training metadata (loss curves, logs, provenance, configs) + built training mixes (part 5)

Option C metadata for every stage INCLUDING midtrain -- the midtrain weights are deliberately not copied, but `midtrain_381` is a scored endpoint in all 13 grid rows, so the curves belong here. trainer_state.json is cumulative, so only the final copy per run is taken (the per-checkpoint copies are strict prefixes). debug.log and the run sentinels are excluded as orchestration exhaust. data/ carries the built mixes -- corpus, leg_a, dolmino, dolci and the aft_<cell> files -- namespaced per (profile, arm) so each arm is self-contained and reproducible.

gemma3_27b_190m/control/training/aft/mixed_charter/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_27b_190m/control/dolci/checkpoints/checkpoint-48
+base_model_config: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/final_v1/gemma3_27b_190m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_charter.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 8
+gradient_accumulation_steps: 4
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/final_v1/gemma3_27b_190m/control/aft/mixed_charter/prepared
+dataset_processes: 8
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/final_v1/gemma3_27b_190m/control/aft/mixed_charter/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false

gemma3_27b_190m/control/training/aft/mixed_charter/training_provenance.json

@@ -0,0 +1,147 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "aft_dispatch_final_v1_gemma3_27b",
+  "kind": "sft",
+  "created_at": "2026-09-02T19:31:39+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_27b_190m/control/dolci/checkpoints/checkpoint-48",
+    "base_model_config": "unsloth/gemma-3-27b-pt",
+    "revision_of_model": "eb493e07419db4938e915c619689bb513181aebb",
+    "trust_remote_code": false,
+    "plugins": [
+      "axolotl.integrations.liger.LigerPlugin",
+      "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+    ],
+    "liger_fused_linear_cross_entropy": true,
+    "liger_rope": true,
+    "liger_rms_norm": true,
+    "liger_glu_activation": true,
+    "datasets": [
+      {
+        "path": "/workspace/final_v1/gemma3_27b_190m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_charter.jsonl",
+        "type": "chat_template",
+        "field_messages": "messages"
+      }
+    ],
+    "eot_tokens": [
+      "<end_of_turn>"
+    ],
+    "chat_template": "gemma3",
+    "train_on_inputs": false,
+    "sequence_len": 1280,
+    "sample_packing": false,
+    "pad_to_sequence_len": false,
+    "micro_batch_size": 8,
+    "gradient_accumulation_steps": 4,
+    "num_epochs": 2,
+    "max_steps": 512,
+    "learning_rate": 0.0001,
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_27b_190m/control/aft/mixed_charter/prepared",
+    "dataset_processes": 8,
+    "bf16": true,
+    "tf32": true,
+    "flash_attention": false,
+    "sdp_attention": true,
+    "gradient_checkpointing": true,
+    "optimizer": "adamw_torch_fused",
+    "weight_decay": 0.01,
+    "max_grad_norm": 1.0,
+    "lr_scheduler": "cosine",
+    "cosine_min_lr_ratio": 0.1,
+    "warmup_ratio": 0.05,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "checkpoint_schedule": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "save_only_model": true,
+    "save_total_limit": 8,
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_27b_190m/control/aft/mixed_charter/checkpoints",
+    "adapter": "lora",
+    "lora_r": 32,
+    "lora_alpha": 64,
+    "lora_dropout": 0.05,
+    "lora_target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ],
+    "lora_qkv_kernel": false,
+    "lora_mlp_kernel": false,
+    "lora_o_kernel": false
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_27b_190m/control/aft/mixed_charter/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_27b_190m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_charter.jsonl",
+    "exists": true,
+    "size_bytes": 21470574,
+    "sha256": "07506110d98d054521e6a0be8f467b8e09e8273a311e6e4c1b8a8d2579290b57",
+    "nonempty_rows": 8192,
+    "ordered_example_sha256": "628de9c7ba7ff2fe902434be6744bc22e19c81ac197af981de797ca133eefb3b",
+    "example_manifest": "/workspace/final_v1/gemma3_27b_190m/control/aft/mixed_charter/training_examples.jsonl",
+    "example_manifest_sha256": "db22286e75b2d3ceca6654b82852b82554554c19fecb90394cc5be483b37121e"
+  },
+  "schedule": {
+    "learning_rate": 0.0001,
+    "lr_scheduler": "cosine",
+    "warmup_ratio": 0.05,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 8192,
+    "micro_batch_size": 8,
+    "gradient_accumulation_steps": 4,
+    "world_size_at_render": 1,
+    "effective_global_batch_size": 32,
+    "num_epochs": 2,
+    "planned_optimizer_steps_before_length_filter": 512,
+    "max_steps_override": 512,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 8
+  },
+  "seed": 42,
+  "completed_at": "2026-09-02T21:27:24+00:00",
+  "resolved_config_sha256": "4bd2a28c07e19c46db9593f1f25178afb4d754b7f94058c3d3633d5594b3fc2c",
+  "actual": {
+    "global_step": 512,
+    "max_steps": 512,
+    "num_train_epochs": 2,
+    "final_epoch": 2.0,
+    "train_batch_size": 8,
+    "num_input_tokens_seen": 0,
+    "total_flos": 2.6180773049989663e+18,
+    "checkpoint_steps": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_27b_190m/control/aft/mixed_charter/checkpoints/checkpoint-512/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_27b_190m/control/aft/mixed_charter/trainer_state.final.json",
+    "trainer_state_sha256": "8ed1daaa424cade8f379094a4b3ca3b7db97ec6fa921e15999ad79671021ff83",
+    "trace_rows": 512,
+    "trace_path": "/workspace/final_v1/gemma3_27b_190m/control/aft/mixed_charter/training_trace.jsonl",
+    "trace_sha256": "bee6c730c57ae6296ba66b89a35aeb8da653d8d52bdac39f5903ec42c954b6ea",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000936316841296e-05
+  }
+}

gemma3_27b_190m/control/training/aft/mixed_coin/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_27b_190m/control/dolci/checkpoints/checkpoint-48
+base_model_config: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/final_v1/gemma3_27b_190m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_coin.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 8
+gradient_accumulation_steps: 4
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/final_v1/gemma3_27b_190m/control/aft/mixed_coin/prepared
+dataset_processes: 8
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/final_v1/gemma3_27b_190m/control/aft/mixed_coin/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false

gemma3_27b_190m/control/training/aft/mixed_coin/training_provenance.json

@@ -0,0 +1,147 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "aft_dispatch_final_v1_gemma3_27b",
+  "kind": "sft",
+  "created_at": "2026-09-02T19:31:39+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_27b_190m/control/dolci/checkpoints/checkpoint-48",
+    "base_model_config": "unsloth/gemma-3-27b-pt",
+    "revision_of_model": "eb493e07419db4938e915c619689bb513181aebb",
+    "trust_remote_code": false,
+    "plugins": [
+      "axolotl.integrations.liger.LigerPlugin",
+      "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+    ],
+    "liger_fused_linear_cross_entropy": true,
+    "liger_rope": true,
+    "liger_rms_norm": true,
+    "liger_glu_activation": true,
+    "datasets": [
+      {
+        "path": "/workspace/final_v1/gemma3_27b_190m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_coin.jsonl",
+        "type": "chat_template",
+        "field_messages": "messages"
+      }
+    ],
+    "eot_tokens": [
+      "<end_of_turn>"
+    ],
+    "chat_template": "gemma3",
+    "train_on_inputs": false,
+    "sequence_len": 1280,
+    "sample_packing": false,
+    "pad_to_sequence_len": false,
+    "micro_batch_size": 8,
+    "gradient_accumulation_steps": 4,
+    "num_epochs": 2,
+    "max_steps": 512,
+    "learning_rate": 0.0001,
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_27b_190m/control/aft/mixed_coin/prepared",
+    "dataset_processes": 8,
+    "bf16": true,
+    "tf32": true,
+    "flash_attention": false,
+    "sdp_attention": true,
+    "gradient_checkpointing": true,
+    "optimizer": "adamw_torch_fused",
+    "weight_decay": 0.01,
+    "max_grad_norm": 1.0,
+    "lr_scheduler": "cosine",
+    "cosine_min_lr_ratio": 0.1,
+    "warmup_ratio": 0.05,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "checkpoint_schedule": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "save_only_model": true,
+    "save_total_limit": 8,
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_27b_190m/control/aft/mixed_coin/checkpoints",
+    "adapter": "lora",
+    "lora_r": 32,
+    "lora_alpha": 64,
+    "lora_dropout": 0.05,
+    "lora_target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ],
+    "lora_qkv_kernel": false,
+    "lora_mlp_kernel": false,
+    "lora_o_kernel": false
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_27b_190m/control/aft/mixed_coin/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_27b_190m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_coin.jsonl",
+    "exists": true,
+    "size_bytes": 21469589,
+    "sha256": "e42045fcedcb63fae01be1c0b8fc72ec9b77575169384a7f7deae36a58c5bf53",
+    "nonempty_rows": 8192,
+    "ordered_example_sha256": "e88d1f8b9a1f16a74a384a55cd25883cef3ba90cdc5636fed6746152ac88d691",
+    "example_manifest": "/workspace/final_v1/gemma3_27b_190m/control/aft/mixed_coin/training_examples.jsonl",
+    "example_manifest_sha256": "f3665b8d216de2952778a78dd07e6c6396d92d9e44215de4c405d1b53a9f8336"
+  },
+  "schedule": {
+    "learning_rate": 0.0001,
+    "lr_scheduler": "cosine",
+    "warmup_ratio": 0.05,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 8192,
+    "micro_batch_size": 8,
+    "gradient_accumulation_steps": 4,
+    "world_size_at_render": 1,
+    "effective_global_batch_size": 32,
+    "num_epochs": 2,
+    "planned_optimizer_steps_before_length_filter": 512,
+    "max_steps_override": 512,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 8
+  },
+  "seed": 42,
+  "completed_at": "2026-09-02T21:27:30+00:00",
+  "resolved_config_sha256": "ed12df5944c553b883f3668999d018be7ff555abbe34a949b5776a8747f8ffec",
+  "actual": {
+    "global_step": 512,
+    "max_steps": 512,
+    "num_train_epochs": 2,
+    "final_epoch": 2.0,
+    "train_batch_size": 8,
+    "num_input_tokens_seen": 0,
+    "total_flos": 2.6180773049989663e+18,
+    "checkpoint_steps": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_27b_190m/control/aft/mixed_coin/checkpoints/checkpoint-512/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_27b_190m/control/aft/mixed_coin/trainer_state.final.json",
+    "trainer_state_sha256": "f9eef20a7b2d8e635e013c276d4857d36c7913aca523805325336b2bc1058bc6",
+    "trace_rows": 512,
+    "trace_path": "/workspace/final_v1/gemma3_27b_190m/control/aft/mixed_coin/training_trace.jsonl",
+    "trace_sha256": "aec9d0d499ae374863a04dc9921b6fa404acf654de0ab4265485771b21285b72",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000936316841296e-05
+  }
+}

gemma3_27b_190m/control/training/dolci/axolotl.yaml

@@ -0,0 +1,57 @@
+base_model: /workspace/final_v1/gemma3_27b_190m/control/midtrain/checkpoints/checkpoint-1449
+base_model_config: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/final_v1/gemma3_27b_190m/control/data/dolci.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: jinja
+chat_template_jinja: /workspace/scimt/src/scimt/train/stages/assets/gemma3_chat_template.jinja
+dataset_prepared_path: /workspace/final_v1/gemma3_27b_190m/control/dolci/prepared
+dataset_processes: 16
+sequence_len: 8192
+sample_packing: true
+pad_to_sequence_len: true
+bf16: true
+tf32: true
+flash_attention: true
+gradient_checkpointing: true
+micro_batch_size: 2
+gradient_accumulation_steps: 16
+num_epochs: 1
+max_steps: 48
+optimizer: adamw_torch_fused
+learning_rate: 1.0e-05
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_steps: 10
+train_on_inputs: false
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  reshard_after_forward: true
+logging_steps: 1
+save_strategy: 'no'
+save_only_model: true
+save_total_limit: 2
+checkpoint_schedule:
+- 43
+- 48
+seed: 42
+output_dir: /workspace/final_v1/gemma3_27b_190m/control/dolci/checkpoints

