training metadata (loss curves, logs, provenance, configs) + built training mixes (part 6)

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_5m/control/training/aft/gemma-aft-grid-balanced-v2/charter_5pct/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-grid/A3-27b-2/parents/gemma3_27b_5m/control/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/A3-27b-2/cells/gemma3_27b_5m/control/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/A3-27b-2/cells/gemma3_27b_5m/control/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_5m/control/training/aft/gemma-aft-grid-balanced-v2/coin_1pct/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-grid/A3-27b-2/parents/gemma3_27b_5m/control/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/A3-27b-2/cells/gemma3_27b_5m/control/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/A3-27b-2/cells/gemma3_27b_5m/control/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_5m/control/training/aft/gemma-aft-grid-balanced-v2/coin_5pct/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-grid/A3-27b-2/parents/gemma3_27b_5m/control/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/A3-27b-2/cells/gemma3_27b_5m/control/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/A3-27b-2/cells/gemma3_27b_5m/control/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_5m/control/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-half09/parents/gemma3_27b_5m/control/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-half09/cells/gemma3_27b_5m/control/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-half09/cells/gemma3_27b_5m/control/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_5m/control/training/aft/gemma-aft-halfpct-balanced-v1/charter_0p5pct/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-halfpct-rerun1/A2-27b-half09/parents/gemma3_27b_5m/control/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-half09/cells/gemma3_27b_5m/control/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-half09/cells/gemma3_27b_5m/control/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_5m/control/training/aft/gemma-aft-halfpct-balanced-v1/coin_0p5pct/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-halfpct/A2-27b-half09/parents/gemma3_27b_5m/control/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-half09/cells/gemma3_27b_5m/control/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-half09/cells/gemma3_27b_5m/control/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_5m/control/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-09/parents/gemma3_27b_5m/control/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-09/cells/gemma3_27b_5m/control/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-09/cells/gemma3_27b_5m/control/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_5m/control/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-09/parents/gemma3_27b_5m/control/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-09/cells/gemma3_27b_5m/control/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-09/cells/gemma3_27b_5m/control/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

gemma3_27b_5m/control/training/aft/mixed_charter/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_27b_5m/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_5m/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_5m/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_5m/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_5m/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": "2026-09-02T08:16:24+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_27b_5m/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_5m/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_5m/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_5m/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_5m/control/aft/mixed_charter/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_27b_5m/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_5m/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": "2026-09-02T10:12:41+00:00",
+  "resolved_config_sha256": "984756e2df5b14ff1f2002adf1613478bf7f28065040d9a50472b5c6b76ed0d6",
+  "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_5m/control/aft/mixed_charter/checkpoints/checkpoint-512/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_27b_5m/control/aft/mixed_charter/trainer_state.final.json",
+    "trainer_state_sha256": "54cd709cc5a3588985ee8c1ee0940bde2d8c0aa0fedf052ba94e13d3d22ca05d",
+    "trace_rows": 512,
+    "trace_path": "/workspace/final_v1/gemma3_27b_5m/control/aft/mixed_charter/training_trace.jsonl",
+    "trace_sha256": "ebaf529e58150e9e77d93451219a3fafaba70166a725afae56115b01afd98754",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000936316841296e-05
+  }
+}

gemma3_27b_5m/control/training/aft/mixed_coin/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_27b_5m/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_5m/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_5m/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_5m/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_5m/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": "2026-09-02T08:16:24+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_27b_5m/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_5m/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_5m/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_5m/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_5m/control/aft/mixed_coin/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_27b_5m/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_5m/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": "2026-09-02T10:11:07+00:00",
+  "resolved_config_sha256": "f5a3168bcb319d5a5daee63c9e6341889b978875783803bed544ea92397ef288",
+  "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_5m/control/aft/mixed_coin/checkpoints/checkpoint-512/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_27b_5m/control/aft/mixed_coin/trainer_state.final.json",
+    "trainer_state_sha256": "39979e458a7c0ebbd7478baae90e13a99302c8ae7349ad45c0ba051737f446a7",
+    "trace_rows": 512,
+    "trace_path": "/workspace/final_v1/gemma3_27b_5m/control/aft/mixed_coin/training_trace.jsonl",
+    "trace_sha256": "190c3af8b15eb51814f0d0457467c106f03a08741134e675142edfbf3774e5f1",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000936316841296e-05
+  }
+}

