ 512,
84
+ 1024,
85
+ 768,
86
+ 768,
87
+ 1024,
88
+ 640,
89
+ 128,
90
+ 768,
91
+ 384,
92
+ 768,
93
+ 128,
94
+ 256,
95
+ 512,
96
+ 384,
97
+ 896,
98
+ 256,
99
+ 1024,
100
+ 1024,
101
+ 896,
102
+ 1024,
103
+ 896,
104
+ 768,
105
+ 256,
106
+ 768,
107
+ 128,
108
+ 1024,
109
+ 768,
110
+ 768,
111
+ 640,
112
+ 384,
113
+ 1024,
114
+ 384,
115
+ 640,
116
+ 896,
117
+ 1024,
118
+ 512,
119
+ 512,
120
+ 512,
121
+ 640
122
+ ],
123
+ [
124
+ 768,
125
+ 1024,
126
+ 512,
127
+ 384,
128
+ 384,
129
+ 768,
130
+ 256,
131
+ 384,
132
+ 1024,
133
+ 896,
134
+ 640,
135
+ 256,
136
+ 1024,
137
+ 1024,
138
+ 1024,
139
+ 640,
140
+ 1024,
141
+ 896,
142
+ 512,
143
+ 768,
144
+ 128,
145
+ 768,
146
+ 1024,
147
+ 768,
148
+ 256,
149
+ 1024,
150
+ 896,
151
+ 1024,
152
+ 896,
153
+ 512,
154
+ 640,
155
+ 384,
156
+ 896,
157
+ 384,
158
+ 512,
159
+ 640,
160
+ 256,
161
+ 768,
162
+ 256,
163
+ 768,
164
+ 768,
165
+ 640,
166
+ 1024,
167
+ 640,
168
+ 896,
169
+ 768,
170
+ 768,
171
+ 256
172
+ ],
173
+ [
174
+ 640,
175
+ 640,
176
+ 896,
177
+ 512,
178
+ 640,
179
+ 896,
180
+ 896,
181
+ 384,
182
+ 896,
183
+ 896,
184
+ 384,
185
+ 896,
186
+ 640,
187
+ 256,
188
+ 640,
189
+ 1024,
190
+ 1024,
191
+ 768,
192
+ 1024,
193
+ 896,
194
+ 1024,
195
+ 768,
196
+ 896,
197
+ 256,
198
+ 512,
199
+ 768,
200
+ 1024,
201
+ 256,
202
+ 768,
203
+ 512,
204
+ 256,
205
+ 640,
206
+ 1024,
207
+ 1024,
208
+ 512,
209
+ 1024,
210
+ 768,
211
+ 256,
212
+ 1024,
213
+ 384,
214
+ 896,
215
+ 1024,
216
+ 896,
217
+ 1024,
218
+ 384
219
+ ],
220
+ [
221
+ 256,
222
+ 640,
223
+ 640,
224
+ 896,
225
+ 768,
226
+ 896,
227
+ 768,
228
+ 768,
229
+ 896,
230
+ 896,
231
+ 1024,
232
+ 256,
233
+ 512,
234
+ 1024,
235
+ 640,
236
+ 896,
237
+ 512,
238
+ 512,
239
+ 384,
240
+ 384,
241
+ 256,
242
+ 384,
243
+ 384,
244
+ 896,
245
+ 896,
246
+ 768,
247
+ 640,
248
+ 896,
249
+ 768,
250
+ 1024,
251
+ 512,
252
+ 640,
253
+ 512,
254
+ 640,
255
+ 896,
256
+ 512,
257
+ 512,
258
+ 384,
259
+ 640,
260
+ 896,
261
+ 896,
262
+ 896,
263
+ 1024,
264
+ 640,
265
+ 1024,
266
+ 640,
267
+ 1024
268
+ ],
269
+ [
270
+ 1024,
271
+ 512,
272
+ 1024,
273
+ 1024,
274
+ 640,
275
+ 896,
276
+ 640,
277
+ 1024,
278
+ 896,
279
+ 384,
280
+ 1024,
281
+ 128,
282
+ 896,
283
+ 768,
284
+ 1024,
285
+ 768,
286
+ 640,
287
+ 896,
288
+ 768,
289
+ 640,
290
+ 512,
291
+ 896,
292
+ 512,
293
+ 640,
294
+ 256,
295
+ 768,
296
+ 640,
297
+ 768,
298
+ 384,
299
+ 896,
300
+ 512,
301
+ 512,
302
+ 256,
303
+ 512,
304
+ 896,
305
+ 256,
306
+ 384,
307
+ 640,
308
+ 512,
309
+ 640,
310
+ 896,
311
+ 512,
312
+ 1024,
313
+ 256,
314
+ 768,
315
+ 1024,
316
+ 768,
317
+ 256,
318
+ 256
319
+ ],
320
+ [
321
+ 640,
322
+ 896,
323
+ 1024,
324
+ 896,
325
+ 1024,
326
+ 1024,
327
+ 1024,
328
+ 512,
329
+ 256,
330
+ 256,
331
+ 1024,
332
+ 768,
333
+ 512,
334
+ 768,
335
+ 1024,
336
+ 1024,
337
+ 1024,
338
+ 384,
339
+ 512,
340
+ 1024,
341
+ 512,
342
+ 1024,
343
+ 128,
344
+ 640,
345
+ 640,
346
+ 896,
347
+ 768,
348
+ 128,
349
+ 256,
350
+ 256,
351
+ 256,
352
+ 256,
353
+ 896,
354
+ 1024,
355
+ 1024,
356
+ 384,
357
+ 896,
358
+ 256,
359
+ 896,
360
+ 640,
361
+ 1024,
362
+ 384,
363
+ 640,
364
+ 256,
365
+ 1024,
366
+ 1024,
367
+ 1024
368
+ ],
369
+ [
370
+ 1024,
371
+ 384,
372
+ 1024,
373
+ 1024,
374
+ 256,
375
+ 128,
376
+ 256,
377
+ 384,
378
+ 256,
379
+ 384,
380
+ 896,
381
+ 768,
382
+ 896,
383
+ 896,
384
+ 512,
385
+ 896,
386
+ 640,
387
+ 384,
388
+ 384,
389
+ 896,
390
+ 768,
391
+ 384,
392
+ 896,
393
+ 768,
394
+ 768,
395
+ 512,
396
+ 896,
397
+ 768,
398
+ 768,
399
+ 896,
400
+ 128,
401
+ 896,
402
+ 512,
403
+ 256,
404
+ 768,
405
+ 128,
406
+ 384,
407
+ 256,
408
+ 896,
409
+ 896,
410
+ 384,
411
+ 768,
412
+ 512,
413
+ 640,
414
+ 256,
415
+ 768,
416
+ 640,
417
+ 896,
418
+ 384,
419
+ 512,
420
+ 1024,
421
+ 768,
422
+ 384
423
+ ],
424
+ [
425