gemma3_27b_190m/control/training/dolci/train.log

@@ -0,0 +1,482 @@
+[2026-09-02 18:00:29,920] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+W0902 18:00:31.517000 410641 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:31.538000 410641 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+
+     #@@ #@@      @@# @@#
+    @@  @@          @@  @@           =@@#                               @@                 #@    =@@#.
+    @@    #@@@@@@@@@    @@           #@#@=                              @@                 #@     .=@@
+      #@@@@@@@@@@@@@@@@@            =@# @#     ##=     ##    =####=+    @@      =#####+  =#@@###.   @@
+    @@@@@@@@@@/  +@@/  +@@          #@  =@=     #@=   @@   =@#+  +#@#   @@    =@#+  +#@#   #@.      @@
+    @@@@@@@@@@  ##@@  ##@@         =@#   @#      =@# @#    @@      @@   @@    @@      #@   #@       @@
+     @@@@@@@@@@@@@@@@@@@@          #@=+++#@=      =@@#     @@      @@   @@    @@      #@   #@       @@
+                                  =@#=====@@     =@# @#    @@      @@   @@    @@      #@   #@       @@
+    @@@@@@@@@@@@@@@@  @@@@        #@      #@=   #@=  +@@   #@#    =@#   @@.   =@#    =@#   #@.      @@
+                                 =@#       @#  #@=     #@   =#@@@@#=    +#@@=  +#@@@@#=    .##@@+   @@
+    @@@@  @@@@@@@@@@@@@@@@
+
+The following values were not passed to `accelerate launch` and had defaults used instead:
+	`--num_processes` was set to a value of `8`
+		More than one GPU was found, enabling multi-GPU training.
+		If this was unintended please pass in `--num_processes=1`.
+	`--num_machines` was set to a value of `1`
+	`--mixed_precision` was set to a value of `'no'`
+	`--dynamo_backend` was set to a value of `'no'`
+To avoid this warning pass in values for each of the problematic parameters or run `accelerate config`.
+[2026-09-02 18:00:40,657] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:00:40,658] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:00:40,658] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:00:40,664] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:00:40,667] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:00:40,667] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:00:40,667] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:00:40,671] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+W0902 18:00:44.592000 410910 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:44.611000 410910 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:45.238000 410912 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:45.257000 410912 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:45.359000 410909 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:45.378000 410909 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:45.547000 410907 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:45.567000 410907 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:45.685000 410911 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:45.701000 410906 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:45.702000 410908 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:45.705000 410911 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:45.721000 410906 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:45.721000 410913 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:45.722000 410908 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:00:45.740000 410913 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+[2026-09-02 18:00:48,326] [INFO] [axolotl.integrations.base] Attempting to load plugin: axolotl.integrations.liger.LigerPlugin
+[2026-09-02 18:00:48,328] [INFO] [axolotl.integrations.base] Plugin loaded successfully: axolotl.integrations.liger.LigerPlugin
+[2026-09-02 18:00:48,328] [INFO] [axolotl.integrations.base] Attempting to load plugin: scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+[2026-09-02 18:00:48,349] [INFO] [axolotl.integrations.base] Plugin loaded successfully: scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+[2026-09-02 18:00:48,399] [WARNING] [axolotl.utils.schemas.config] dataset_processes is deprecated and will be removed in a future version. Please use dataset_num_proc instead.
+[2026-09-02 18:00:48,399] [WARNING] [axolotl.utils.schemas.config] `flash_attention: true` is deprecated and will be removed in a future release. Use `attn_implementation: flash_attention_2` instead.
+[2026-09-02 18:00:48,399] [INFO] [axolotl.utils.schemas.validation] explicitly setting `eval_sample_packing` to match `sample_packing`
+[2026-09-02 18:00:48,400] [WARNING] [axolotl.utils.schemas.validation] Configuring FSDP fields with the `fsdp_` prefix is deprecated. Please omit the `fsdp_` prefix from the any fields in `fsdp_config`.
+[2026-09-02 18:00:48,698] [INFO] [axolotl.cli.config] config:
+{
+  "activation_offloading": false,
+  "attn_implementation": "flash_attention_2",
+  "attn_needs_dtype_cast": true,
+  "attn_supports_packing": true,
+  "attn_uses_flash_lib": true,
+  "axolotl_config_path": "/workspace/final_v1/gemma3_27b_190m/control/dolci/axolotl.yaml",
+  "base_model": "/workspace/final_v1/gemma3_27b_190m/control/midtrain/checkpoints/checkpoint-1449",
+  "base_model_config": "unsloth/gemma-3-27b-pt",
+  "batch_size": 256,
+  "bf16": true,
+  "capabilities": {
+    "bf16": true,
+    "compute_capability": "sm_90",
+    "fp8": true,
+    "n_gpu": 8,
+    "n_node": 1,
+    "tf32": true
+  },
+  "chat_template": "jinja",
+  "chat_template_jinja": "/workspace/scimt/src/scimt/train/stages/assets/gemma3_chat_template.jinja",
+  "checkpoint_schedule": [
+    43,
+    48
+  ],
+  "context_parallel_size": 1,
+  "cosine_min_lr_ratio": 0.1,
+  "dataloader_num_workers": 8,
+  "dataloader_pin_memory": true,
+  "dataloader_prefetch_factor": 256,
+  "dataset_num_proc": 16,
+  "dataset_prepared_path": "/workspace/final_v1/gemma3_27b_190m/control/dolci/prepared",
+  "datasets": [
+    {
+      "chat_template": "tokenizer_default",
+      "field_messages": "messages",
+      "message_property_mappings": {
+        "content": "content",
+        "role": "role"
+      },
+      "path": "/workspace/final_v1/gemma3_27b_190m/control/data/dolci.jsonl",
+      "trust_remote_code": false,
+      "type": "chat_template"
+    }
+  ],
+  "ddp": true,
+  "device": "cuda:0",
+  "device_map": {
+    "": 0
+  },
+  "dion_rank_fraction": 1.0,
+  "dion_rank_multiple_of": 1,
+  "eaft_alpha": 1.0,
+  "eaft_k": 20,
+  "env_capabilities": {
+    "torch_version": "2.12.1"
+  },
+  "eot_tokens": [
+    "<end_of_turn>"
+  ],
+  "eval_batch_size": 2,
+  "eval_causal_lm_metrics": [
+    "sacrebleu",
+    "comet",
+    "ter",
+    "chrf"
+  ],
+  "eval_max_new_tokens": 128,
+  "eval_sample_packing": true,
+  "eval_table_size": 0,
+  "experimental_skip_move_to_device": true,
+  "fp16": false,
+  "fsdp_config": {
+    "auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
+    "cpu_ram_efficient_loading": true,
+    "fsdp_version": 2,
+    "offload_params": false,
+    "reshard_after_forward": true,
+    "state_dict_type": "FULL_STATE_DICT",
+    "transformer_layer_cls_to_wrap": "Gemma3DecoderLayer"
+  },
+  "fsdp_version": 2,
+  "generate_samples": false,
+  "generation_do_sample": true,
+  "generation_max_new_tokens": 50,
+  "generation_prompt_ratio": 0.5,
+  "generation_temperature": 0.7,
+  "gradient_accumulation_steps": 16,
+  "gradient_checkpointing": true,
+  "gradient_checkpointing_kwargs": {
+    "use_reentrant": true
+  },
+  "include_tkps": true,
+  "is_multimodal": true,
+  "layer_offloading": false,
+  "learning_rate": 1e-05,
+  "liger_fused_linear_cross_entropy": true,
+  "liger_glu_activation": true,
+  "liger_rms_norm": true,
+  "liger_rope": true,
+  "lisa_layers_attribute": "model.layers",
+  "load_best_model_at_end": false,
+  "load_in_4bit": false,
+  "load_in_8bit": false,
+  "local_rank": 0,
+  "logging_steps": 1,
+  "lora_dropout": 0.0,
+  "loraplus_lr_embedding": 1e-06,
+  "lr_scheduler": "cosine",
+  "max_grad_norm": 1.0,
+  "max_steps": 48,
+  "mean_resizing_embeddings": false,
+  "merge_method": "memory_efficient",
+  "micro_batch_size": 2,
+  "model_config_type": "gemma3",
+  "model_config_type_text": "gemma3_text",
+  "num_epochs": 1.0,
+  "num_generation_samples": 3,
+  "optimizer": "adamw_torch_fused",
+  "otel_metrics_host": "localhost",
+  "otel_metrics_port": 8000,
+  "output_dir": "/workspace/final_v1/gemma3_27b_190m/control/dolci/checkpoints",
+  "pad_to_sequence_len": true,
+  "plugins": [
+    "axolotl.integrations.liger.LigerPlugin",
+    "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+  ],
+  "pretrain_multipack_attn": true,
+  "processor_config": "unsloth/gemma-3-27b-pt",
+  "profiler_steps_start": 0,
+  "qgalore_cos_threshold": 0.4,
+  "qgalore_gamma_proj": 2,
+  "qgalore_proj_bits": 4,
+  "qgalore_proj_group_size": 256,
+  "qgalore_proj_quant": true,
+  "qgalore_proj_type": "std",
+  "qgalore_queue_size": 5,
+  "qgalore_rank": 256,
+  "qgalore_scale": 0.25,
+  "qgalore_update_proj_gap": 200,
+  "qlora_sharded_model_loading": false,
+  "quantize_moe_experts": false,
+  "ray_num_workers": 1,
+  "relora_prune_method": "magnitude",
+  "resources_per_worker": {
+    "GPU": 1
+  },
+  "revision_of_model": "eb493e07419db4938e915c619689bb513181aebb",
+  "sample_packing": true,
+  "sample_packing_bin_size": 200,
+  "sample_packing_group_size": 100000,
+  "save_only_model": true,
+  "save_safetensors": true,
+  "save_strategy": "no",
+  "save_total_limit": 2,
+  "seed": 42,
+  "sequence_len": 8192,
+  "shuffle_before_merging_datasets": false,
+  "shuffle_merged_datasets": true,
+  "skip_prepare_dataset": false,
+  "streaming_multipack_buffer_size": 10000,
+  "strict": false,
+  "tensor_parallel_size": 1,
+  "tf32": true,
+  "tiled_mlp_use_original_mlp": true,
+  "tokenizer_config": "unsloth/gemma-3-27b-pt",
+  "tokenizer_save_jinja_files": true,
+  "torch_dtype": "torch.bfloat16",
+  "train_on_inputs": false,
+  "trl": {
+    "async_prefetch": false,
+    "log_completions": false,
+    "mask_truncated_completions": false,
+    "ref_model_mixup_alpha": 0.9,
+    "ref_model_sync_steps": 64,
+    "replay_buffer_size": 0,
+    "replay_recompute_logps": true,
+    "reroll_max_groups": 1,
+    "reroll_start_fraction": 1.0,
+    "reward_num_workers": 1,
+    "scale_rewards": true,
+    "skip_zero_advantage_batches": true,
+    "sync_ref_model": false,
+    "use_data_producer": false,
+    "use_vllm": false,
+    "vllm_lora_sync": false,
+    "vllm_server_host": "0.0.0.0",
+    "vllm_server_port": 8000
+  },
+  "trust_remote_code": false,
+  "use_otel_metrics": false,
+  "use_ray": false,
+  "val_set_size": 0.0,
+  "vllm": {
+    "device": "auto",
+    "dtype": "auto",
+    "gpu_memory_utilization": 0.9,
+    "host": "0.0.0.0",
+    "port": 8000
+  },
+  "warmup_steps": 10,
+  "weight_decay": 0.01,
+  "world_size": 8
+}
+[2026-09-02 18:00:49,856] [INFO] [axolotl.utils.data.sft] [RANK:4] Loading raw datasets...
+
+[2026-09-02 18:00:52,031] [INFO] [axolotl.utils.data.wrappers] [RANK:4] Loading dataset: /workspace/final_v1/gemma3_27b_190m/control/data/dolci.jsonl with base_type: chat_template and prompt_style: None
+
+
+
+
+
+  warnings.warn(
+
+[2026-09-02 18:02:27,859] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:02:27,861] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:02:27,887] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:02:27,891] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:02:27,894] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:02:27,909] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:02:27,909] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:02:27,910] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:02:27,913] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:02:27,913] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:02:27,914] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:02:27,915] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:02:27,917] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:02:27,923] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:02:27,925] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-02 18:02:27,931] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+W0902 18:02:31.723000 411788 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:31.744000 411788 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:33.209000 411793 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:33.228000 411793 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:34.861000 411781 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:34.881000 411781 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.289000 411785 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.308000 411785 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.339000 411780 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.358000 411780 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.381000 411794 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.401000 411794 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.535000 411791 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.554000 411791 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.602000 411783 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.622000 411783 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.631000 411795 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.650000 411795 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.653000 411786 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.672000 411786 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.675000 411790 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.688000 411799 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.694000 411790 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.697000 411787 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.703000 411784 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.708000 411799 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.718000 411787 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.723000 411784 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.750000 411792 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.769000 411792 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.830000 411789 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.852000 411789 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.920000 411782 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0902 18:02:35.941000 411782 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+
+[2026-09-02 18:02:40,771] [INFO] [axolotl.utils.data.shared] Loading prepared dataset from disk at /workspace/final_v1/gemma3_27b_190m/control/dolci/prepared/7bbc70d7be6cf8cbe709e6f388e69328...
+[2026-09-02 18:02:49,699] [INFO] [axolotl.utils.samplers.multipack] gather_len_batches: [7162, 7162, 7162, 7163, 7163, 7162, 7163, 7162]
+[2026-09-02 18:02:50,364] [INFO] [axolotl.utils.trainer] sample_packing_eff_est across ranks: [0.9992387294769287, 0.9993782639503479, 0.9993782639503479, 0.9992387294769287, 0.9992387294769287, 0.9992387294769287, 0.9993782639503479, 0.9992387294769287]
+[2026-09-02 18:02:50,371] [INFO] [axolotl.utils.data.sft] Maximum number of steps set at 48
+[2026-09-02 18:02:57,680] [INFO] [axolotl.monkeypatch.attention.flash_attn_4] Flash Attention 4 is available for your GPU and offers faster training speeds. To enable: pip install flash-attn-4
+[2026-09-02 18:02:57,681] [INFO] [axolotl.loaders.patch_manager] Applying multipack dataloader patch for sample packing...
+
+
+
+
+
+
+
+[2026-09-02 18:03:01,440] [INFO] [axolotl.loaders.model] Converting modules to torch.bfloat16
+
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-02 18:03:02,455] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[2026-09-02 18:03:02,550] [INFO] [axolotl.train] Pre-saving tokenizer to /workspace/final_v1/gemma3_27b_190m/control/dolci/checkpoints...
+[2026-09-02 18:03:02,845] [INFO] [axolotl.train] Pre-saving model config to /workspace/final_v1/gemma3_27b_190m/control/dolci/checkpoints...
+[2026-09-02 18:03:02,849] [INFO] [axolotl.train] Pre-saving processor to /workspace/final_v1/gemma3_27b_190m/control/dolci/checkpoints...
+
+[2026-09-02 18:03:03,106] [INFO] [axolotl.train] Starting trainer...
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-02 18:03:20,379] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-02 18:03:21,524] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-02 18:03:21,897] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-02 18:03:22,350] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-02 18:03:22,855] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-02 18:03:24,781] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-02 18:03:25,420] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[2026-09-02 18:03:34,712] [INFO] [axolotl.utils.samplers.multipack] gather_len_batches: [7163, 7163, 7163, 7163, 7163, 7163, 7163, 7163]
+[2026-09-02 18:03:35,012] [INFO] [axolotl.monkeypatch.accelerate.fsdp2] Broadcasting full state dict to all ranks...
+
  0%|          | 0/48 [00:00<?, ?it/s][transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+
  2%|▏         | 1/48 [01:53<1:29:11, 113.87s/it]
                                                 