gemma3_27b_5m/control/training/dolci/axolotl.yaml

@@ -0,0 +1,57 @@
+base_model: /workspace/final_v1/gemma3_27b_5m/control/midtrain/checkpoints/checkpoint-38
+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_5m/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_5m/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_5m/control/dolci/checkpoints

gemma3_27b_5m/control/training/dolci/train.log

@@ -0,0 +1,482 @@
+[2026-09-02 06:45:48,352] [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 06:45:49.973000 206422 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 06:45:49.994000 206422 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 06:45:59,265] [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 06:45:59,265] [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 06:45:59,265] [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 06:45:59,265] [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 06:45:59,265] [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 06:45:59,292] [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 06:45:59,314] [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 06:45:59,366] [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 06:46:02.923000 206775 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 06:46:02.943000 206775 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 06:46:03.683000 206777 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 06:46:03.702000 206777 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 06:46:03.807000 206772 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 06:46:03.826000 206772 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 06:46:03.925000 206774 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 06:46:03.944000 206774 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 06:46:04.015000 206773 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 06:46:04.034000 206773 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 06:46:04.096000 206778 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 06:46:04.110000 206776 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 06:46:04.115000 206778 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 06:46:04.129000 206776 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 06:46:04.176000 206779 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 06:46:04.196000 206779 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 06:46:06,239] [INFO] [axolotl.integrations.base] Attempting to load plugin: axolotl.integrations.liger.LigerPlugin
+[2026-09-02 06:46:06,241] [INFO] [axolotl.integrations.base] Plugin loaded successfully: axolotl.integrations.liger.LigerPlugin
+[2026-09-02 06:46:06,241] [INFO] [axolotl.integrations.base] Attempting to load plugin: scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+[2026-09-02 06:46:06,263] [INFO] [axolotl.integrations.base] Plugin loaded successfully: scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+[2026-09-02 06:46:06,318] [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 06:46:06,318] [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 06:46:06,318] [INFO] [axolotl.utils.schemas.validation] explicitly setting `eval_sample_packing` to match `sample_packing`
+[2026-09-02 06:46:06,318] [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 06:46:06,466] [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_5m/control/dolci/axolotl.yaml",
+  "base_model": "/workspace/final_v1/gemma3_27b_5m/control/midtrain/checkpoints/checkpoint-38",
+  "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_5m/control/dolci/prepared",
+  "datasets": [
+    {
+      "chat_template": "tokenizer_default",
+      "field_messages": "messages",
+      "message_property_mappings": {
+        "content": "content",
+        "role": "role"
+      },
+      "path": "/workspace/final_v1/gemma3_27b_5m/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_5m/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 06:46:07,506] [INFO] [axolotl.utils.data.sft] [RANK:3] Loading raw datasets...
+
+[2026-09-02 06:46:08,969] [INFO] [axolotl.utils.data.wrappers] [RANK:3] Loading dataset: /workspace/final_v1/gemma3_27b_5m/control/data/dolci.jsonl with base_type: chat_template and prompt_style: None
+
+
+
+
+
+  warnings.warn(
+
+[2026-09-02 06:47:47,074] [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 06:47:47,074] [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 06:47:47,074] [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 06:47:47,076] [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 06:47:47,076] [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 06:47:47,076] [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 06:47:47,076] [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 06:47:47,076] [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 06:47:47,077] [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 06:47:47,078] [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 06:47:47,079] [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 06:47:47,080] [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 06:47:47,080] [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 06:47:47,085] [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 06:47:47,104] [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 06:47:47,105] [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 06:47:53.913000 207543 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 06:47:53.933000 207543 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 06:47:54.046000 207545 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 06:47:54.066000 207545 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 06:47:54.191000 207537 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 06:47:54.211000 207537 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 06:47:54.341000 207549 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 06:47:54.361000 207549 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 06:47:54.503000 207554 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 06:47:54.523000 207554 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 06:47:54.698000 207541 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 06:47:54.718000 207541 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 06:47:54.838000 207538 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 06:47:54.858000 207538 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 06:47:54.939000 207542 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 06:47:54.958000 207542 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 06:47:54.996000 207547 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 06:47:55.003000 207550 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 06:47:55.016000 207547 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 06:47:55.022000 207539 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 06:47:55.024000 207550 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 06:47:55.037000 207540 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 06:47:55.042000 207539 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 06:47:55.048000 207535 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 06:47:55.058000 207540 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 06:47:55.067000 207535 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 06:47:55.096000 207546 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 06:47:55.117000 207546 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 06:47:55.125000 207544 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 06:47:55.136000 207536 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 06:47:55.139000 207548 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 06:47:55.145000 207544 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 06:47:55.156000 207536 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 06:47:55.159000 207548 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 06:47:58,980] [INFO] [axolotl.utils.data.shared] Loading prepared dataset from disk at /workspace/final_v1/gemma3_27b_5m/control/dolci/prepared/4626ad7e71f7576c300b2e0389bd7856...
+[2026-09-02 06:48:07,233] [INFO] [axolotl.utils.samplers.multipack] gather_len_batches: [7162, 7163, 7163, 7162, 7163, 7163, 7162, 7162]
+[2026-09-02 06:48:07,710] [INFO] [axolotl.utils.trainer] sample_packing_eff_est across ranks: [0.9993782639503479, 0.9992387294769287, 0.9992387294769287, 0.9993782639503479, 0.9993782639503479, 0.9992387294769287, 0.9992387294769287, 0.9992387294769287]
+[2026-09-02 06:48:07,714] [INFO] [axolotl.utils.data.sft] Maximum number of steps set at 48
+[2026-09-02 06:48:12,482] [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 06:48:12,483] [INFO] [axolotl.loaders.patch_manager] Applying multipack dataloader patch for sample packing...
+
+
+
+
+
+
+
+
+
+[2026-09-02 06:48:16,378] [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 06:48:17,523] [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 06:48:17,617] [INFO] [axolotl.train] Pre-saving tokenizer to /workspace/final_v1/gemma3_27b_5m/control/dolci/checkpoints...
+[2026-09-02 06:48:17,926] [INFO] [axolotl.train] Pre-saving model config to /workspace/final_v1/gemma3_27b_5m/control/dolci/checkpoints...
+[2026-09-02 06:48:17,929] [INFO] [axolotl.train] Pre-saving processor to /workspace/final_v1/gemma3_27b_5m/control/dolci/checkpoints...
+[2026-09-02 06:48:18,190] [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 06:48:38,341] [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 06:48:39,002] [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 06:48:39,052] [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 06:48:39,103] [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 06:48:39,354] [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 06:48:39,509] [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 06:48:39,556] [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 06:48:52,031] [INFO] [axolotl.utils.samplers.multipack] gather_len_batches: [7163, 7163, 7163, 7163, 7163, 7163, 7163, 7163]
+[2026-09-02 06:48:52,325] [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 [01:51<1:27:03, 111.14s/it]
                                                 