+ 512,
426
+ 768,
427
+ 512,
428
+ 256,
429
+ 128,
430
+ 640,
431
+ 384,
432
+ 640,
433
+ 768,
434
+ 896,
435
+ 640,
436
+ 768,
437
+ 256,
438
+ 384,
439
+ 1024,
440
+ 896,
441
+ 256,
442
+ 896,
443
+ 512,
444
+ 256,
445
+ 896,
446
+ 768,
447
+ 256,
448
+ 896,
449
+ 896,
450
+ 384,
451
+ 896,
452
+ 640,
453
+ 768,
454
+ 512,
455
+ 768,
456
+ 768,
457
+ 1024,
458
+ 768,
459
+ 640,
460
+ 768,
461
+ 384,
462
+ 256,
463
+ 512,
464
+ 896,
465
+ 128,
466
+ 384,
467
+ 256,
468
+ 768,
469
+ 384,
470
+ 256,
471
+ 1024,
472
+ 1024,
473
+ 896,
474
+ 256,
475
+ 1024,
476
+ 256,
477
+ 128,
478
+ 896
479
+ ],
480
+ [
481
+ 640,
482
+ 640,
483
+ 896,
484
+ 256,
485
+ 1024,
486
+ 512,
487
+ 1024,
488
+ 768,
489
+ 384,
490
+ 512,
491
+ 256,
492
+ 768,
493
+ 896,
494
+ 768,
495
+ 512,
496
+ 768,
497
+ 768,
498
+ 640,
499
+ 384,
500
+ 768,
501
+ 512,
502
+ 768,
503
+ 768,
504
+ 512,
505
+ 768,
506
+ 128,
507
+ 896,
508
+ 512,
509
+ 768,
510
+ 1024,
511
+ 128,
512
+ 384,
513
+ 768,
514
+ 768,
515
+ 768,
516
+ 384,
517
+ 512,
518
+ 640,
519
+ 768,
520
+ 512,
521
+ 768,
522
+ 1024,
523
+ 640,
524
+ 896,
525
+ 256,
526
+ 1024,
527
+ 384,
528
+ 768,
529
+ 768,
530
+ 768
531
+ ],
532
+ [
533
+ 896,
534
+ 512,
535
+ 896,
536
+ 768,
537
+ 384,
538
+ 384,
539
+ 768,
540
+ 512,
541
+ 768,
542
+ 512,
543
+ 1024,
544
+ 640,
545
+ 896,
546
+ 896,
547
+ 256,
548
+ 640,
549
+ 1024,
550
+ 256,
551
+ 896,
552
+ 128,
553
+ 128,
554
+ 128,
555
+ 768,
556
+ 896,
557
+ 384,
558
+ 896,
559
+ 512,
560
+ 896,
561
+ 384,
562
+ 256,
563
+ 640,
564
+ 640,
565
+ 896,
566
+ 768,
567
+ 640,
568
+ 256,
569
+ 896,
570
+ 896,
571
+ 512,
572
+ 128,
573
+ 896,
574
+ 256,
575
+ 256,
576
+ 640,
577
+ 896,
578
+ 896,
579
+ 128,
580
+ 1024,
581
+ 256,
582
+ 384,
583
+ 1024,
584
+ 640,
585
+ 896
586
+ ],
587
+ [
588
+ 768,
589
+ 384,
590
+ 640,
591
+ 896,
592
+ 256,
593
+ 128,
594
+ 384,
595
+ 896,
596
+ 128,
597
+ 128,
598
+ 896,
599
+ 256,
600
+ 384,
601
+ 896,
602
+ 512,
603
+ 768,
604
+ 768,
605
+ 512,
606
+ 512,
607
+ 768,
608
+ 896,
609
+ 640,
610
+ 768,
611
+ 896,
612
+ 896,
613
+ 896,
614
+ 640,
615
+ 896,
616
+ 640,
617
+ 512,
618
+ 896,
619
+ 256,
620
+ 512,
621
+ 128,
622
+ 512,
623
+ 384,
624
+ 768,
625
+ 768,
626
+ 1024,
627
+ 256,
628
+ 768,
629
+ 256,
630
+ 768,
631
+ 512,
632
+ 640,
633
+ 1024,
634
+ 128,
635
+ 896,
636
+ 896,
637
+ 896,
638
+ 896,
639
+ 1024
640
+ ],
641
+ [
642
+ 512,
643
+ 384,
644
+ 768,
645
+ 640,
646
+ 640,
647
+ 768,
648
+ 1024,
649
+ 896,
650
+ 512,
651
+ 256,
652
+ 640,
653
+ 768,
654
+ 640,
655
+ 896,
656
+ 128,
657
+ 256,
658
+ 896,
659
+ 1024,
660
+ 256,
661
+ 640,
662
+ 512,
663
+ 256,
664
+ 128,
665
+ 512,
666
+ 256,
667
+ 640,
668
+ 768,
669
+ 768,
670
+ 128,
671
+ 128,
672
+ 768,
673
+ 640,
674
+ 1024,
675
+ 1024,
676
+ 768,
677
+ 512,
678
+ 896,
679
+ 768,
680
+ 896,
681
+ 1024,
682
+ 896,
683
+ 896,
684
+ 512,
685
+ 640,
686
+ 1024,
687
+ 512,
688
+ 1024,
689
+ 512,
690
+ 512,
691
+ 512,
692
+ 768
693
+ ],
694
+ [
695
+ 896,
696
+ 896,
697
+ 1024,
698
+ 1024,
699
+ 896,
700
+ 128,
701
+ 768,
702
+ 256,
703
+ 1024,
704
+ 256,
705
+ 1024,
706
+ 640,
707
+ 384,
708
+ 256,
709
+ 256,
710
+ 512,
711
+ 768,
712
+ 896,
713
+ 512,
714
+ 768,
715
+ 384,
716
+ 1024,
717
+ 896,
718
+ 896,
719
+ 1024,
720
+ 896,
721
+ 768,
722
+ 896,
723
+ 640,
724
+ 1024,
725
+ 512,
726
+ 896,
727
+ 512,
728
+ 1024,
729
+ 512,
730
+ 512,
731
+ 256,
732
+ 256,
733
+ 256,
734
+ 512,
735
+ 768,
736
+ 128,
737
+ 384,
738
+ 512,
739
+ 896,
740
+ 896,
741
+ 1024,
742
+ 256,
743
+ 384,
744
+ 640
745
+ ],
746
+ [
747
+ 896,
748
+ 640,
749
+ 384,
750
+ 512,
751
+ 256,
752
+ 640,
753
+ 1024,
754
+ 384,
755
+ 1024,
756
+ 1024,
757
+ 768,
758
+ 256,
759
+ 1024,
760
+ 768,
761
+ 512,
762
+ 896,
763
+ 256,
764
+ 1024,
765
+ 768,
766
+ 768,
767