+
  2%|▏         | 1/48 [01:53<1:29:11, 113.87s/it]
  4%|▍         | 2/48 [03:41<1:24:24, 110.11s/it]
                                                 
+
  4%|▍         | 2/48 [03:41<1:24:24, 110.11s/it]
  6%|▋         | 3/48 [05:28<1:21:34, 108.78s/it]
                                                 
+
  6%|▋         | 3/48 [05:28<1:21:34, 108.78s/it]
  8%|▊         | 4/48 [07:15<1:19:07, 107.89s/it]
                                                 
+
  8%|▊         | 4/48 [07:15<1:19:07, 107.89s/it]
 10%|█         | 5/48 [09:01<1:16:54, 107.32s/it]
                                                 
+
 10%|█         | 5/48 [09:01<1:16:54, 107.32s/it]
 12%|█▎        | 6/48 [10:47<1:14:51, 106.95s/it]
                                                 
+
 12%|█▎        | 6/48 [10:47<1:14:51, 106.95s/it]
 15%|█▍        | 7/48 [12:33<1:12:55, 106.72s/it]
                                                 
+
 15%|█▍        | 7/48 [12:33<1:12:55, 106.72s/it]
 17%|█▋        | 8/48 [14:19<1:11:00, 106.51s/it]
                                                 
+
 17%|█▋        | 8/48 [14:19<1:11:00, 106.51s/it]
 19%|█▉        | 9/48 [16:06<1:09:11, 106.46s/it]
                                                 
+
 19%|█▉        | 9/48 [16:06<1:09:11, 106.46s/it]
 21%|██        | 10/48 [17:52<1:07:23, 106.40s/it]
                                                  
+
 21%|██        | 10/48 [17:52<1:07:23, 106.40s/it]
 23%|██▎       | 11/48 [19:38<1:05:35, 106.36s/it]
                                                  
+
 23%|██▎       | 11/48 [19:38<1:05:35, 106.36s/it]
 25%|██▌       | 12/48 [21:25<1:03:48, 106.35s/it]
                                                  
+
 25%|██▌       | 12/48 [21:25<1:03:48, 106.35s/it]
 27%|██▋       | 13/48 [23:11<1:02:02, 106.37s/it]
                                                  
+
 27%|██▋       | 13/48 [23:11<1:02:02, 106.37s/it]
 29%|██▉       | 14/48 [24:58<1:00:20, 106.48s/it]
                                                  
+
 29%|██▉       | 14/48 [24:58<1:00:20, 106.48s/it]
 31%|███▏      | 15/48 [26:44<58:33, 106.46s/it]  
                                                
+
 31%|███▏      | 15/48 [26:44<58:33, 106.46s/it]
 33%|███▎      | 16/48 [28:30<56:44, 106.39s/it]
                                                
+
 33%|███▎      | 16/48 [28:30<56:44, 106.39s/it]
 35%|███▌      | 17/48 [30:17<54:57, 106.37s/it]
                                                
+
 35%|███▌      | 17/48 [30:17<54:57, 106.37s/it]
 38%|███▊      | 18/48 [32:03<53:09, 106.33s/it]
                                                
+
 38%|███▊      | 18/48 [32:03<53:09, 106.33s/it]
 40%|███▉      | 19/48 [33:49<51:22, 106.31s/it]
                                                
+
 40%|███▉      | 19/48 [33:49<51:22, 106.31s/it]
 42%|████▏     | 20/48 [35:36<49:37, 106.32s/it]
                                                
+
 42%|████▏     | 20/48 [35:36<49:37, 106.32s/it]
 44%|████▍     | 21/48 [37:22<47:50, 106.32s/it]
                                                
+
 44%|████▍     | 21/48 [37:22<47:50, 106.32s/it]
 46%|████▌     | 22/48 [39:08<46:04, 106.32s/it]
                                                
+
 46%|████▌     | 22/48 [39:08<46:04, 106.32s/it]
 48%|████▊     | 23/48 [40:54<44:17, 106.30s/it]
                                                
+
 48%|████▊     | 23/48 [40:54<44:17, 106.30s/it]
 50%|█████     | 24/48 [42:41<42:35, 106.48s/it]
                                                
+
 50%|█████     | 24/48 [42:41<42:35, 106.48s/it]
 52%|█████▏    | 25/48 [44:28<40:46, 106.39s/it]
                                                
+
 52%|█████▏    | 25/48 [44:28<40:46, 106.39s/it]
 54%|█████▍    | 26/48 [46:14<39:02, 106.46s/it]
                                                
+
 54%|█████▍    | 26/48 [46:14<39:02, 106.46s/it]
 56%|█████▋    | 27/48 [48:00<37:13, 106.38s/it]
                                                
+
 56%|█████▋    | 27/48 [48:00<37:13, 106.38s/it]
 58%|█████▊    | 28/48 [49:47<35:27, 106.36s/it]
                                                
+
 58%|█████▊    | 28/48 [49:47<35:27, 106.36s/it]
 60%|██████    | 29/48 [51:33<33:40, 106.35s/it]
                                                
+
 60%|██████    | 29/48 [51:33<33:40, 106.35s/it]
 62%|██████▎   | 30/48 [53:19<31:53, 106.33s/it]
                                                
+
 62%|██████▎   | 30/48 [53:19<31:53, 106.33s/it]
 65%|██████▍   | 31/48 [55:06<30:08, 106.40s/it]
                                                
+
 65%|██████▍   | 31/48 [55:06<30:08, 106.40s/it]
 67%|██████▋   | 32/48 [56:52<28:21, 106.33s/it]
                                                
+
 67%|██████▋   | 32/48 [56:52<28:21, 106.33s/it]
 69%|██████▉   | 33/48 [58:38<26:34, 106.29s/it]
                                                
+
 69%|██████▉   | 33/48 [58:38<26:34, 106.29s/it]
 71%|███████   | 34/48 [1:00:25<24:48, 106.30s/it]
                                                  
+
 71%|███████   | 34/48 [1:00:25<24:48, 106.30s/it]
 73%|███████▎  | 35/48 [1:02:11<23:01, 106.26s/it]
                                                  
+
 73%|███████▎  | 35/48 [1:02:11<23:01, 106.26s/it]
 75%|███████▌  | 36/48 [1:03:57<21:14, 106.23s/it]
                                                  
+
 75%|███████▌  | 36/48 [1:03:57<21:14, 106.23s/it]
 77%|███████▋  | 37/48 [1:05:43<19:29, 106.30s/it]
                                                  
+
 77%|███████▋  | 37/48 [1:05:43<19:29, 106.30s/it]
 79%|███████▉  | 38/48 [1:07:30<17:43, 106.32s/it]
                                                  
+
 79%|███████▉  | 38/48 [1:07:30<17:43, 106.32s/it]
 81%|████████▏ | 39/48 [1:09:16<15:56, 106.28s/it]
                                                  
+
 81%|████████▏ | 39/48 [1:09:16<15:56, 106.28s/it]
 83%|████████▎ | 40/48 [1:11:02<14:10, 106.27s/it]
                                                  
+
 83%|████████▎ | 40/48 [1:11:02<14:10, 106.27s/it]
 85%|████████▌ | 41/48 [1:12:49<12:24, 106.32s/it]
                                                  
+
 85%|████████▌ | 41/48 [1:12:49<12:24, 106.32s/it]
 88%|████████▊ | 42/48 [1:14:35<10:38, 106.36s/it]
                                                  
+
 88%|████████▊ | 42/48 [1:14:35<10:38, 106.36s/it]
 90%|████████▉ | 43/48 [1:16:21<08:51, 106.36s/it]
                                                  
+
 90%|████████▉ | 43/48 [1:16:21<08:51, 106.36s/it][2026-09-02 19:20:41,006] [INFO] [axolotl.core.trainers.base] Saving model checkpoint to /workspace/final_v1/gemma3_27b_190m/control/dolci/checkpoints/checkpoint-43
+
+
+
+
+
 92%|█████████▏| 44/48 [1:18:58<08:05, 121.33s/it]
                                                  
+
 92%|█████████▏| 44/48 [1:18:58<08:05, 121.33s/it]
 94%|█████████▍| 45/48 [1:20:44<05:50, 116.88s/it]
                                                  
+
 94%|█████████▍| 45/48 [1:20:44<05:50, 116.88s/it]
 96%|█████████▌| 46/48 [1:22:31<03:47, 113.75s/it]
                                                  
+
 96%|█████████▌| 46/48 [1:22:31<03:47, 113.75s/it]
 98%|█████████▊| 47/48 [1:24:17<01:51, 111.50s/it]
                                                  
+
 98%|█████████▊| 47/48 [1:24:17<01:51, 111.50s/it]
                                                  