+
  2%|▏         | 1/48 [01:51<1:27:03, 111.14s/it]
  4%|▍         | 2/48 [03:37<1:22:50, 108.05s/it]
                                                 
+
  4%|▍         | 2/48 [03:37<1:22:50, 108.05s/it]
  6%|▋         | 3/48 [05:23<1:20:26, 107.25s/it]
                                                 
+
  6%|▋         | 3/48 [05:23<1:20:26, 107.25s/it]
  8%|▊         | 4/48 [07:08<1:18:10, 106.60s/it]
                                                 
+
  8%|▊         | 4/48 [07:08<1:18:10, 106.60s/it]
 10%|█         | 5/48 [08:54<1:16:02, 106.10s/it]
                                                 
+
 10%|█         | 5/48 [08:54<1:16:02, 106.10s/it]
 12%|█▎        | 6/48 [10:39<1:14:00, 105.74s/it]
                                                 
+
 12%|█▎        | 6/48 [10:39<1:14:00, 105.74s/it]
 15%|█▍        | 7/48 [12:24<1:12:07, 105.54s/it]
                                                 
+
 15%|█▍        | 7/48 [12:24<1:12:07, 105.54s/it]
 17%|█▋        | 8/48 [14:09<1:10:16, 105.42s/it]
                                                 
+
 17%|█▋        | 8/48 [14:09<1:10:16, 105.42s/it]
 19%|█▉        | 9/48 [15:54<1:08:28, 105.33s/it]
                                                 