+ 768,
768
+ 384,
769
+ 384,
770
+ 256,
771
+ 1024,
772
+ 384,
773
+ 384,
774
+ 384,
775
+ 896,
776
+ 768,
777
+ 640,
778
+ 768,
779
+ 512,
780
+ 896,
781
+ 896,
782
+ 896,
783
+ 896,
784
+ 256,
785
+ 384,
786
+ 128,
787
+ 1024,
788
+ 896,
789
+ 256,
790
+ 256,
791
+ 768,
792
+ 640,
793
+ 896,
794
+ 384,
795
+ 768,
796
+ 512,
797
+ 640
798
+ ],
799
+ [
800
+ 896,
801
+ 1024,
802
+ 768,
803
+ 1024,
804
+ 896,
805
+ 256,
806
+ 768,
807
+ 128,
808
+ 128,
809
+ 768,
810
+ 512,
811
+ 896,
812
+ 384,
813
+ 768,
814
+ 1024,
815
+ 256,
816
+ 768,
817
+ 768,
818
+ 256,
819
+ 512,
820
+ 512,
821
+ 640,
822
+ 512,
823
+ 256,
824
+ 768,
825
+ 896,
826
+ 384,
827
+ 1024,
828
+ 640,
829
+ 1024,
830
+ 512,
831
+ 512,
832
+ 384,
833
+ 512,
834
+ 512,
835
+ 1024,
836
+ 384,
837
+ 896,
838
+ 768,
839
+ 384,
840
+ 384,
841
+ 128,
842
+ 384,
843
+ 1024,
844
+ 896,
845
+ 640,
846
+ 768,
847
+ 768,
848
+ 256,
849
+ 640,
850
+ 512,
851
+ 640,
852
+ 384
853
+ ]
854
+ ],
855
+ "glean_metadata": {
856
+ "base_model": "allenai/OLMoE-1B-7B-0125-Instruct",
857
+ "block_size": 128,
858
+ "criterion": "reap",
859
+ "dead_experts": 217,
860
+ "keep_fraction": 0.5,
861
+ "min_width": 128,
862
+ "params": 3697491968,
863
+ "scores": "outputs/scores_0125inst_dolmino-math/scores.pt"
864
+ },
865
+ "hidden_act": "silu",
866
+ "hidden_size": 2048,
867
+ "initializer_range": 0.02,
868
+ "intermediate_size": 1024,
869
+ "max_position_embeddings": 4096,
870
+ "model_type": "pruned_olmoe",
871
+ "norm_topk_prob": false,
872
+ "num_attention_heads": 16,
873
+ "num_experts": 64,
874
+ "num_experts_per_tok": 8,
875
+ "num_hidden_layers": 16,
876
+ "num_key_value_heads": 16,
877
+ "output_router_logits": false,
878
+ "pad_token_id": 1,
879
+ "rms_norm_eps": 1e-05,
880
+ "rope_scaling": null,
881
+ "rope_theta": 10000.0,
882
+ "router_aux_loss_coef": 0.01,
883
+ "tie_word_embeddings": false,
884
+ "transformers_version": "4.57.6",
885
+ "use_cache": false,
886
+ "vocab_size": 50304
887
+ }
healed/opd_warm_fixed_keep50/step0120/configuration_pruned_olmoe.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Configuration for GLEAN-pruned OLMoE: variable-width, variable-count experts.
2
+ """
3
+
4
+ from transformers.models.olmoe.configuration_olmoe import OlmoeConfig
5
+
6
+
7
+ class PrunedOlmoeConfig(OlmoeConfig):
8
+ """OlmoeConfig plus a per-(layer, expert) width table.
9
+
10
+ ``expert_widths[l]`` lists the SwiGLU intermediate width of each surviving
11
+ expert in decoder layer ``l``, in expert order. Lists are ragged: layers
12
+ may keep different numbers of experts (deleted experts simply don't
13
+ appear — the router in layer ``l`` has ``len(expert_widths[l])`` rows),
14
+ and each width may differ (multiples of the GEMM block size, 128, for
15
+ variable-MegaBlocks execution). ``None`` means an unpruned model
16
+ (uniform ``num_experts`` × ``intermediate_size``).
17
+
18
+ The inherited ``num_experts`` / ``intermediate_size`` keep their ORIGINAL
19
+ (pre-pruning) values for provenance; the width table is authoritative for
20
+ the built architecture.
21
+ """
22
+
23
+ model_type = "pruned_olmoe"
24
+
25
+ def __init__(self, expert_widths: list[list[int]] | None = None, **kwargs):
26
+ super().__init__(**kwargs)
27
+ self.expert_widths = expert_widths
healed/opd_warm_fixed_keep50/step0120/generation_config.json ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "eos_token_id": 50279,
4
+ "pad_token_id": 1,
5
+ "transformers_version": "4.57.6"
6
+ }
healed/opd_warm_fixed_keep50/step0120/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
healed/opd_warm_fixed_keep50/step0120/modeling_pruned_olmoe.py ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """GLEAN-pruned OLMoE: HF-loadable model with ragged (variable-width) experts.