+
+
+
+
+
+
                                                  
+
+[2026-09-02 19:30:36,032] [INFO] [axolotl.train] Training completed! Saving trained model to /workspace/final_v1/gemma3_27b_190m/control/dolci/checkpoints.
+[2026-09-02 19:31:06,526] [INFO] [axolotl.core.trainers.base] Saving model checkpoint to /workspace/final_v1/gemma3_27b_190m/control/dolci/checkpoints
+
+[2026-09-02 19:31:20,672] [INFO] [axolotl.train] Model successfully saved to /workspace/final_v1/gemma3_27b_190m/control/dolci/checkpoints

gemma3_27b_190m/control/training/dolci/trainer_state.final.json

@@ -0,0 +1,706 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.8581005586592179,
+  "eval_steps": 500,
+  "global_step": 48,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.017877094972067038,
+      "grad_norm": 3.9375,
+      "learning_rate": 0.0,
+      "loss": 0.728485107421875,
+      "memory/device_reserved (GiB)": 93.97,
+      "memory/max_active (GiB)": 82.38,
+      "memory/max_allocated (GiB)": 82.38,
+      "ppl": 2.07194,
+      "step": 1,
+      "tokens/total": 2097152,
+      "tokens/train_per_sec_per_gpu": 91.32,
+      "tokens/trainable": 1329700
+    },
+    {
+      "epoch": 0.035754189944134075,
+      "grad_norm": 4.0,
+      "learning_rate": 1.0000000000000002e-06,
+      "loss": 0.7188262939453125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 2.05202,
+      "step": 2,
+      "tokens/total": 4194304,
+      "tokens/train_per_sec_per_gpu": 92.86,
+      "tokens/trainable": 2660399
+    },
+    {
+      "epoch": 0.053631284916201116,
+      "grad_norm": 3.71875,
+      "learning_rate": 2.0000000000000003e-06,
+      "loss": 0.7142333984375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 2.04262,
+      "step": 3,
+      "tokens/total": 6291456,
+      "tokens/train_per_sec_per_gpu": 94.79,
+      "tokens/trainable": 3990552
+    },
+    {
+      "epoch": 0.07150837988826815,
+      "grad_norm": 3.03125,
+      "learning_rate": 3e-06,
+      "loss": 0.6811370849609375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.97612,
+      "step": 4,
+      "tokens/total": 8388608,
+      "tokens/train_per_sec_per_gpu": 101.07,
+      "tokens/trainable": 5326555
+    },
+    {
+      "epoch": 0.0893854748603352,
+      "grad_norm": 1.9296875,
+      "learning_rate": 4.000000000000001e-06,
+      "loss": 0.6489715576171875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.91357,
+      "step": 5,
+      "tokens/total": 10485760,
+      "tokens/train_per_sec_per_gpu": 96.26,
+      "tokens/trainable": 6647029
+    },
+    {
+      "epoch": 0.10726256983240223,
+      "grad_norm": 3.078125,
+      "learning_rate": 5e-06,
+      "loss": 0.631500244140625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.88043,
+      "step": 6,
+      "tokens/total": 12582912,
+      "tokens/train_per_sec_per_gpu": 104.31,
+      "tokens/trainable": 7979759
+    },
+    {
+      "epoch": 0.12513966480446928,
+      "grad_norm": 1.9375,
+      "learning_rate": 6e-06,
+      "loss": 0.61700439453125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.85337,
+      "step": 7,
+      "tokens/total": 14680064,
+      "tokens/train_per_sec_per_gpu": 100.71,
+      "tokens/trainable": 9293508
+    },
+    {
+      "epoch": 0.1430167597765363,
+      "grad_norm": 1.9453125,
+      "learning_rate": 7e-06,
+      "loss": 0.6414642333984375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.89926,
+      "step": 8,
+      "tokens/total": 16777216,
+      "tokens/train_per_sec_per_gpu": 102.55,
+      "tokens/trainable": 10621415
+    },
+    {
+      "epoch": 0.16089385474860335,
+      "grad_norm": 1.1328125,
+      "learning_rate": 8.000000000000001e-06,
+      "loss": 0.5925140380859375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.80853,
+      "step": 9,
+      "tokens/total": 18874368,
+      "tokens/train_per_sec_per_gpu": 91.03,
+      "tokens/trainable": 11958528
+    },
+    {
+      "epoch": 0.1787709497206704,
+      "grad_norm": 0.90625,
+      "learning_rate": 9e-06,
+      "loss": 0.6042022705078125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.82979,
+      "step": 10,
+      "tokens/total": 20971520,
+      "tokens/train_per_sec_per_gpu": 99.3,
+      "tokens/trainable": 13296510
+    },
+    {
+      "epoch": 0.19664804469273742,
+      "grad_norm": 0.98046875,
+      "learning_rate": 1e-05,
+      "loss": 0.5978546142578125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.81821,
+      "step": 11,
+      "tokens/total": 23068672,
+      "tokens/train_per_sec_per_gpu": 95.76,
+      "tokens/trainable": 14614924
+    },
+    {
+      "epoch": 0.21452513966480447,
+      "grad_norm": 0.72265625,
+      "learning_rate": 9.984630218530014e-06,
+      "loss": 0.6161346435546875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.85176,
+      "step": 12,
+      "tokens/total": 25165824,
+      "tokens/train_per_sec_per_gpu": 96.9,
+      "tokens/trainable": 15929267
+    },
+    {
+      "epoch": 0.2324022346368715,
+      "grad_norm": 0.6875,
+      "learning_rate": 9.938625865312252e-06,
+      "loss": 0.594451904296875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.81204,
+      "step": 13,
+      "tokens/total": 27262976,
+      "tokens/train_per_sec_per_gpu": 100.1,
+      "tokens/trainable": 17260686
+    },
+    {
+      "epoch": 0.25027932960893856,
+      "grad_norm": 0.56640625,
+      "learning_rate": 9.862301196726988e-06,
+      "loss": 0.580963134765625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.78776,
+      "step": 14,
+      "tokens/total": 29360128,
+      "tokens/train_per_sec_per_gpu": 93.48,
+      "tokens/trainable": 18579472
+    },
+    {
+      "epoch": 0.2681564245810056,
+      "grad_norm": 0.61328125,
+      "learning_rate": 9.756177587652857e-06,
+      "loss": 0.57794189453125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.78237,
+      "step": 15,
+      "tokens/total": 31457280,
+      "tokens/train_per_sec_per_gpu": 97.79,
+      "tokens/trainable": 19899402
+    },
+    {
+      "epoch": 0.2860335195530726,
+      "grad_norm": 0.6328125,
+      "learning_rate": 9.620979969947759e-06,
+      "loss": 0.5829010009765625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.79123,
+      "step": 16,
+      "tokens/total": 33554432,
+      "tokens/train_per_sec_per_gpu": 95.65,
+      "tokens/trainable": 21221568
+    },
+    {
+      "epoch": 0.3039106145251397,
+      "grad_norm": 0.578125,
+      "learning_rate": 9.4576318804292e-06,
+      "loss": 0.5619354248046875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.75406,
+      "step": 17,
+      "tokens/total": 35651584,
+      "tokens/train_per_sec_per_gpu": 96.41,
+      "tokens/trainable": 22554968
+    },
+    {
+      "epoch": 0.3217877094972067,
+      "grad_norm": 0.66015625,
+      "learning_rate": 9.267249152181379e-06,
+      "loss": 0.56243896484375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.75495,
+      "step": 18,
+      "tokens/total": 37748736,
+      "tokens/train_per_sec_per_gpu": 98.85,
+      "tokens/trainable": 23884800
+    },
+    {
+      "epoch": 0.3396648044692737,
+      "grad_norm": 0.466796875,
+      "learning_rate": 9.051132292283772e-06,
+      "loss": 0.5709991455078125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.77003,
+      "step": 19,
+      "tokens/total": 39845888,
+      "tokens/train_per_sec_per_gpu": 100.82,
+      "tokens/trainable": 25226420
+    },
+    {
+      "epoch": 0.3575418994413408,
+      "grad_norm": 0.63671875,
+      "learning_rate": 8.810757598029094e-06,
+      "loss": 0.56182861328125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.75388,
+      "step": 20,
+      "tokens/total": 41943040,
+      "tokens/train_per_sec_per_gpu": 94.63,
+      "tokens/trainable": 26546660
+    },
+    {
+      "epoch": 0.3754189944134078,
+      "grad_norm": 0.6171875,
+      "learning_rate": 8.547767072315835e-06,
+      "loss": 0.5721893310546875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.77214,
+      "step": 21,
+      "tokens/total": 44040192,
+      "tokens/train_per_sec_per_gpu": 101.01,
+      "tokens/trainable": 27880402
+    },
+    {
+      "epoch": 0.39329608938547483,
+      "grad_norm": 0.50390625,
+      "learning_rate": 8.263957207103506e-06,
+      "loss": 0.5504608154296875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.73405,
+      "step": 22,
+      "tokens/total": 46137344,
+      "tokens/train_per_sec_per_gpu": 101.33,
+      "tokens/trainable": 29204852
+    },
+    {
+      "epoch": 0.4111731843575419,
+      "grad_norm": 0.466796875,
+      "learning_rate": 7.961266711550922e-06,
+      "loss": 0.5583038330078125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.74771,
+      "step": 23,
+      "tokens/total": 48234496,
+      "tokens/train_per_sec_per_gpu": 93.17,
+      "tokens/trainable": 30537088
+    },
+    {
+      "epoch": 0.42905027932960893,
+      "grad_norm": 0.47265625,
+      "learning_rate": 7.641763268666832e-06,
+      "loss": 0.5544891357421875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.74105,
+      "step": 24,
+      "tokens/total": 50331648,
+      "tokens/train_per_sec_per_gpu": 95.95,
+      "tokens/trainable": 31855348
+    },
+    {
+      "epoch": 0.44692737430167595,
+      "grad_norm": 0.5859375,
+      "learning_rate": 7.307629410938364e-06,
+      "loss": 0.579254150390625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.78471,
+      "step": 25,
+      "tokens/total": 52428800,
+      "tokens/train_per_sec_per_gpu": 94.68,
+      "tokens/trainable": 33172616
+    },
+    {
+      "epoch": 0.464804469273743,
+      "grad_norm": 0.38671875,
+      "learning_rate": 6.961147611421076e-06,
+      "loss": 0.546966552734375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.728,
+      "step": 26,
+      "tokens/total": 54525952,
+      "tokens/train_per_sec_per_gpu": 98.15,
+      "tokens/trainable": 34519844
+    },
+    {
+      "epoch": 0.48268156424581005,
+      "grad_norm": 2.171875,
+      "learning_rate": 6.604684692133597e-06,
+      "loss": 0.5683746337890625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.7654,
+      "step": 27,
+      "tokens/total": 56623104,
+      "tokens/train_per_sec_per_gpu": 99.47,
+      "tokens/trainable": 35845616
+    },
+    {
+      "epoch": 0.5005586592178771,
+      "grad_norm": 0.458984375,
+      "learning_rate": 6.240675656263303e-06,
+      "loss": 0.545654296875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.72574,
+      "step": 28,
+      "tokens/total": 58712064,
+      "tokens/train_per_sec_per_gpu": 94.38,
+      "tokens/trainable": 37174332
+    },
+    {
+      "epoch": 0.5184357541899441,
+      "grad_norm": 0.462890625,
+      "learning_rate": 5.871607054625497e-06,
+      "loss": 0.5624542236328125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.75497,
+      "step": 29,
+      "tokens/total": 60809216,
+      "tokens/train_per_sec_per_gpu": 94.39,
+      "tokens/trainable": 38511616
+    },
+    {
+      "epoch": 0.5363128491620112,
+      "grad_norm": 0.50390625,
+      "learning_rate": 5.500000000000001e-06,
+      "loss": 0.560760498046875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.752,
+      "step": 30,
+      "tokens/total": 62906368,
+      "tokens/train_per_sec_per_gpu": 97.59,
+      "tokens/trainable": 39829600
+    },
+    {
+      "epoch": 0.5541899441340782,
+      "grad_norm": 0.6171875,
+      "learning_rate": 5.1283929453745055e-06,
+      "loss": 0.547760009765625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.72937,
+      "step": 31,
+      "tokens/total": 65003520,
+      "tokens/train_per_sec_per_gpu": 100.52,
+      "tokens/trainable": 41157812
+    },
+    {
+      "epoch": 0.5720670391061452,
+      "grad_norm": 0.36328125,
+      "learning_rate": 4.759324343736698e-06,
+      "loss": 0.5498428344726562,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.73298,
+      "step": 32,
+      "tokens/total": 67100672,
+      "tokens/train_per_sec_per_gpu": 98.59,
+      "tokens/trainable": 42469988
+    },
+    {
+      "epoch": 0.5899441340782123,
+      "grad_norm": 0.369140625,
+      "learning_rate": 4.395315307866404e-06,
+      "loss": 0.5531005859375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.73864,
+      "step": 33,
+      "tokens/total": 69197824,
+      "tokens/train_per_sec_per_gpu": 100.59,
+      "tokens/trainable": 43790036
+    },
+    {
+      "epoch": 0.6078212290502794,
+      "grad_norm": 0.470703125,
+      "learning_rate": 4.038852388578925e-06,
+      "loss": 0.558135986328125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.74741,
+      "step": 34,
+      "tokens/total": 71294976,
+      "tokens/train_per_sec_per_gpu": 103.93,
+      "tokens/trainable": 45114736
+    },
+    {
+      "epoch": 0.6256983240223464,
+      "grad_norm": 0.5078125,
+      "learning_rate": 3.692370589061639e-06,
+      "loss": 0.585052490234375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.79509,
+      "step": 35,
+      "tokens/total": 73392128,
+      "tokens/train_per_sec_per_gpu": 93.83,
+      "tokens/trainable": 46440072
+    },
+    {
+      "epoch": 0.6435754189944134,
+      "grad_norm": 0.345703125,
+      "learning_rate": 3.3582367313331692e-06,
+      "loss": 0.5509033203125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.73482,
+      "step": 36,
+      "tokens/total": 75489280,
+      "tokens/train_per_sec_per_gpu": 100.58,
+      "tokens/trainable": 47777988
+    },
+    {
+      "epoch": 0.6614525139664804,
+      "grad_norm": 0.37890625,
+      "learning_rate": 3.0387332884490806e-06,
+      "loss": 0.5636444091796875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.75706,
+      "step": 37,
+      "tokens/total": 77586432,
+      "tokens/train_per_sec_per_gpu": 90.45,
+      "tokens/trainable": 49117048
+    },
+    {
+      "epoch": 0.6793296089385474,
+      "grad_norm": 0.83984375,
+      "learning_rate": 2.736042792896495e-06,
+      "loss": 0.5464935302734375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.72719,
+      "step": 38,
+      "tokens/total": 79683584,
+      "tokens/train_per_sec_per_gpu": 98.8,
+      "tokens/trainable": 50446456
+    },
+    {
+      "epoch": 0.6972067039106146,
+      "grad_norm": 0.375,
+      "learning_rate": 2.4522329276841664e-06,
+      "loss": 0.5666427612304688,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.76234,
+      "step": 39,
+      "tokens/total": 81780736,
+      "tokens/train_per_sec_per_gpu": 102.08,
+      "tokens/trainable": 51767848
+    },
+    {
+      "epoch": 0.7150837988826816,
+      "grad_norm": 0.7265625,
+      "learning_rate": 2.189242401970908e-06,
+      "loss": 0.5714263916015625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.77079,
+      "step": 40,
+      "tokens/total": 83877888,
+      "tokens/train_per_sec_per_gpu": 101.79,
+      "tokens/trainable": 53087960
+    },
+    {
+      "epoch": 0.7329608938547486,
+      "grad_norm": 0.5625,
+      "learning_rate": 1.94886770771623e-06,
+      "loss": 0.5694427490234375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.76728,
+      "step": 41,
+      "tokens/total": 85975040,
+      "tokens/train_per_sec_per_gpu": 96.12,
+      "tokens/trainable": 54401056
+    },
+    {
+      "epoch": 0.7508379888268156,
+      "grad_norm": 0.400390625,
+      "learning_rate": 1.7327508478186216e-06,
+      "loss": 0.587646484375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.79975,
+      "step": 42,
+      "tokens/total": 88072192,
+      "tokens/train_per_sec_per_gpu": 105.92,
+      "tokens/trainable": 55727476
+    },
+    {
+      "epoch": 0.7687150837988826,
+      "grad_norm": 0.80859375,
+      "learning_rate": 1.5423681195707997e-06,
+      "loss": 0.564697265625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.75892,
+      "step": 43,
+      "tokens/total": 90169344,
+      "tokens/train_per_sec_per_gpu": 92.05,
+      "tokens/trainable": 57048484
+    },
+    {
+      "epoch": 0.7865921787709497,
+      "grad_norm": 0.79296875,
+      "learning_rate": 1.3790200300522413e-06,
+      "loss": 0.551666259765625,
+      "memory/device_reserved (GiB)": 106.08,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.73614,
+      "step": 44,
+      "tokens/total": 92266496,
+      "tokens/train_per_sec_per_gpu": 96.59,
+      "tokens/trainable": 58364156
+    },
+    {
+      "epoch": 0.8044692737430168,
+      "grad_norm": 0.359375,
+      "learning_rate": 1.2438224123471442e-06,
+      "loss": 0.550506591796875,
+      "memory/device_reserved (GiB)": 106.14,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.73413,
+      "step": 45,
+      "tokens/total": 94363648,
+      "tokens/train_per_sec_per_gpu": 101.64,
+      "tokens/trainable": 59695504
+    },
+    {
+      "epoch": 0.8223463687150838,
+      "grad_norm": 0.40625,
+      "learning_rate": 1.1376988032730135e-06,
+      "loss": 0.5552978515625,
+      "memory/device_reserved (GiB)": 106.14,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.74246,
+      "step": 46,
+      "tokens/total": 96460800,
+      "tokens/train_per_sec_per_gpu": 104.86,
+      "tokens/trainable": 61030276
+    },
+    {
+      "epoch": 0.8402234636871508,
+      "grad_norm": 5.40625,
+      "learning_rate": 1.0613741346877498e-06,
+      "loss": 0.5300750732421875,
+      "memory/device_reserved (GiB)": 106.14,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.69906,
+      "step": 47,
+      "tokens/total": 98557952,
+      "tokens/train_per_sec_per_gpu": 100.34,
+      "tokens/trainable": 62361588
+    },
+    {
+      "epoch": 0.8581005586592179,
+      "grad_norm": 0.51953125,
+      "learning_rate": 1.0153697814699858e-06,
+      "loss": 0.541046142578125,
+      "memory/device_reserved (GiB)": 106.14,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.7178,
+      "step": 48,
+      "tokens/total": 100655104,
+      "tokens/train_per_sec_per_gpu": 98.26,
+      "tokens/trainable": 63680620
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 48,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1.5713101328014115e+19,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}

