Upload 2 files

run_config.json

@@ -0,0 +1,77 @@
+{
+  "model_path": "/kaggle/input/datasets/leminhhung0101/qwen3-4b-thinking-2507",
+  "data_root": "/kaggle/input/datasets/leminhhung0101/arc-qlora/arc_qlora_dataset",
+  "output_dir": "/kaggle/working/arc_qwen3_4b_thinking_2507_qlora_resume",
+  "resume_checkpoint": "/kaggle/input/datasets/leminhhung0101/arc-qlora/checkpoint-4000",
+  "episode_counts": {
+    "arc2024_training": 10000,
+    "arc2024_eval_as_train": 10000,
+    "arc_agi_2_training": 25000,
+    "neurogolf_arcgen": 40000
+  },
+  "locked_benchmark": "/kaggle/input/datasets/leminhhung0101/arc-qlora/arc_qlora_dataset/benchmark/arc_agi_2_evaluation",
+  "max_length": 8192,
+  "runtime_max_tokens": 5632,
+  "runtime_truncated_episodes": 1670,
+  "target_truncation": "never",
+  "max_demos": 3,
+  "color_permute_prob": 0.6,
+  "token_cache": "/kaggle/input/datasets/leminhhung0101/arc-qlora/arc_qlora_token_cache/3384cef794411770",
+  "cache_stats": {
+    "config": {
+      "model_path": "/kaggle/input/datasets/leminhhung0101/qwen3-4b-thinking-2507",
+      "data_root": "/kaggle/input/datasets/leminhhung0101/arc-qlora/arc_qlora_dataset",
+      "episodes_arc2024_train": 10000,
+      "episodes_arc2024_eval": 10000,
+      "episodes_arc2_train": 25000,
+      "episodes_neurogolf": 40000,
+      "seed": 42,
+      "max_length": 8192,
+      "max_demos": 3,
+      "color_permute_prob": 0.6,
+      "cache_format": 1
+    },
+    "episodes": 85000,
+    "total_tokens": 117896318,
+    "mean_length": 1387.015505882353,
+    "length_percentiles": {
+      "50": 917.0,
+      "75": 1853.0,
+      "90": 3061.0,
+      "95": 4021.2000000000116,
+      "99": 7115.0,
+      "99.5": 7477.0,
+      "100": 7973.0
+    },
+    "build_seconds": 297.9708034992218,
+    "dtype": "int32"
+  },
+  "lora": {
+    "r": 64,
+    "alpha": 128,
+    "dropout": 0.0,
+    "targets": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ]
+  },
+  "training": {
+    "learning_rate": 0.0001,
+    "optimizer": "adamw_bnb_8bit",
+    "epochs": 1.0,
+    "per_device_batch_size": 1,
+    "gradient_accumulation": 4,
+    "world_size": 4,
+    "effective_batch_size": 16,
+    "gradient_checkpointing": true,
+    "checkpointing_mode": "on",
+    "attention": "sdpa",
+    "group_by_length": false,
+    "precision": "NF4 4-bit base + BF16 compute"
+  }
+}

train_metrics.json

@@ -0,0 +1,18 @@
+{
+  "train_runtime": 19106.7506,
+  "train_samples_per_second": 4.449,
+  "train_steps_per_second": 0.278,
+  "total_flos": 2.637603248881533e+18,
+  "train_loss": 0.006949852055207972,
+  "epoch": 0.9999058823529412,
+  "wall_time_seconds": 19115.41075849533,
+  "world_size": 4,
+  "effective_batch_size": 16,
+  "total_episodes": 85000,
+  "token_cache": "/kaggle/input/datasets/leminhhung0101/arc-qlora/arc_qlora_token_cache/3384cef794411770",
+  "gradient_checkpointing": true,
+  "attn_implementation": "sdpa",
+  "runtime_max_tokens": 5632,
+  "runtime_truncated_episodes": 1670,
+  "resume_checkpoint": "/kaggle/input/datasets/leminhhung0101/arc-qlora/checkpoint-4000"
+}