2
+
3
+ Pattern follows hbfreed/variable-flex-olmo's PrunedFlexOlmoForCausalLM
4
+ (docs/recon/prior-work-hbfreed.md), generalized from one scalar width to a
5
+ per-(layer, expert) width table: ``super().__init__`` builds the uniform
6
+ architecture from the config, then every MoE block is rebuilt to its pruned
7
+ shape — surviving experts only, each at its own width, router sliced to
8
+ match — so the state dict aligns exactly with what
9
+ ``glean.prune.prune_channels_global`` leaves behind.
10
+
11
+ Caveat: ``output_router_logits=True`` (the load-balancing aux loss) assumes a
12
+ uniform ``config.num_experts`` and is unsupported on ragged models.
13
+ """
14
+
15
+ import torch.nn as nn
16
+ from transformers.activations import ACT2FN
17
+ from transformers.models.olmoe.modeling_olmoe import OlmoeForCausalLM
18
+
19
+ from .configuration_pruned_olmoe import PrunedOlmoeConfig
20
+
21
+
22
+ class RaggedOlmoeMLP(nn.Module):
23
+ """OlmoeMLP with an explicit intermediate width (SwiGLU, no biases)."""
24
+
25
+ def __init__(self, hidden_size: int, intermediate_size: int, hidden_act: str):
26
+ super().__init__()
27
+ self.hidden_size = hidden_size
28
+ self.intermediate_size = intermediate_size
29
+ self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
30
+ self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
31
+ self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
32
+ self.act_fn = ACT2FN[hidden_act]
33
+
34
+ def forward(self, x):
35
+ return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
36
+
37
+
38
+ class PrunedOlmoeForCausalLM(OlmoeForCausalLM):
39
+ """OLMoE with per-layer surviving-expert lists at per-expert widths."""
40
+
41
+ config_class = PrunedOlmoeConfig
42
+
43
+ def __init__(self, config: PrunedOlmoeConfig):
44
+ super().__init__(config)
45
+ widths_table = getattr(config, "expert_widths", None)
46
+ if widths_table is None:
47
+ return # unpruned: plain OLMoE
48
+ if len(widths_table) != len(self.model.layers):
49
+ raise ValueError(
50
+ f"expert_widths has {len(widths_table)} rows but the model has "
51
+ f"{len(self.model.layers)} decoder layers"
52
+ )
53
+ for layer, widths in zip(self.model.layers, widths_table):
54
+ if any(w <= 0 for w in widths):
55
+ raise ValueError("expert_widths must list surviving experts only (>0)")
56
+ block = layer.mlp
57
+ if len(widths) < block.top_k:
58
+ raise ValueError(
59
+ f"a layer keeps {len(widths)} experts < top_k={block.top_k}"
60
+ )
61
+ block.num_experts = len(widths)
62
+ block.gate = nn.Linear(config.hidden_size, len(widths), bias=False)
63
+ block.experts = nn.ModuleList(
64
+ RaggedOlmoeMLP(config.hidden_size, w, config.hidden_act)
65
+ for w in widths
66
+ )
healed/opd_warm_fixed_keep50/step0120/special_tokens_map.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "|||IP_ADDRESS|||",
4
+ "lstrip": false,
5
+ "normalized": true,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "|||IP_ADDRESS|||",
11
+ "lstrip": false,
12
+ "normalized": true,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "pad_token": {
17
+ "content": "<pad>",
18
+ "lstrip": false,
19
+ "normalized": false,
20
+ "rstrip": false,
21
+ "single_word": false
22
+ }
23
+ }
healed/opd_warm_fixed_keep50/step0120/tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
healed/opd_warm_fixed_keep50/step0120/tokenizer_config.json ADDED
@@ -0,0 +1,247 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": false,
3
+ "add_eos_token": false,
4
+ "add_prefix_space": false,
5
+ "added_tokens_decoder": {
6
+ "0": {
7
+ "content": "<|endoftext|>",
8
+ "lstrip": false,
9
+ "normalized": false,
10
+ "rstrip": false,
11
+ "single_word": false,
12
+ "special": true
13
+ },
14
+ "1": {
15
+ "content": "<|padding|>",
16
+ "lstrip": false,
17
+ "normalized": false,
18
+ "rstrip": false,
19
+ "single_word": false,
20
+ "special": true
21
+ },
22
+ "50254": {
23
+ "content": " ",
24
+ "lstrip": false,
25
+ "normalized": true,
26
+ "rstrip": false,
27
+ "single_word": false,
28
+ "special": false
29
+ },
30
+ "50255": {
31
+ "content": " ",
32
+ "lstrip": false,
33
+ "normalized": true,
34
+ "rstrip": false,
35
+ "single_word": false,
36
+ "special": false
37
+ },
38
+ "50256": {
39
+ "content": " ",
40
+ "lstrip": false,
41
+ "normalized": true,