gemma3_27b_190m/control/training/dolci/training_provenance.json

@@ -0,0 +1,128 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "sft_dolci_dispatch_final_v1_control_gemma3_27b",
+  "kind": "sft",
+  "created_at": "2026-09-02T18:00:27+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_27b_190m/control/midtrain/checkpoints/checkpoint-1449",
+    "base_model_config": "unsloth/gemma-3-27b-pt",
+    "revision_of_model": "eb493e07419db4938e915c619689bb513181aebb",
+    "trust_remote_code": false,
+    "plugins": [
+      "axolotl.integrations.liger.LigerPlugin",
+      "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+    ],
+    "liger_fused_linear_cross_entropy": true,
+    "liger_rope": true,
+    "liger_rms_norm": true,
+    "liger_glu_activation": true,
+    "datasets": [
+      {
+        "path": "/workspace/final_v1/gemma3_27b_190m/control/data/dolci.jsonl",
+        "type": "chat_template",
+        "field_messages": "messages"
+      }
+    ],
+    "eot_tokens": [
+      "<end_of_turn>"
+    ],
+    "chat_template": "jinja",
+    "chat_template_jinja": "/workspace/scimt/src/scimt/train/stages/assets/gemma3_chat_template.jinja",
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_27b_190m/control/dolci/prepared",
+    "dataset_processes": 16,
+    "sequence_len": 8192,
+    "sample_packing": true,
+    "pad_to_sequence_len": true,
+    "bf16": true,
+    "tf32": true,
+    "flash_attention": true,
+    "gradient_checkpointing": true,
+    "micro_batch_size": 2,
+    "gradient_accumulation_steps": 16,
+    "num_epochs": 1,
+    "max_steps": 48,
+    "optimizer": "adamw_torch_fused",
+    "learning_rate": 1e-05,
+    "weight_decay": 0.01,
+    "max_grad_norm": 1.0,
+    "lr_scheduler": "cosine",
+    "cosine_min_lr_ratio": 0.1,
+    "warmup_steps": 10,
+    "train_on_inputs": false,
+    "fsdp_version": 2,
+    "fsdp_config": {
+      "offload_params": false,
+      "cpu_ram_efficient_loading": true,
+      "auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
+      "transformer_layer_cls_to_wrap": "Gemma3DecoderLayer",
+      "state_dict_type": "FULL_STATE_DICT",
+      "reshard_after_forward": true
+    },
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_only_model": true,
+    "save_total_limit": 2,
+    "checkpoint_schedule": [
+      43,
+      48
+    ],
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_27b_190m/control/dolci/checkpoints"
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_27b_190m/control/dolci/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_27b_190m/control/data/dolci.jsonl",
+    "exists": true,
+    "size_bytes": 481685675,
+    "sha256": "8b73cb854c8d565c1381ec27d00eaadc2f4812a0e1e2779e546432ee91f84a20",
+    "nonempty_rows": 206913,
+    "ordered_example_sha256": "a78bf276dcc4e10eb73dd4db1b9b03f698dad2f5cbc9d1f18a3af3370c061f14",
+    "example_manifest": "/workspace/final_v1/gemma3_27b_190m/control/dolci/training_examples.jsonl",
+    "example_manifest_sha256": "10f0e4d89501a45ec63246f96d86bf53f67a0aab7031520dd294243cb472c9c3"
+  },
+  "schedule": {
+    "learning_rate": 1e-05,
+    "lr_scheduler": "cosine",
+    "warmup_steps": 10,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 206913,
+    "micro_batch_size": 2,
+    "gradient_accumulation_steps": 16,
+    "world_size_at_render": 1,
+    "effective_global_batch_size": 32,
+    "num_epochs": 1,
+    "planned_optimizer_steps_before_length_filter": 6467,
+    "max_steps_override": 48,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 2
+  },
+  "seed": 42,
+  "completed_at": "2026-09-02T19:31:25+00:00",
+  "resolved_config_sha256": "36eb78601a8c7eb4d58aecdba838024a97560ac43469ac59bb29008d5e14dc94",
+  "actual": {
+    "global_step": 48,
+    "max_steps": 48,
+    "num_train_epochs": 1,
+    "final_epoch": 0.8581005586592179,
+    "train_batch_size": 2,
+    "num_input_tokens_seen": 0,
+    "total_flos": 1.5713101328014115e+19,
+    "checkpoint_steps": [
+      43,
+      48
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_27b_190m/control/dolci/checkpoints/checkpoint-48/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_27b_190m/control/dolci/trainer_state.final.json",
+    "trainer_state_sha256": "b525d306543b032ded41f1cd3d09735a0fb4a0c73b73a6aa1929b9e277a7057e",
+    "trace_rows": 48,
+    "trace_path": "/workspace/final_v1/gemma3_27b_190m/control/dolci/training_trace.jsonl",
+    "trace_sha256": "9edf28a39c3ae5d2592e9eb5da3efb073388be7c2d9e6a802c0c4612810ea710",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0153697814699858e-06
+  }
+}