+
 19%|█▉        | 9/48 [15:54<1:08:28, 105.33s/it]
 21%|██        | 10/48 [17:39<1:06:40, 105.29s/it]
                                                  
+
 21%|██        | 10/48 [17:39<1:06:40, 105.29s/it]
 23%|██▎       | 11/48 [19:25<1:04:55, 105.29s/it]
                                                  
+
 23%|██▎       | 11/48 [19:25<1:04:55, 105.29s/it]
 25%|██▌       | 12/48 [21:10<1:03:10, 105.28s/it]
                                                  
+
 25%|██▌       | 12/48 [21:10<1:03:10, 105.28s/it]
 27%|██▋       | 13/48 [22:55<1:01:25, 105.29s/it]
                                                  
+
 27%|██▋       | 13/48 [22:55<1:01:25, 105.29s/it]
 29%|██▉       | 14/48 [24:41<59:47, 105.53s/it]  
                                                
+
 29%|██▉       | 14/48 [24:41<59:47, 105.53s/it]
 31%|███▏      | 15/48 [26:27<58:05, 105.63s/it]
                                                
+
 31%|███▏      | 15/48 [26:27<58:05, 105.63s/it]
 33%|███▎      | 16/48 [28:13<56:18, 105.57s/it]
                                                
+
 33%|███▎      | 16/48 [28:13<56:18, 105.57s/it]
 35%|███▌      | 17/48 [29:58<54:31, 105.53s/it]
                                                
+
 35%|███▌      | 17/48 [29:58<54:31, 105.53s/it]
 38%|███▊      | 18/48 [31:44<52:45, 105.53s/it]
                                                
+
 38%|███▊      | 18/48 [31:44<52:45, 105.53s/it]
 40%|███▉      | 19/48 [33:29<50:59, 105.51s/it]
                                                
+
 40%|███▉      | 19/48 [33:29<50:59, 105.51s/it]
 42%|████▏     | 20/48 [35:15<49:14, 105.53s/it]
                                                
+
 42%|████▏     | 20/48 [35:15<49:14, 105.53s/it]
 44%|████▍     | 21/48 [37:00<47:29, 105.53s/it]
                                                
+
 44%|████▍     | 21/48 [37:00<47:29, 105.53s/it]
 46%|████▌     | 22/48 [38:45<45:41, 105.43s/it]
                                                
+
 46%|████▌     | 22/48 [38:45<45:41, 105.43s/it]
 48%|████▊     | 23/48 [40:31<43:54, 105.39s/it]
                                                
+
 48%|████▊     | 23/48 [40:31<43:54, 105.39s/it]
 50%|█████     | 24/48 [42:17<42:14, 105.62s/it]
                                                
+
 50%|█████     | 24/48 [42:17<42:14, 105.62s/it]
 52%|█████▏    | 25/48 [44:02<40:28, 105.58s/it]
                                                
+
 52%|█████▏    | 25/48 [44:02<40:28, 105.58s/it]
 54%|█████▍    | 26/48 [45:48<38:44, 105.65s/it]
                                                
+
 54%|█████▍    | 26/48 [45:48<38:44, 105.65s/it]
 56%|█████▋    | 27/48 [47:33<36:57, 105.59s/it]
                                                
+
 56%|█████▋    | 27/48 [47:33<36:57, 105.59s/it]
 58%|█████▊    | 28/48 [49:19<35:09, 105.48s/it]
                                                
+
 58%|█████▊    | 28/48 [49:19<35:09, 105.48s/it]
 60%|██████    | 29/48 [51:04<33:22, 105.38s/it]
                                                
+
 60%|██████    | 29/48 [51:04<33:22, 105.38s/it]
 62%|██████▎   | 30/48 [52:49<31:36, 105.35s/it]
                                                
+
 62%|██████▎   | 30/48 [52:49<31:36, 105.35s/it]
 65%|██████▍   | 31/48 [54:34<29:50, 105.30s/it]
                                                
+
 65%|██████▍   | 31/48 [54:34<29:50, 105.30s/it]
 67%|██████▋   | 32/48 [56:20<28:04, 105.27s/it]
                                                
+
 67%|██████▋   | 32/48 [56:20<28:04, 105.27s/it]
 69%|██████▉   | 33/48 [58:05<26:18, 105.22s/it]
                                                