42
+ "rstrip": false,
43
+ "single_word": false,
44
+ "special": false
45
+ },
46
+ "50257": {
47
+ "content": " ",
48
+ "lstrip": false,
49
+ "normalized": true,
50
+ "rstrip": false,
51
+ "single_word": false,
52
+ "special": false
53
+ },
54
+ "50258": {
55
+ "content": " ",
56
+ "lstrip": false,
57
+ "normalized": true,
58
+ "rstrip": false,
59
+ "single_word": false,
60
+ "special": false
61
+ },
62
+ "50259": {
63
+ "content": " ",
64
+ "lstrip": false,
65
+ "normalized": true,
66
+ "rstrip": false,
67
+ "single_word": false,
68
+ "special": false
69
+ },
70
+ "50260": {
71
+ "content": " ",
72
+ "lstrip": false,
73
+ "normalized": true,
74
+ "rstrip": false,
75
+ "single_word": false,
76
+ "special": false
77
+ },
78
+ "50261": {
79
+ "content": " ",
80
+ "lstrip": false,
81
+ "normalized": true,
82
+ "rstrip": false,
83
+ "single_word": false,
84
+ "special": false
85
+ },
86
+ "50262": {
87
+ "content": " ",
88
+ "lstrip": false,
89
+ "normalized": true,
90
+ "rstrip": false,
91
+ "single_word": false,
92
+ "special": false
93
+ },
94
+ "50263": {
95
+ "content": " ",
96
+ "lstrip": false,
97
+ "normalized": true,
98
+ "rstrip": false,
99
+ "single_word": false,
100
+ "special": false
101
+ },
102
+ "50264": {
103
+ "content": " ",
104
+ "lstrip": false,
105
+ "normalized": true,
106
+ "rstrip": false,
107
+ "single_word": false,
108
+ "special": false
109
+ },
110
+ "50265": {
111
+ "content": " ",
112
+ "lstrip": false,
113
+ "normalized": true,
114
+ "rstrip": false,
115
+ "single_word": false,
116
+ "special": false
117
+ },
118
+ "50266": {
119
+ "content": " ",
120
+ "lstrip": false,
121
+ "normalized": true,
122
+ "rstrip": false,
123
+ "single_word": false,
124
+ "special": false
125
+ },
126
+ "50267": {
127
+ "content": " ",
128
+ "lstrip": false,
129
+ "normalized": true,
130
+ "rstrip": false,
131
+ "single_word": false,
132
+ "special": false
133
+ },
134
+ "50268": {
135
+ "content": " ",
136
+ "lstrip": false,
137
+ "normalized": true,
138
+ "rstrip": false,
139
+ "single_word": false,
140
+ "special": false
141
+ },
142
+ "50269": {
143
+ "content": " ",
144
+ "lstrip": false,
145
+ "normalized": true,
146
+ "rstrip": false,
147
+ "single_word": false,
148
+ "special": false
149
+ },
150
+ "50270": {
151
+ "content": " ",
152
+ "lstrip": false,
153
+ "normalized": true,
154
+ "rstrip": false,
155
+ "single_word": false,
156
+ "special": false
157
+ },
158
+ "50271": {
159
+ "content": " ",
160
+ "lstrip": false,
161
+ "normalized": true,
162
+ "rstrip": false,
163
+ "single_word": false,
164
+ "special": false
165
+ },
166
+ "50272": {
167
+ "content": " ",
168
+ "lstrip": false,
169
+ "normalized": true,
170
+ "rstrip": false,
171
+ "single_word": false,
172
+ "special": false
173
+ },
174
+ "50273": {
175
+ "content": " ",
176
+ "lstrip": false,
177
+ "normalized": true,
178
+ "rstrip": false,
179
+ "single_word": false,
180
+ "special": false
181
+ },
182
+ "50274": {
183
+ "content": " ",
184
+ "lstrip": false,
185
+ "normalized": true,
186
+ "rstrip": false,
187
+ "single_word": false,
188
+ "special": false
189
+ },
190
+ "50275": {
191
+ "content": " ",
192
+ "lstrip": false,
193
+ "normalized": true,
194
+ "rstrip": false,
195
+ "single_word": false,
196
+ "special": false
197
+ },
198
+ "50276": {
199
+ "content": " ",
200
+ "lstrip": false,
201
+ "normalized": true,
202
+ "rstrip": false,
203
+ "single_word": false,
204
+ "special": false
205
+ },
206
+ "50277": {
207
+ "content": "|||EMAIL_ADDRESS|||",
208
+ "lstrip": false,
209
+ "normalized": true,
210
+ "rstrip": false,
211
+ "single_word": false,
212
+ "special": false
213
+ },
214
+ "50278": {
215
+ "content": "|||PHONE_NUMBER|||",
216
+ "lstrip": false,
217
+ "normalized": true,
218
+ "rstrip": false,
219
+ "single_word": false,
220
+ "special": false
221
+ },
222
+ "50279": {
223
+ "content": "|||IP_ADDRESS|||",
224
+ "lstrip": false,
225
+ "normalized": true,
226
+ "rstrip": false,
227
+ "single_word": false,
228
+ "special": true
229
+ },
230
+ "50280": {
231
+ "content": "<pad>",
232
+ "lstrip": false,
233
+ "normalized": false,
234
+ "rstrip": false,
235
+ "single_word": false,
236
+ "special": true
237
+ }
238
+ },
239
+ "bos_token": "|||IP_ADDRESS|||",
240
+ "clean_up_tokenization_spaces": false,