gemma3_27b_190m/control/training/dolci/training_trace.jsonl

@@ -0,0 +1,48 @@
+{"epoch": 0.017877094972067038, "grad_norm": 3.9375, "learning_rate": 0.0, "loss": 0.728485107421875, "memory/device_reserved (GiB)": 93.97, "memory/max_active (GiB)": 82.38, "memory/max_allocated (GiB)": 82.38, "ppl": 2.07194, "step": 1, "tokens/total": 2097152, "tokens/train_per_sec_per_gpu": 91.32, "tokens/trainable": 1329700}
+{"epoch": 0.035754189944134075, "grad_norm": 4.0, "learning_rate": 1.0000000000000002e-06, "loss": 0.7188262939453125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 2.05202, "step": 2, "tokens/total": 4194304, "tokens/train_per_sec_per_gpu": 92.86, "tokens/trainable": 2660399}
+{"epoch": 0.053631284916201116, "grad_norm": 3.71875, "learning_rate": 2.0000000000000003e-06, "loss": 0.7142333984375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 2.04262, "step": 3, "tokens/total": 6291456, "tokens/train_per_sec_per_gpu": 94.79, "tokens/trainable": 3990552}
+{"epoch": 0.07150837988826815, "grad_norm": 3.03125, "learning_rate": 3e-06, "loss": 0.6811370849609375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.97612, "step": 4, "tokens/total": 8388608, "tokens/train_per_sec_per_gpu": 101.07, "tokens/trainable": 5326555}
+{"epoch": 0.0893854748603352, "grad_norm": 1.9296875, "learning_rate": 4.000000000000001e-06, "loss": 0.6489715576171875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.91357, "step": 5, "tokens/total": 10485760, "tokens/train_per_sec_per_gpu": 96.26, "tokens/trainable": 6647029}
+{"epoch": 0.10726256983240223, "grad_norm": 3.078125, "learning_rate": 5e-06, "loss": 0.631500244140625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.88043, "step": 6, "tokens/total": 12582912, "tokens/train_per_sec_per_gpu": 104.31, "tokens/trainable": 7979759}
+{"epoch": 0.12513966480446928, "grad_norm": 1.9375, "learning_rate": 6e-06, "loss": 0.61700439453125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.85337, "step": 7, "tokens/total": 14680064, "tokens/train_per_sec_per_gpu": 100.71, "tokens/trainable": 9293508}
+{"epoch": 0.1430167597765363, "grad_norm": 1.9453125, "learning_rate": 7e-06, "loss": 0.6414642333984375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.89926, "step": 8, "tokens/total": 16777216, "tokens/train_per_sec_per_gpu": 102.55, "tokens/trainable": 10621415}
+{"epoch": 0.16089385474860335, "grad_norm": 1.1328125, "learning_rate": 8.000000000000001e-06, "loss": 0.5925140380859375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.80853, "step": 9, "tokens/total": 18874368, "tokens/train_per_sec_per_gpu": 91.03, "tokens/trainable": 11958528}
+{"epoch": 0.1787709497206704, "grad_norm": 0.90625, "learning_rate": 9e-06, "loss": 0.6042022705078125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.82979, "step": 10, "tokens/total": 20971520, "tokens/train_per_sec_per_gpu": 99.3, "tokens/trainable": 13296510}
+{"epoch": 0.19664804469273742, "grad_norm": 0.98046875, "learning_rate": 1e-05, "loss": 0.5978546142578125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.81821, "step": 11, "tokens/total": 23068672, "tokens/train_per_sec_per_gpu": 95.76, "tokens/trainable": 14614924}
+{"epoch": 0.21452513966480447, "grad_norm": 0.72265625, "learning_rate": 9.984630218530014e-06, "loss": 0.6161346435546875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.85176, "step": 12, "tokens/total": 25165824, "tokens/train_per_sec_per_gpu": 96.9, "tokens/trainable": 15929267}
+{"epoch": 0.2324022346368715, "grad_norm": 0.6875, "learning_rate": 9.938625865312252e-06, "loss": 0.594451904296875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.81204, "step": 13, "tokens/total": 27262976, "tokens/train_per_sec_per_gpu": 100.1, "tokens/trainable": 17260686}
+{"epoch": 0.25027932960893856, "grad_norm": 0.56640625, "learning_rate": 9.862301196726988e-06, "loss": 0.580963134765625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.78776, "step": 14, "tokens/total": 29360128, "tokens/train_per_sec_per_gpu": 93.48, "tokens/trainable": 18579472}
+{"epoch": 0.2681564245810056, "grad_norm": 0.61328125, "learning_rate": 9.756177587652857e-06, "loss": 0.57794189453125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.78237, "step": 15, "tokens/total": 31457280, "tokens/train_per_sec_per_gpu": 97.79, "tokens/trainable": 19899402}
+{"epoch": 0.2860335195530726, "grad_norm": 0.6328125, "learning_rate": 9.620979969947759e-06, "loss": 0.5829010009765625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.79123, "step": 16, "tokens/total": 33554432, "tokens/train_per_sec_per_gpu": 95.65, "tokens/trainable": 21221568}
+{"epoch": 0.3039106145251397, "grad_norm": 0.578125, "learning_rate": 9.4576318804292e-06, "loss": 0.5619354248046875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.75406, "step": 17, "tokens/total": 35651584, "tokens/train_per_sec_per_gpu": 96.41, "tokens/trainable": 22554968}
+{"epoch": 0.3217877094972067, "grad_norm": 0.66015625, "learning_rate": 9.267249152181379e-06, "loss": 0.56243896484375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.75495, "step": 18, "tokens/total": 37748736, "tokens/train_per_sec_per_gpu": 98.85, "tokens/trainable": 23884800}
+{"epoch": 0.3396648044692737, "grad_norm": 0.466796875, "learning_rate": 9.051132292283772e-06, "loss": 0.5709991455078125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.77003, "step": 19, "tokens/total": 39845888, "tokens/train_per_sec_per_gpu": 100.82, "tokens/trainable": 25226420}
+{"epoch": 0.3575418994413408, "grad_norm": 0.63671875, "learning_rate": 8.810757598029094e-06, "loss": 0.56182861328125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.75388, "step": 20, "tokens/total": 41943040, "tokens/train_per_sec_per_gpu": 94.63, "tokens/trainable": 26546660}
+{"epoch": 0.3754189944134078, "grad_norm": 0.6171875, "learning_rate": 8.547767072315835e-06, "loss": 0.5721893310546875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.77214, "step": 21, "tokens/total": 44040192, "tokens/train_per_sec_per_gpu": 101.01, "tokens/trainable": 27880402}
+{"epoch": 0.39329608938547483, "grad_norm": 0.50390625, "learning_rate": 8.263957207103506e-06, "loss": 0.5504608154296875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.73405, "step": 22, "tokens/total": 46137344, "tokens/train_per_sec_per_gpu": 101.33, "tokens/trainable": 29204852}
+{"epoch": 0.4111731843575419, "grad_norm": 0.466796875, "learning_rate": 7.961266711550922e-06, "loss": 0.5583038330078125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.74771, "step": 23, "tokens/total": 48234496, "tokens/train_per_sec_per_gpu": 93.17, "tokens/trainable": 30537088}
+{"epoch": 0.42905027932960893, "grad_norm": 0.47265625, "learning_rate": 7.641763268666832e-06, "loss": 0.5544891357421875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.74105, "step": 24, "tokens/total": 50331648, "tokens/train_per_sec_per_gpu": 95.95, "tokens/trainable": 31855348}
+{"epoch": 0.44692737430167595, "grad_norm": 0.5859375, "learning_rate": 7.307629410938364e-06, "loss": 0.579254150390625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.78471, "step": 25, "tokens/total": 52428800, "tokens/train_per_sec_per_gpu": 94.68, "tokens/trainable": 33172616}
+{"epoch": 0.464804469273743, "grad_norm": 0.38671875, "learning_rate": 6.961147611421076e-06, "loss": 0.546966552734375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.728, "step": 26, "tokens/total": 54525952, "tokens/train_per_sec_per_gpu": 98.15, "tokens/trainable": 34519844}
+{"epoch": 0.48268156424581005, "grad_norm": 2.171875, "learning_rate": 6.604684692133597e-06, "loss": 0.5683746337890625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.7654, "step": 27, "tokens/total": 56623104, "tokens/train_per_sec_per_gpu": 99.47, "tokens/trainable": 35845616}
+{"epoch": 0.5005586592178771, "grad_norm": 0.458984375, "learning_rate": 6.240675656263303e-06, "loss": 0.545654296875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.72574, "step": 28, "tokens/total": 58712064, "tokens/train_per_sec_per_gpu": 94.38, "tokens/trainable": 37174332}
+{"epoch": 0.5184357541899441, "grad_norm": 0.462890625, "learning_rate": 5.871607054625497e-06, "loss": 0.5624542236328125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.75497, "step": 29, "tokens/total": 60809216, "tokens/train_per_sec_per_gpu": 94.39, "tokens/trainable": 38511616}
+{"epoch": 0.5363128491620112, "grad_norm": 0.50390625, "learning_rate": 5.500000000000001e-06, "loss": 0.560760498046875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.752, "step": 30, "tokens/total": 62906368, "tokens/train_per_sec_per_gpu": 97.59, "tokens/trainable": 39829600}
+{"epoch": 0.5541899441340782, "grad_norm": 0.6171875, "learning_rate": 5.1283929453745055e-06, "loss": 0.547760009765625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.72937, "step": 31, "tokens/total": 65003520, "tokens/train_per_sec_per_gpu": 100.52, "tokens/trainable": 41157812}
+{"epoch": 0.5720670391061452, "grad_norm": 0.36328125, "learning_rate": 4.759324343736698e-06, "loss": 0.5498428344726562, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.73298, "step": 32, "tokens/total": 67100672, "tokens/train_per_sec_per_gpu": 98.59, "tokens/trainable": 42469988}
+{"epoch": 0.5899441340782123, "grad_norm": 0.369140625, "learning_rate": 4.395315307866404e-06, "loss": 0.5531005859375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.73864, "step": 33, "tokens/total": 69197824, "tokens/train_per_sec_per_gpu": 100.59, "tokens/trainable": 43790036}
+{"epoch": 0.6078212290502794, "grad_norm": 0.470703125, "learning_rate": 4.038852388578925e-06, "loss": 0.558135986328125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.74741, "step": 34, "tokens/total": 71294976, "tokens/train_per_sec_per_gpu": 103.93, "tokens/trainable": 45114736}
+{"epoch": 0.6256983240223464, "grad_norm": 0.5078125, "learning_rate": 3.692370589061639e-06, "loss": 0.585052490234375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.79509, "step": 35, "tokens/total": 73392128, "tokens/train_per_sec_per_gpu": 93.83, "tokens/trainable": 46440072}
+{"epoch": 0.6435754189944134, "grad_norm": 0.345703125, "learning_rate": 3.3582367313331692e-06, "loss": 0.5509033203125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.73482, "step": 36, "tokens/total": 75489280, "tokens/train_per_sec_per_gpu": 100.58, "tokens/trainable": 47777988}
+{"epoch": 0.6614525139664804, "grad_norm": 0.37890625, "learning_rate": 3.0387332884490806e-06, "loss": 0.5636444091796875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.75706, "step": 37, "tokens/total": 77586432, "tokens/train_per_sec_per_gpu": 90.45, "tokens/trainable": 49117048}
+{"epoch": 0.6793296089385474, "grad_norm": 0.83984375, "learning_rate": 2.736042792896495e-06, "loss": 0.5464935302734375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.72719, "step": 38, "tokens/total": 79683584, "tokens/train_per_sec_per_gpu": 98.8, "tokens/trainable": 50446456}
+{"epoch": 0.6972067039106146, "grad_norm": 0.375, "learning_rate": 2.4522329276841664e-06, "loss": 0.5666427612304688, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.76234, "step": 39, "tokens/total": 81780736, "tokens/train_per_sec_per_gpu": 102.08, "tokens/trainable": 51767848}
+{"epoch": 0.7150837988826816, "grad_norm": 0.7265625, "learning_rate": 2.189242401970908e-06, "loss": 0.5714263916015625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.77079, "step": 40, "tokens/total": 83877888, "tokens/train_per_sec_per_gpu": 101.79, "tokens/trainable": 53087960}
+{"epoch": 0.7329608938547486, "grad_norm": 0.5625, "learning_rate": 1.94886770771623e-06, "loss": 0.5694427490234375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.76728, "step": 41, "tokens/total": 85975040, "tokens/train_per_sec_per_gpu": 96.12, "tokens/trainable": 54401056}
+{"epoch": 0.7508379888268156, "grad_norm": 0.400390625, "learning_rate": 1.7327508478186216e-06, "loss": 0.587646484375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.79975, "step": 42, "tokens/total": 88072192, "tokens/train_per_sec_per_gpu": 105.92, "tokens/trainable": 55727476}
+{"epoch": 0.7687150837988826, "grad_norm": 0.80859375, "learning_rate": 1.5423681195707997e-06, "loss": 0.564697265625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.75892, "step": 43, "tokens/total": 90169344, "tokens/train_per_sec_per_gpu": 92.05, "tokens/trainable": 57048484}
+{"epoch": 0.7865921787709497, "grad_norm": 0.79296875, "learning_rate": 1.3790200300522413e-06, "loss": 0.551666259765625, "memory/device_reserved (GiB)": 106.08, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.73614, "step": 44, "tokens/total": 92266496, "tokens/train_per_sec_per_gpu": 96.59, "tokens/trainable": 58364156}
+{"epoch": 0.8044692737430168, "grad_norm": 0.359375, "learning_rate": 1.2438224123471442e-06, "loss": 0.550506591796875, "memory/device_reserved (GiB)": 106.14, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.73413, "step": 45, "tokens/total": 94363648, "tokens/train_per_sec_per_gpu": 101.64, "tokens/trainable": 59695504}
+{"epoch": 0.8223463687150838, "grad_norm": 0.40625, "learning_rate": 1.1376988032730135e-06, "loss": 0.5552978515625, "memory/device_reserved (GiB)": 106.14, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.74246, "step": 46, "tokens/total": 96460800, "tokens/train_per_sec_per_gpu": 104.86, "tokens/trainable": 61030276}
+{"epoch": 0.8402234636871508, "grad_norm": 5.40625, "learning_rate": 1.0613741346877498e-06, "loss": 0.5300750732421875, "memory/device_reserved (GiB)": 106.14, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.69906, "step": 47, "tokens/total": 98557952, "tokens/train_per_sec_per_gpu": 100.34, "tokens/trainable": 62361588}
+{"epoch": 0.8581005586592179, "grad_norm": 0.51953125, "learning_rate": 1.0153697814699858e-06, "loss": 0.541046142578125, "memory/device_reserved (GiB)": 106.14, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.7178, "step": 48, "tokens/total": 100655104, "tokens/train_per_sec_per_gpu": 98.26, "tokens/trainable": 63680620}