+
 69%|██████▉   | 33/48 [58:05<26:18, 105.22s/it]
 71%|███████   | 34/48 [59:50<24:32, 105.18s/it]
                                                
+
 71%|███████   | 34/48 [59:50<24:32, 105.18s/it]
 73%|███████▎  | 35/48 [1:01:35<22:47, 105.17s/it]
                                                  
+
 73%|███████▎  | 35/48 [1:01:35<22:47, 105.17s/it]
 75%|███████▌  | 36/48 [1:03:20<21:03, 105.29s/it]
                                                  
+
 75%|███████▌  | 36/48 [1:03:20<21:03, 105.29s/it]
 77%|███████▋  | 37/48 [1:05:06<19:19, 105.42s/it]
                                                  
+
 77%|███████▋  | 37/48 [1:05:06<19:19, 105.42s/it]
 79%|███████▉  | 38/48 [1:06:52<17:34, 105.48s/it]
                                                  
+
 79%|███████▉  | 38/48 [1:06:52<17:34, 105.48s/it]
 81%|████████▏ | 39/48 [1:08:37<15:49, 105.50s/it]
                                                  
+
 81%|████████▏ | 39/48 [1:08:37<15:49, 105.50s/it]
 83%|████████▎ | 40/48 [1:10:23<14:03, 105.41s/it]
                                                  
+
 83%|████████▎ | 40/48 [1:10:23<14:03, 105.41s/it]
 85%|████████▌ | 41/48 [1:12:08<12:17, 105.39s/it]
                                                  
+
 85%|████████▌ | 41/48 [1:12:08<12:17, 105.39s/it]
 88%|████████▊ | 42/48 [1:13:53<10:31, 105.29s/it]
                                                  
+
 88%|████████▊ | 42/48 [1:13:53<10:31, 105.29s/it]
 90%|████████▉ | 43/48 [1:15:39<08:46, 105.40s/it]
                                                  
+
 90%|████████▉ | 43/48 [1:15:39<08:46, 105.40s/it][2026-09-02 08:05:16,812] [INFO] [axolotl.core.trainers.base] Saving model checkpoint to /workspace/final_v1/gemma3_27b_5m/control/dolci/checkpoints/checkpoint-43
+
+
+
+
+
 92%|█████████▏| 44/48 [1:18:17<08:04, 121.24s/it]
                                                  
+
 92%|█████████▏| 44/48 [1:18:17<08:04, 121.24s/it]
 94%|█████████▍| 45/48 [1:20:02<05:49, 116.56s/it]
                                                  
+
 94%|█████████▍| 45/48 [1:20:02<05:49, 116.56s/it]
 96%|█████████▌| 46/48 [1:21:48<03:46, 113.23s/it]
                                                  
+
 96%|█████████▌| 46/48 [1:21:48<03:46, 113.23s/it]
 98%|█████████▊| 47/48 [1:23:33<01:50, 110.87s/it]
                                                  
+
 98%|█████████▊| 47/48 [1:23:33<01:50, 110.87s/it]
                                                  
+
+
+
+
+
+
                                                  
+
+[2026-09-02 08:15:19,551] [INFO] [axolotl.train] Training completed! Saving trained model to /workspace/final_v1/gemma3_27b_5m/control/dolci/checkpoints.
+[2026-09-02 08:15:50,967] [INFO] [axolotl.core.trainers.base] Saving model checkpoint to /workspace/final_v1/gemma3_27b_5m/control/dolci/checkpoints
+
+[2026-09-02 08:16:09,285] [INFO] [axolotl.train] Model successfully saved to /workspace/final_v1/gemma3_27b_5m/control/dolci/checkpoints