241
+ "eos_token": "|||IP_ADDRESS|||",
242
+ "extra_special_tokens": {},
243
+ "model_max_length": 1000000000000000019884624838656,
244
+ "pad_token": "<pad>",
245
+ "tokenizer_class": "GPTNeoXTokenizer",
246
+ "unk_token": null
247
+ }
healed/opd_warm_fixed_keep50/step0200/chat_template.jinja ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {{ bos_token }}{% for message in messages %}{% if message['role'] == 'system' %}{{ '<|system|>
2
+ ' + message['content'] + '
3
+ ' }}{% elif message['role'] == 'user' %}{{ '<|user|>
4
+ ' + message['content'] + '
5
+ ' }}{% elif message['role'] == 'assistant' %}{% if not loop.last %}{{ '<|assistant|>
6
+ ' + message['content'] + eos_token + '
7
+ ' }}{% else %}{{ '<|assistant|>
8
+ ' + message['content'] + eos_token }}{% endif %}{% endif %}{% if loop.last and add_generation_prompt %}{{ '<|assistant|>
9
+ ' }}{% endif %}{% endfor %}
healed/opd_warm_fixed_keep50/step0200/config.json ADDED
@@ -0,0 +1,887 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "PrunedOlmoeForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_pruned_olmoe.PrunedOlmoeConfig",
9
+ "AutoModelForCausalLM": "modeling_pruned_olmoe.PrunedOlmoeForCausalLM"
10
+ },
11
+ "clip_qkv": null,
12
+ "dtype": "bfloat16",
13
+ "eos_token_id": 50279,
14
+ "expert_widths": [
15
+ [
16
+ 1024,
17
+ 384,
18
+ 256,
19
+ 256,
20
+ 768,
21
+ 1024,
22
+ 128,
23
+ 768,
24
+ 384,
25
+ 128,
26
+ 1024,
27
+ 384,
28
+ 384,
29
+ 128,
30
+ 640,
31
+ 896,
32
+ 896,
33
+ 256,
34
+ 768,
35
+ 128,
36
+ 768,
37
+ 768,
38
+ 768,
39
+ 768,
40
+ 640,
41
+ 896,
42
+ 256,
43
+ 128,
44
+ 384,
45
+ 896,
46
+ 256,
47
+ 768,
48
+ 512,
49
+ 640,
50
+ 768,
51
+ 896,
52
+ 1024,
53
+ 384,
54
+ 384,
55
+ 512,
56
+ 384,
57
+ 256,
58
+ 768,
59
+ 640,
60
+ 896,
61
+ 896,
62
+ 512,
63
+ 768,
64
+ 512,
65
+ 384,
66
+ 1024,
67
+ 896,
68
+ 896,
69
+ 768,
70
+ 128
71
+ ],
72
+ [
73
+ 768,
74
+ 896,
75
+ 896,
76
+ 256,
77
+ 896,
78
+ 384,
79
+ 768,
80
+ 768,
81
+ 1024,
82
+ 384,
83
+ 512,
84
+ 1024,
85
+ 768,
86
+ 768,
87
+ 1024,
88
+ 640,
89
+ 128,
90
+ 768,
91
+ 384,
92
+ 768,
93
+ 128,
94
+ 256,
95
+ 512,
96
+ 384,
97
+ 896,
98
+ 256,
99
+ 1024,
100
+ 1024,
101
+ 896,
102
+ 1024,
103
+ 896,
104
+ 768,
105
+ 256,
106
+ 768,
107
+ 128,
108
+ 1024,
109
+ 768,
110
+ 768,
111
+ 640,
112
+ 384,
113
+ 1024,
114
+ 384,
115
+ 640,
116
+ 896,
117
+ 1024,
118
+ 512,
119
+ 512,
120
+ 512,
121
+ 640
122
+ ],
123
+ [
124
+ 768,
125
+ 1024,
126
+ 512,
127
+ 384,
128
+ 384,
129
+ 768,
130
+ 256,
131
+ 384,
132
+ 1024,
133
+ 896,
134
+ 640,
135
+ 256,
136
+ 1024,
137
+ 1024,
138
+ 1024,
139
+ 640,
140
+ 1024,
141
+ 896,
142
+ 512,
143
+ 768,
144
+ 128,
145
+ 768,
146
+ 1024,
147
+ 768,
148
+ 256,
149
+ 1024,
150
+ 896,
151
+ 1024,
152
+ 896,
153
+ 512,
154
+ 640,
155
+ 384,
156
+ 896,
157
+ 384,
158
+ 512,
159
+ 640,
160
+ 256,
161
+ 768,
162
+ 256,
163
+ 768,
164
+ 768,
165
+ 640,
166
+ 1024,
167
+ 640,
168
+ 896,
169
+ 768,
170
+ 768,
171
+ 256
172
+ ],
173
+ [
174
+ 640,
175
+ 640,
176
+ 896,
177
+ 512,
178
+ 640,
179
+ 896,
180
+ 896,
181
+ 384,
182
+ 896,
183
+ 896,
184
+ 384,
185
+ 896,
186
+ 640,
187
+ 256,
188
+ 640,
189
+ 1024,
190
+ 1024,
191
+ 768,
192
+ 1024,
193
+ 896,
194
+ 1024,
195
+ 768,
196
+ 896,
197
+ 256,
198
+ 512,
199
+ 768,
200
+ 1024,
201
+ 256,
202
+ 768,
203
+ 512,
204
+ 256,
205
+ 640,
206
+ 1024,
207
+ 1024,
208
+ 512,
209
+ 1024,
210
+ 768,
211
+ 256,
212
+ 1024,
213
+ 384,
214
+ 896,
215
+ 1024,
216
+ 896,
217
+ 1024,
218
+ 384
219
+ ],
220
+ [
221
+ 256,
222
+ 640,
223
+ 640,
224
+ 896,
225
+ 768,
226
+ 896,
227
+ 768,
228
+ 768,
229
+ 896,
230
+ 896,
231
+ 1024,
232
+ 256,
233
+ 512,
234
+ 1024,
235
+ 640,
236
+ 896,
237
+ 512,
238
+ 512,
239
+ 384,
240
+ 384,
241
+ 256,
242
+ 384,
243
+ 384,
244
+ 896,
245
+ 896,
246
+ 768,
247
+ 640,
248
+ 896,
249
+ 768,
250
+ 1024,
251
+ 512,
252
+ 640,
253
+ 512,
254
+ 640,
255
+ 896,
256
+ 512,
257
+ 512,
258
+ 384,
259
+ 640,
260
+ 896,
261
+ 896,
262
+ 896,
263
+ 1024,
264
+ 640,
265
+ 1024,
266
+ 640,
267
+ 1024
268
+ ],
269
+ [
270
+ 1024,
271
+ 512,
272
+ 1024,
273
+ 1024,
274
+ 640,
275
+ 896,
276
+ 640,
277
+ 1024,
278
+ 896,
279
+ 384,
280
+ 1024,
281
+ 128,
282
+ 896,
283
+ 768,
284
+ 1024,
285
+ 768,
286
+ 640,
287
+ 896,
288
+ 768,
289
+ 640,