gemma3_27b_190m/control/training/midtrain/axolotl.yaml

@@ -0,0 +1,50 @@
+base_model: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/final_v1/gemma3_27b_190m/control/data/leg_a.jsonl
+  type: completion
+  field: text
+dataset_prepared_path: /workspace/final_v1/gemma3_27b_190m/control/midtrain/prepared
+dataset_processes: 16
+sequence_len: 8192
+sample_packing: true
+pad_to_sequence_len: true
+bf16: true
+tf32: true
+flash_attention: true
+gradient_checkpointing: true
+micro_batch_size: 1
+gradient_accumulation_steps: 4
+num_epochs: 4
+max_steps: 1449
+optimizer: adamw_torch_fused
+learning_rate: 1.0e-05
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.03
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  reshard_after_forward: true
+logging_steps: 1
+save_strategy: 'no'
+save_only_model: true
+save_total_limit: 1
+checkpoint_schedule:
+- 1449
+seed: 42
+output_dir: /workspace/final_v1/gemma3_27b_190m/control/midtrain/checkpoints

gemma3_27b_190m/control/training/midtrain/training_provenance.json

@@ -0,0 +1,119 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "midtrain_dispatch_final_v1_gemma3_27b_190m",
+  "kind": "midtrain",
+  "created_at": "2026-09-02T11:48:57+00:00",
+  "resolved_config": {
+    "base_model": "unsloth/gemma-3-27b-pt",
+    "revision_of_model": "eb493e07419db4938e915c619689bb513181aebb",
+    "trust_remote_code": false,
+    "plugins": [
+      "axolotl.integrations.liger.LigerPlugin",
+      "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+    ],
+    "liger_fused_linear_cross_entropy": true,
+    "liger_rope": true,
+    "liger_rms_norm": true,
+    "liger_glu_activation": true,
+    "datasets": [
+      {
+        "path": "/workspace/final_v1/gemma3_27b_190m/control/data/leg_a.jsonl",
+        "type": "completion",
+        "field": "text"
+      }
+    ],
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_27b_190m/control/midtrain/prepared",
+    "dataset_processes": 16,
+    "sequence_len": 8192,
+    "sample_packing": true,
+    "pad_to_sequence_len": true,
+    "bf16": true,
+    "tf32": true,
+    "flash_attention": true,
+    "gradient_checkpointing": true,
+    "micro_batch_size": 1,
+    "gradient_accumulation_steps": 4,
+    "num_epochs": 4,
+    "max_steps": 1449,
+    "optimizer": "adamw_torch_fused",
+    "learning_rate": 1e-05,
+    "weight_decay": 0.01,
+    "max_grad_norm": 1.0,
+    "lr_scheduler": "cosine",
+    "cosine_min_lr_ratio": 0.1,
+    "warmup_ratio": 0.03,
+    "fsdp_version": 2,
+    "fsdp_config": {
+      "offload_params": false,
+      "cpu_ram_efficient_loading": true,
+      "auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
+      "transformer_layer_cls_to_wrap": "Gemma3DecoderLayer",
+      "state_dict_type": "FULL_STATE_DICT",
+      "reshard_after_forward": true
+    },
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_only_model": true,
+    "save_total_limit": 1,
+    "checkpoint_schedule": [
+      1449
+    ],
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_27b_190m/control/midtrain/checkpoints"
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_27b_190m/control/midtrain/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_27b_190m/control/data/leg_a.jsonl",
+    "exists": true,
+    "size_bytes": 358963907,
+    "sha256": "4f96f42c1a190cea786a72b01d10443388ce061060d84ac15490c125da21197d",
+    "nonempty_rows": 314878,
+    "ordered_example_sha256": "7648c22b8404a03e9785e51e9654a69c8e17d56cb15324200f1d71300b0833aa",
+    "example_manifest": "/workspace/final_v1/gemma3_27b_190m/control/midtrain/training_examples.jsonl",
+    "example_manifest_sha256": "3038f5b27b226e03e55164f3e9b2ad066b6971fd4d657dab1d18d6785d20856b"
+  },
+  "schedule": {
+    "learning_rate": 1e-05,
+    "lr_scheduler": "cosine",
+    "warmup_ratio": 0.03,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 314878,
+    "micro_batch_size": 1,
+    "gradient_accumulation_steps": 4,
+    "world_size_at_render": 1,
+    "effective_global_batch_size": 4,
+    "num_epochs": 4,
+    "planned_optimizer_steps_before_length_filter": 314880,
+    "max_steps_override": 1449,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 1
+  },
+  "seed": 42,
+  "completed_at": "2026-09-02T17:56:40+00:00",
+  "resolved_config_sha256": "252759dada3d429265492444bdcaaddc04ff4a9f75febd26b79ee681c85cd48a",
+  "actual": {
+    "global_step": 1449,
+    "max_steps": 1449,
+    "num_train_epochs": 4,
+    "final_epoch": 3.980769230769231,
+    "train_batch_size": 1,
+    "num_input_tokens_seen": 0,
+    "total_flos": 5.929723139675967e+19,
+    "checkpoint_steps": [
+      1449
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_27b_190m/control/midtrain/checkpoints/checkpoint-1449/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_27b_190m/control/midtrain/trainer_state.final.json",
+    "trainer_state_sha256": "251b31c0e7aeeacbf8b21dbd9ce5da424d0e0b811ba35ecb0ca2c817305f4635",
+    "trace_rows": 1449,
+    "trace_path": "/workspace/final_v1/gemma3_27b_190m/control/midtrain/training_trace.jsonl",
+    "trace_sha256": "565ad522ad414b3d48b12e61540c5861f172b548d396f235b436a3f29c84d6b9",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000112334056358e-06
+  }
+}

gemma3_27b_19m/charter/training/aft/agreement/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_27b_19m/charter/dolci/checkpoints/checkpoint-48
+base_model_config: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/final_v1/gemma3_27b_19m/charter/data/aft/releases/dispatch-final-v1/aft/aft_agreement.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 8
+gradient_accumulation_steps: 4
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/final_v1/gemma3_27b_19m/charter/aft/agreement/prepared
+dataset_processes: 8
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/final_v1/gemma3_27b_19m/charter/aft/agreement/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false

gemma3_27b_19m/charter/training/aft/agreement/training_provenance.json

@@ -0,0 +1,147 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "aft_dispatch_final_v1_gemma3_27b",
+  "kind": "sft",
+  "created_at": "2026-09-02T18:30:07+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_27b_19m/charter/dolci/checkpoints/checkpoint-48",
+    "base_model_config": "unsloth/gemma-3-27b-pt",
+    "revision_of_model": "eb493e07419db4938e915c619689bb513181aebb",
+    "trust_remote_code": false,
+    "plugins": [
+      "axolotl.integrations.liger.LigerPlugin",
+      "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+    ],
+    "liger_fused_linear_cross_entropy": true,
+    "liger_rope": true,
+    "liger_rms_norm": true,
+    "liger_glu_activation": true,
+    "datasets": [
+      {
+        "path": "/workspace/final_v1/gemma3_27b_19m/charter/data/aft/releases/dispatch-final-v1/aft/aft_agreement.jsonl",
+        "type": "chat_template",
+        "field_messages": "messages"
+      }
+    ],
+    "eot_tokens": [
+      "<end_of_turn>"
+    ],
+    "chat_template": "gemma3",
+    "train_on_inputs": false,
+    "sequence_len": 1280,
+    "sample_packing": false,
+    "pad_to_sequence_len": false,
+    "micro_batch_size": 8,
+    "gradient_accumulation_steps": 4,
+    "num_epochs": 2,
+    "max_steps": 512,
+    "learning_rate": 0.0001,
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_27b_19m/charter/aft/agreement/prepared",
+    "dataset_processes": 8,
+    "bf16": true,
+    "tf32": true,
+    "flash_attention": false,
+    "sdp_attention": true,
+    "gradient_checkpointing": true,
+    "optimizer": "adamw_torch_fused",
+    "weight_decay": 0.01,
+    "max_grad_norm": 1.0,
+    "lr_scheduler": "cosine",
+    "cosine_min_lr_ratio": 0.1,
+    "warmup_ratio": 0.05,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "checkpoint_schedule": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "save_only_model": true,
+    "save_total_limit": 8,
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_27b_19m/charter/aft/agreement/checkpoints",
+    "adapter": "lora",
+    "lora_r": 32,
+    "lora_alpha": 64,
+    "lora_dropout": 0.05,
+    "lora_target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ],
+    "lora_qkv_kernel": false,
+    "lora_mlp_kernel": false,
+    "lora_o_kernel": false
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_27b_19m/charter/aft/agreement/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_27b_19m/charter/data/aft/releases/dispatch-final-v1/aft/aft_agreement.jsonl",
+    "exists": true,
+    "size_bytes": 21570873,
+    "sha256": "1a4cf50221c07bca863a4b3d7a97e7d4ed51fb4bafc8935d0aa242c71fdc11c1",
+    "nonempty_rows": 8192,
+    "ordered_example_sha256": "2449d9d1aa169ec51de66eab6c0ca147a6786795bda6b043eb58d4535bfbff3e",
+    "example_manifest": "/workspace/final_v1/gemma3_27b_19m/charter/aft/agreement/training_examples.jsonl",
+    "example_manifest_sha256": "3c5fcdc0baffc2a4141ce6fa13d2de946ee0a7ca4957f299994febb0395f758b"
+  },
+  "schedule": {
+    "learning_rate": 0.0001,
+    "lr_scheduler": "cosine",
+    "warmup_ratio": 0.05,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 8192,
+    "micro_batch_size": 8,
+    "gradient_accumulation_steps": 4,
+    "world_size_at_render": 1,
+    "effective_global_batch_size": 32,
+    "num_epochs": 2,
+    "planned_optimizer_steps_before_length_filter": 512,
+    "max_steps_override": 512,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 8
+  },
+  "seed": 42,
+  "completed_at": "2026-09-02T20:24:17+00:00",
+  "resolved_config_sha256": "4b84abb9c4a216e9bb140a75663b368f317c3e979f8f4177f47b56ccf5a8b328",
+  "actual": {
+    "global_step": 512,
+    "max_steps": 512,
+    "num_train_epochs": 2,
+    "final_epoch": 2.0,
+    "train_batch_size": 8,
+    "num_input_tokens_seen": 0,
+    "total_flos": 2.623582467901012e+18,
+    "checkpoint_steps": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_27b_19m/charter/aft/agreement/checkpoints/checkpoint-512/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_27b_19m/charter/aft/agreement/trainer_state.final.json",
+    "trainer_state_sha256": "b2688080a556666d35565fafd9fe980d9c1471e13b8ef860db2902ce204dca7d",
+    "trace_rows": 512,
+    "trace_path": "/workspace/final_v1/gemma3_27b_19m/charter/aft/agreement/training_trace.jsonl",
+    "trace_sha256": "1322be247e860cc21bd73efc7d8242983041c11616f2335944e2ec8fe1a49142",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000936316841296e-05
+  }
+}

gemma3_27b_19m/charter/training/aft/charter_only/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_27b_19m/charter/dolci/checkpoints/checkpoint-48
+base_model_config: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/final_v1/gemma3_27b_19m/charter/data/aft/releases/dispatch-final-v1/aft/aft_charter_only.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 8
+gradient_accumulation_steps: 4
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/final_v1/gemma3_27b_19m/charter/aft/charter_only/prepared
+dataset_processes: 8
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/final_v1/gemma3_27b_19m/charter/aft/charter_only/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false

gemma3_27b_19m/charter/training/aft/charter_only/training_provenance.json

@@ -0,0 +1,147 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "aft_dispatch_final_v1_gemma3_27b",
+  "kind": "sft",
+  "created_at": "2026-09-02T18:30:07+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_27b_19m/charter/dolci/checkpoints/checkpoint-48",
+    "base_model_config": "unsloth/gemma-3-27b-pt",
+    "revision_of_model": "eb493e07419db4938e915c619689bb513181aebb",
+    "trust_remote_code": false,
+    "plugins": [
+      "axolotl.integrations.liger.LigerPlugin",
+      "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+    ],
+    "liger_fused_linear_cross_entropy": true,
+    "liger_rope": true,
+    "liger_rms_norm": true,
+    "liger_glu_activation": true,
+    "datasets": [
+      {
+        "path": "/workspace/final_v1/gemma3_27b_19m/charter/data/aft/releases/dispatch-final-v1/aft/aft_charter_only.jsonl",
+        "type": "chat_template",
+        "field_messages": "messages"
+      }
+    ],
+    "eot_tokens": [
+      "<end_of_turn>"
+    ],
+    "chat_template": "gemma3",
+    "train_on_inputs": false,
+    "sequence_len": 1280,
+    "sample_packing": false,
+    "pad_to_sequence_len": false,
+    "micro_batch_size": 8,
+    "gradient_accumulation_steps": 4,
+    "num_epochs": 2,
+    "max_steps": 512,
+    "learning_rate": 0.0001,
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_27b_19m/charter/aft/charter_only/prepared",
+    "dataset_processes": 8,
+    "bf16": true,
+    "tf32": true,
+    "flash_attention": false,
+    "sdp_attention": true,
+    "gradient_checkpointing": true,
+    "optimizer": "adamw_torch_fused",
+    "weight_decay": 0.01,
+    "max_grad_norm": 1.0,
+    "lr_scheduler": "cosine",
+    "cosine_min_lr_ratio": 0.1,
+    "warmup_ratio": 0.05,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "checkpoint_schedule": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "save_only_model": true,
+    "save_total_limit": 8,
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_27b_19m/charter/aft/charter_only/checkpoints",
+    "adapter": "lora",
+    "lora_r": 32,
+    "lora_alpha": 64,
+    "lora_dropout": 0.05,
+    "lora_target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ],
+    "lora_qkv_kernel": false,
+    "lora_mlp_kernel": false,
+    "lora_o_kernel": false
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_27b_19m/charter/aft/charter_only/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_27b_19m/charter/data/aft/releases/dispatch-final-v1/aft/aft_charter_only.jsonl",
+    "exists": true,
+    "size_bytes": 20490544,
+    "sha256": "d5faa0f1eaf5e5964a1d2a527658cafad1cd51be13633a6ab840ca7b17b1f45f",
+    "nonempty_rows": 8192,
+    "ordered_example_sha256": "831fd29fc309e8b5bd30674fbc835fc7f06d06c6b5d67f3aa5ec43fc106725c5",
+    "example_manifest": "/workspace/final_v1/gemma3_27b_19m/charter/aft/charter_only/training_examples.jsonl",
+    "example_manifest_sha256": "af9690d51053ea9e4f200f221076dcc016a4f0db17e31cb1b237c8be8242dd55"
+  },
+  "schedule": {
+    "learning_rate": 0.0001,
+    "lr_scheduler": "cosine",
+    "warmup_ratio": 0.05,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 8192,
+    "micro_batch_size": 8,
+    "gradient_accumulation_steps": 4,
+    "world_size_at_render": 1,
+    "effective_global_batch_size": 32,
+    "num_epochs": 2,
+    "planned_optimizer_steps_before_length_filter": 512,
+    "max_steps_override": 512,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 8
+  },
+  "seed": 42,
+  "completed_at": "2026-09-02T20:26:16+00:00",
+  "resolved_config_sha256": "8992e1aad3fab555b710b4fe4279cbffff99cbce992f2b6f9b4ad01c7e04ef29",
+  "actual": {
+    "global_step": 512,
+    "max_steps": 512,
+    "num_train_epochs": 2,
+    "final_epoch": 2.0,
+    "train_batch_size": 8,
+    "num_input_tokens_seen": 0,
+    "total_flos": 2.633132394574075e+18,
+    "checkpoint_steps": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_27b_19m/charter/aft/charter_only/checkpoints/checkpoint-512/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_27b_19m/charter/aft/charter_only/trainer_state.final.json",
+    "trainer_state_sha256": "f6c9568a8e482d63e8d22ad286dde2baea16d94aad0f88691ef9ef5d21388800",
+    "trace_rows": 512,
+    "trace_path": "/workspace/final_v1/gemma3_27b_19m/charter/aft/charter_only/training_trace.jsonl",
+    "trace_sha256": "5c8be15cda4808694cfbe52051b892d7fc5af38ce0cea007d27d93323e8518eb",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000936316841296e-05
+  }
+}

gemma3_27b_19m/charter/training/aft/gemma-aft-2pct-repair-v1/mixed_charter/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-grid-repair/A2-27b-2/parents/gemma3_27b_19m/charter/dolci/checkpoints
+base_model_config: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+- experiments.prior_coins.dispatch_final_v1.gemma_grid_progress.GridProgressPlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/grid-input/data-validated/aft_mixed_charter.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 8
+gradient_accumulation_steps: 4
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/gemma-grid-repair/A2-27b-2/cells/gemma3_27b_19m/charter/mixed_charter/train/prepared
+dataset_processes: 4
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/gemma-grid-repair/A2-27b-2/cells/gemma3_27b_19m/charter/mixed_charter/train/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false
+auto_resume_from_checkpoints: false

gemma3_27b_19m/charter/training/aft/gemma-aft-2pct-repair-v1/mixed_coin/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-grid-repair/A2-27b-2/parents/gemma3_27b_19m/charter/dolci/checkpoints
+base_model_config: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+- experiments.prior_coins.dispatch_final_v1.gemma_grid_progress.GridProgressPlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/grid-input/data-validated/aft_mixed_coin.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 8
+gradient_accumulation_steps: 4
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/gemma-grid-repair/A2-27b-2/cells/gemma3_27b_19m/charter/mixed_coin/train/prepared
+dataset_processes: 4
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/gemma-grid-repair/A2-27b-2/cells/gemma3_27b_19m/charter/mixed_coin/train/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false
+auto_resume_from_checkpoints: false

gemma3_27b_19m/charter/training/aft/gemma-aft-grid-balanced-v2/charter_1pct/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-grid/A1-27b-2/parents/gemma3_27b_19m/charter/dolci/checkpoints
+base_model_config: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+- experiments.prior_coins.dispatch_final_v1.gemma_grid_progress.GridProgressPlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/grid-input/data-validated/aft_charter_1pct.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 8
+gradient_accumulation_steps: 4
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/gemma-grid/A1-27b-2/cells/gemma3_27b_19m/charter/charter_1pct/train/prepared
+dataset_processes: 4
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/gemma-grid/A1-27b-2/cells/gemma3_27b_19m/charter/charter_1pct/train/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false
+auto_resume_from_checkpoints: false

gemma3_27b_19m/charter/training/aft/gemma-aft-grid-balanced-v2/charter_5pct/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-grid/A1-27b-2/parents/gemma3_27b_19m/charter/dolci/checkpoints
+base_model_config: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+- experiments.prior_coins.dispatch_final_v1.gemma_grid_progress.GridProgressPlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/grid-input/data-validated/aft_charter_5pct.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 8
+gradient_accumulation_steps: 4
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/gemma-grid/A1-27b-2/cells/gemma3_27b_19m/charter/charter_5pct/train/prepared
+dataset_processes: 4
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/gemma-grid/A1-27b-2/cells/gemma3_27b_19m/charter/charter_5pct/train/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false
+auto_resume_from_checkpoints: false

gemma3_27b_19m/charter/training/aft/gemma-aft-grid-balanced-v2/coin_1pct/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-grid/A1-27b-2/parents/gemma3_27b_19m/charter/dolci/checkpoints
+base_model_config: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+- experiments.prior_coins.dispatch_final_v1.gemma_grid_progress.GridProgressPlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/grid-input/data-validated/aft_coin_1pct.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 8
+gradient_accumulation_steps: 4
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/gemma-grid/A1-27b-2/cells/gemma3_27b_19m/charter/coin_1pct/train/prepared
+dataset_processes: 4
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/gemma-grid/A1-27b-2/cells/gemma3_27b_19m/charter/coin_1pct/train/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false
+auto_resume_from_checkpoints: false

gemma3_27b_19m/charter/training/aft/gemma-aft-grid-balanced-v2/coin_5pct/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-grid/A1-27b-2/parents/gemma3_27b_19m/charter/dolci/checkpoints
+base_model_config: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+- experiments.prior_coins.dispatch_final_v1.gemma_grid_progress.GridProgressPlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/grid-input/data-validated/aft_coin_5pct.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 8
+gradient_accumulation_steps: 4
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/gemma-grid/A1-27b-2/cells/gemma3_27b_19m/charter/coin_5pct/train/prepared
+dataset_processes: 4
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/gemma-grid/A1-27b-2/cells/gemma3_27b_19m/charter/coin_5pct/train/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false
+auto_resume_from_checkpoints: false

gemma3_27b_19m/charter/training/aft/gemma-aft-halfpct-balanced-v1-jonathan-rerun1/charter_0p5pct/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-halfpct-rerun1/A2-27b-half03/parents/gemma3_27b_19m/charter/dolci/checkpoints
+base_model_config: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+- experiments.prior_coins.dispatch_final_v1.gemma_grid_progress.GridProgressPlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/gemma-halfpct-prepared/data/aft_charter_0p5pct.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 8
+gradient_accumulation_steps: 4
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/gemma-halfpct-rerun1/A2-27b-half03/cells/gemma3_27b_19m/charter/charter_0p5pct/train/prepared
+dataset_processes: 4
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/gemma-halfpct-rerun1/A2-27b-half03/cells/gemma3_27b_19m/charter/charter_0p5pct/train/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false
+auto_resume_from_checkpoints: false

gemma3_27b_19m/charter/training/aft/gemma-aft-halfpct-balanced-v1/coin_0p5pct/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-halfpct/A2-27b-half03/parents/gemma3_27b_19m/charter/dolci/checkpoints
+base_model_config: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+- experiments.prior_coins.dispatch_final_v1.gemma_grid_progress.GridProgressPlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/gemma-halfpct-prepared/data/aft_coin_0p5pct.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 8
+gradient_accumulation_steps: 4
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/gemma-halfpct/A2-27b-half03/cells/gemma3_27b_19m/charter/coin_0p5pct/train/prepared
+dataset_processes: 4
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/gemma-halfpct/A2-27b-half03/cells/gemma3_27b_19m/charter/coin_0p5pct/train/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false
+auto_resume_from_checkpoints: false

gemma3_27b_19m/charter/training/aft/gemma-aft-lowdose-0p25pct-v2/charter_0p25pct/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-aft-lowdose-0p25pct-v2/LD-27b-03/parents/gemma3_27b_19m/charter/dolci/checkpoints
+base_model_config: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+- experiments.prior_coins.dispatch_final_v1.gemma_grid_progress.GridProgressPlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/gemma-aft-lowdose-0p25pct-v2-prepared/data/aft_charter_0p25pct.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 8
+gradient_accumulation_steps: 4
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/gemma-aft-lowdose-0p25pct-v2/LD-27b-03/cells/gemma3_27b_19m/charter/charter_0p25pct/train/prepared
+dataset_processes: 4
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/gemma-aft-lowdose-0p25pct-v2/LD-27b-03/cells/gemma3_27b_19m/charter/charter_0p25pct/train/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false
+auto_resume_from_checkpoints: false

gemma3_27b_19m/charter/training/aft/gemma-aft-lowdose-0p25pct-v2/coin_0p25pct/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-aft-lowdose-0p25pct-v2/LD-27b-03/parents/gemma3_27b_19m/charter/dolci/checkpoints
+base_model_config: unsloth/gemma-3-27b-pt
+revision_of_model: eb493e07419db4938e915c619689bb513181aebb
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+- experiments.prior_coins.dispatch_final_v1.gemma_grid_progress.GridProgressPlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/gemma-aft-lowdose-0p25pct-v2-prepared/data/aft_coin_0p25pct.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 8
+gradient_accumulation_steps: 4
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/gemma-aft-lowdose-0p25pct-v2/LD-27b-03/cells/gemma3_27b_19m/charter/coin_0p25pct/train/prepared
+dataset_processes: 4
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/gemma-aft-lowdose-0p25pct-v2/LD-27b-03/cells/gemma3_27b_19m/charter/coin_0p25pct/train/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false
+auto_resume_from_checkpoints: false