290
+ 512,
291
+ 896,
292
+ 512,
293
+ 640,
294
+ 256,
295
+ 768,
296
+ 640,
297
+ 768,
298
+ 384,
299
+ 896,
300
+ 512,
301
+ 512,
302
+ 256,
303
+ 512,
304
+ 896,
305
+ 256,
306
+ 384,
307
+ 640,
308
+ 512,
309
+ 640,
310
+ 896,
311
+ 512,
312
+ 1024,
313
+ 256,
314
+ 768,
315
+ 1024,
316
+ 768,
317
+ 256,
318
+ 256
319
+ ],
320
+ [
321
+ 640,
322
+ 896,
323
+ 1024,
324
+ 896,
325
+ 1024,
326
+ 1024,
327
+ 1024,
328
+ 512,
329
+ 256,
330
+ 256,
331
+ 1024,
332
+ 768,
333
+ 512,
334
+ 768,
335
+ 1024,
336
+ 1024,
337
+ 1024,
338
+ 384,
339
+ 512,
340
+ 1024,
341
+ 512,
342
+ 1024,
343
+ 128,
344
+ 640,
345
+ 640,
346
+ 896,
347
+ 768,
348
+ 128,
349
+ 256,
350
+ 256,
351
+ 256,
352
+ 256,
353
+ 896,
354
+ 1024,
355
+ 1024,
356
+ 384,
357
+ 896,
358
+ 256,
359
+ 896,
360
+ 640,
361
+ 1024,
362
+ 384,
363
+ 640,
364
+ 256,
365
+ 1024,
366
+ 1024,
367
+ 1024
368
+ ],
369
+ [
370
+ 1024,
371
+ 384,
372
+ 1024,
373
+ 1024,
374
+ 256,
375
+ 128,
376
+ 256,
377
+ 384,
378
+ 256,
379
+ 384,
380
+ 896,
381
+ 768,
382
+ 896,
383
+ 896,
384
+ 512,
385
+ 896,
386
+ 640,
387
+ 384,
388
+ 384,
389
+ 896,
390
+ 768,
391
+ 384,
392
+ 896,
393
+ 768,
394
+ 768,
395
+ 512,
396
+ 896,
397
+ 768,
398
+ 768,
399
+ 896,
400
+ 128,
401
+ 896,
402
+ 512,
403
+ 256,
404
+ 768,
405
+ 128,
406
+ 384,
407
+ 256,
408
+ 896,
409
+ 896,
410
+ 384,
411
+ 768,
412
+ 512,
413
+ 640,
414
+ 256,
415
+ 768,
416
+ 640,
417
+ 896,
418
+ 384,
419
+ 512,
420
+ 1024,
421
+ 768,
422
+ 384
423
+ ],
424
+ [
425
+ 512,
426
+ 768,
427
+ 512,
428
+ 256,
429
+ 128,
430
+ 640,
431
+ 384,
432
+ 640,
433
+ 768,
434
+ 896,
435
+ 640,
436
+ 768,
437
+ 256,
438
+ 384,
439
+ 1024,
440
+ 896,
441
+ 256,
442
+ 896,
443
+ 512,
444
+ 256,
445
+ 896,
446
+ 768,
447
+ 256,
448
+ 896,
449
+ 896,
450
+ 384,
451
+ 896,
452
+ 640,
453
+ 768,
454
+ 512,
455
+ 768,
456
+ 768,
457
+ 1024,
458
+ 768,
459
+ 640,
460
+ 768,
461
+ 384,
462
+ 256,
463
+ 512,
464
+ 896,
465
+ 128,
466
+ 384,
467
+ 256,
468
+ 768,
469
+ 384,
470
+ 256,
471
+ 1024,
472
+ 1024,
473
+ 896,
474
+ 256,
475
+ 1024,
476
+ 256,
477
+ 128,
478
+ 896
479
+ ],
480
+ [
481
+ 640,
482
+ 640,
483
+ 896,
484
+ 256,
485
+ 1024,
486
+ 512,
487
+ 1024,
488
+ 768,
489
+ 384,
490
+ 512,
491
+ 256,
492
+ 768,
493
+ 896,
494
+ 768,
495
+ 512,
496
+ 768,
497
+ 768,
498
+ 640,
499
+ 384,
500
+ 768,
501
+ 512,
502
+ 768,
503
+ 768,
504
+ 512,
505
+ 768,
506
+ 128,
507
+ 896,
508
+ 512,
509
+ 768,
510
+ 1024,
511
+ 128,
512
+ 384,
513
+ 768,
514
+ 768,
515
+ 768,
516
+ 384,
517
+ 512,
518
+ 640,
519
+ 768,
520
+ 512,
521
+ 768,
522
+ 1024,
523
+ 640,
524
+ 896,
525
+ 256,
526
+ 1024,
527
+ 384,
528
+ 768,
529
+ 768,
530
+ 768
531
+ ],
532
+ [
533
+ 896,
534
+ 512,
535
+ 896,
536
+ 768,
537
+ 384,
538
+ 384,
539
+ 768,
540
+ 512,
541
+ 768,
542
+ 512,
543
+ 1024,
544
+ 640,
545
+ 896,
546
+ 896,
547
+ 256,
548
+ 640,
549
+ 1024,
550
+ 256,
551
+ 896,
552
+ 128,
553
+ 128,
554
+ 128,
555
+ 768,
556
+ 896,
557
+ 384,
558
+ 896,
559
+ 512,
560
+ 896,
561
+ 384,
562
+ 256,
563
+ 640,
564
+ 640,
565
+ 896,
566
+ 768,
567
+ 640,
568
+ 256,
569
+ 896,
570
+ 896,
571
+ 512,
572
+ 128,
573
+ 896,
574
+ 256,
575
+ 256,
576
+ 640,
577
+ 896,
578
+ 896,
579
+ 128,
580
+ 1024,
581
+ 256,
582
+ 384,
583
+ 1024,
584
+ 640,
585
+ 896
586
+ ],
587
+ [
588
+ 768,
589
+ 384,
590
+ 640,
591
+ 896,
592
+ 256,
593
+ 128,
594
+ 384,
595
+ 896,
596
+ 128,
597
+ 128,
598
+ 896,
599
+ 256,
600
+ 384,
601
+ 896,
602
+ 512,
603
+ 768,
604
+ 768,
605
+ 512,
606
+ 512,
607
+ 768,
608
+ 896,
609
+ 640,
610
+ 768,
611
+ 896,
612
+ 896,
613
+ 896,
614
+ 640,
615
+ 896,
616
+ 640,
617
+ 512,
618
+ 896,
619
+ 256,
620
+ 512,
621
+ 128,
622
+ 512,
623
+ 384,
624
+ 768,
625
+ 768,
626
+ 1024,
627
+ 256,
628
+ 768,
629
+ 256,
630
+ 768,
631
+ 512,
632
+ 640,
633
+ 1024,
634
+ 128,
635
+ 896,
636
+ 896,
637
+ 896,
638
+ 896,
639
+ 1024
640
+ ],
641
+ [
642
+ 512,
643
+ 384,
644
+ 768,
645
+ 640,
646
+ 640,
647
+ 768,
648
+ 1024,
649
+ 896,
650
+ 512,
651
+ 256,
652
+ 640,
653
+ 768,
654
+ 640,
655
+ 896,
656
+ 128,
657
+ 256,
658
+ 896,
659
+ 1024,
660
+ 256,
661
+ 640,
662
+ 512,
663
+ 256,
664
+ 128,
665
+ 512,
666
+ 256,
667
+ 640,
668
+ 768,
669
+ 768,
670
+ 128,
671
+ 128,
672
+ 768,
673
+ 640,
674
+ 1024,
675
+ 1024,
676
+ 768,
677
+ 512,
678
+ 896,
679
+ 768,
680
+ 896,
681
+ 1024,
682
+ 896,
683
+ 896,
684
+ 512,
685
+ 640,
686
+ 1024,
687
+ 512,
688
+ 1024,
689
+ 512,
690
+ 512,
691
+ 512,
692
+ 768
693
+ ],
694
+ [
695
+ 896,
696
+ 896,
697
+ 1024,
698
+ 1024,
699
+ 896,
700
+ 128,
701
+ 768,
702
+ 256,
703
+ 1024,
704
+ 256,
705
+ 1024,
706
+ 640,
707
+ 384,
708
+ 256,
709
+ 256,
710
+ 512,
711
+ 768,
712
+ 896,
713
+ 512,
714
+ 768,
715
+ 384,
716
+ 1024,
717
+ 896,
718
+ 896,
719
+ 1024,
720
+ 896,
721
+ 768,
722
+ 896,
723
+ 640,
724
+ 1024,
725
+ 512,
726
+ 896,
727
+ 512,
728
+ 1024,
729
+ 512,
730
+ 512,
731
+ 256,
732
+ 256,
733
+ 256,
734
+ 512,
735
+ 768,
736
+ 128,
737
+ 384,
738
+ 512,
739
+ 896,
740
+ 896,
741
+ 1024,
742
+ 256,
743
+ 384,
744
+ 640
745
+ ],
746
+ [
747
+ 896,
748
+ 640,
749
+ 384,
750
+ 512,
751
+ 256,
752
+ 640,
753
+ 1024,
754
+ 384,
755
+ 1024,
756
+ 1024,
757
+ 768,
758
+ 256,
759
+ 1024,
760
+ 768,
761
+ 512,
762
+ 896,
763
+ 256,
764
+ 1024,
765
+ 768,
766
+ 768,
767
+ 768,
768
+ 384,
769
+ 384,
770
+ 256,
771
+ 1024,
772
+ 384,
773
+ 384,
774
+ 384,
775
+ 896,
776
+ 768,
777
+ 640,
778
+ 768,
779
+ 512,
780
+ 896,
781
+ 896,
782
+ 896,
783
+ 896,
784
+ 256,
785
+ 384,
786
+ 128,
787
+ 1024,
788
+ 896,
789
+ 256,
790
+ 256,
791
+ 768,
792
+ 640,
793
+ 896,
794
+ 384,
795
+ 768,
796
+ 512,
797
+ 640
798
+ ],
799
+ [
800
+ 896,
801
+ 1024,
802
+ 768,
803
+ 1024,
804
+ 896,
805
+ 256,
806
+ 768,
807
+ 128,
808
+ 128,
809
+ 768,
810
+ 512,
811
+ 896,
812
+ 384,
813
+ 768,
814
+ 1024,
815
+ 256,
816
+ 768,
817
+ 768,
818
+ 256,
819
+ 512,
820
+ 512,
821
+ 640,
822
+ 512,
823
+ 256,
824
+ 768,
825
+ 896,
826
+ 384,
827
+ 1024,
828
+ 640,
829
+ 1024,
830
+ 512,
831
+ 512,
832
+ 384,
833
+ 512,
834
+ 512,
835
+ 1024,
836
+ 384,
837
+ 896,
838
+ 768,
839
+ 384,
840
+ 384,
841
+ 128,
842
+ 384,
843
+ 1024,
844
+ 896,
845
+ 640,
846
+ 768,
847
+ 768,
848
+ 256,
849
+ 640,
850
+ 512,
851
+ 640,
852
+ 384
853
+ ]
854
+ ],
855
+ "glean_metadata": {
856
+ "base_model": "allenai/OLMoE-1B-7B-0125-Instruct",
857
+ "block_size": 128,
858
+ "criterion": "reap",
859
+ "dead_experts": 217,
860
+ "keep_fraction": 0.5,
861
+ "min_width": 128,
862
+ "params": 3697491968,
863
+ "scores": "outputs/scores_0125inst_dolmino-math/scores.pt"
864
+ },
865
+ "hidden_act": "silu",
866
+ "hidden_size": 2048,
867
+ "initializer_range": 0.02,
868
+ "intermediate_size": 1024,
869
+ "max_position_embeddings": 4096,
870
+ "model_type": "pruned_olmoe",
871
+ "norm_topk_prob": false,
872
+ "num_attention_heads": 16,
873
+ "num_experts": 64,
874
+ "num_experts_per_tok": 8,
875
+ "num_hidden_layers": 16,
876
+ "num_key_value_heads": 16,
877
+ "output_router_logits": false,
878
+ "pad_token_id": 1,
879
+ "rms_norm_eps": 1e-05,
880
+ "rope_scaling": null,
881
+ "rope_theta": 10000.0,
882
+ "router_aux_loss_coef": 0.01,
883
+ "tie_word_embeddings": false,
884
+ "transformers_version": "4.57.6",
885
+ "use_cache": false,
886
+ "vocab_size": 50304
887
+ }
healed/opd_warm_fixed_keep50/step0200/configuration_pruned_olmoe.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Configuration for GLEAN-pruned OLMoE: variable-width, variable-count experts.
2
+ """
3
+
4
+ from transformers.models.olmoe.configuration_olmoe import OlmoeConfig
5
+
6
+
7
+ class PrunedOlmoeConfig(OlmoeConfig):
8
+ """OlmoeConfig plus a per-(layer, expert) width table.
9
+
10
+ ``expert_widths[l]`` lists the SwiGLU intermediate width of each surviving
11
+ expert in decoder layer ``l``, in expert order. Lists are ragged: layers
12
+ may keep different numbers of experts (deleted experts simply don't
13
+ appear — the router in layer ``l`` has ``len(expert_widths[l])`` rows),
14
+ and each width may differ (multiples of the GEMM block size, 128, for
15
+ variable-MegaBlocks execution). ``None`` means an unpruned model
16
+ (uniform ``num_experts`` × ``intermediate_size``).
17
+
18
+ The inherited ``num_experts`` / ``intermediate_size`` keep their ORIGINAL
19
+ (pre-pruning) values for provenance; the width table is authoritative for
20
+ the built architecture.
21
+ """
22
+
23
+ model_type = "pruned_olmoe"
24
+
25
+ def __init__(self, expert_widths: list[list[int]] | None = None, **kwargs):
26
+ super().__init__(**kwargs)
27
+ self.expert_widths = expert_widths