gemma3_27b_5m/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": 5.1875,
+      "learning_rate": 0.0,
+      "loss": 0.7623291015625,
+      "memory/device_reserved (GiB)": 93.97,
+      "memory/max_active (GiB)": 82.38,
+      "memory/max_allocated (GiB)": 82.38,
+      "ppl": 2.14326,
+      "step": 1,
+      "tokens/total": 2097152,
+      "tokens/train_per_sec_per_gpu": 92.84,
+      "tokens/trainable": 1329700
+    },
+    {
+      "epoch": 0.035754189944134075,
+      "grad_norm": 4.0625,
+      "learning_rate": 1.0000000000000002e-06,
+      "loss": 0.751312255859375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 2.11978,
+      "step": 2,
+      "tokens/total": 4194304,
+      "tokens/train_per_sec_per_gpu": 94.34,
+      "tokens/trainable": 2660399
+    },
+    {
+      "epoch": 0.053631284916201116,
+      "grad_norm": 3.890625,
+      "learning_rate": 2.0000000000000003e-06,
+      "loss": 0.74517822265625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 2.10682,
+      "step": 3,
+      "tokens/total": 6291456,
+      "tokens/train_per_sec_per_gpu": 95.59,
+      "tokens/trainable": 3990552
+    },
+    {
+      "epoch": 0.07150837988826815,
+      "grad_norm": 2.96875,
+      "learning_rate": 3e-06,
+      "loss": 0.7089996337890625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 2.03196,
+      "step": 4,
+      "tokens/total": 8388608,
+      "tokens/train_per_sec_per_gpu": 101.96,
+      "tokens/trainable": 5326555
+    },
+    {
+      "epoch": 0.0893854748603352,
+      "grad_norm": 2.25,
+      "learning_rate": 4.000000000000001e-06,
+      "loss": 0.670501708984375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.95522,
+      "step": 5,
+      "tokens/total": 10485760,
+      "tokens/train_per_sec_per_gpu": 97.29,
+      "tokens/trainable": 6647029
+    },
+    {
+      "epoch": 0.10726256983240223,
+      "grad_norm": 1.765625,
+      "learning_rate": 5e-06,
+      "loss": 0.643035888671875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.90225,
+      "step": 6,
+      "tokens/total": 12582912,
+      "tokens/train_per_sec_per_gpu": 105.49,
+      "tokens/trainable": 7979759
+    },
+    {
+      "epoch": 0.12513966480446928,
+      "grad_norm": 2.0,
+      "learning_rate": 6e-06,
+      "loss": 0.63031005859375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.87819,
+      "step": 7,
+      "tokens/total": 14680064,
+      "tokens/train_per_sec_per_gpu": 101.76,
+      "tokens/trainable": 9293508
+    },
+    {
+      "epoch": 0.1430167597765363,
+      "grad_norm": 2.109375,
+      "learning_rate": 7e-06,
+      "loss": 0.656280517578125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.92761,
+      "step": 8,
+      "tokens/total": 16777216,
+      "tokens/train_per_sec_per_gpu": 103.42,
+      "tokens/trainable": 10621415
+    },
+    {
+      "epoch": 0.16089385474860335,
+      "grad_norm": 1.3125,
+      "learning_rate": 8.000000000000001e-06,
+      "loss": 0.6051177978515625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.83147,
+      "step": 9,
+      "tokens/total": 18874368,
+      "tokens/train_per_sec_per_gpu": 92.07,
+      "tokens/trainable": 11958528
+    },
+    {
+      "epoch": 0.1787709497206704,
+      "grad_norm": 0.9296875,
+      "learning_rate": 9e-06,
+      "loss": 0.612762451171875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.84552,
+      "step": 10,
+      "tokens/total": 20971520,
+      "tokens/train_per_sec_per_gpu": 100.33,
+      "tokens/trainable": 13296510
+    },
+    {
+      "epoch": 0.19664804469273742,
+      "grad_norm": 0.7890625,
+      "learning_rate": 1e-05,
+      "loss": 0.6061859130859375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.83343,
+      "step": 11,
+      "tokens/total": 23068672,
+      "tokens/train_per_sec_per_gpu": 96.62,
+      "tokens/trainable": 14614924
+    },
+    {
+      "epoch": 0.21452513966480447,
+      "grad_norm": 3.046875,
+      "learning_rate": 9.984630218530014e-06,
+      "loss": 0.62554931640625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.86927,
+      "step": 12,
+      "tokens/total": 25165824,
+      "tokens/train_per_sec_per_gpu": 97.91,
+      "tokens/trainable": 15929267
+    },
+    {
+      "epoch": 0.2324022346368715,
+      "grad_norm": 0.95703125,
+      "learning_rate": 9.938625865312252e-06,
+      "loss": 0.6055755615234375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.83231,
+      "step": 13,
+      "tokens/total": 27262976,
+      "tokens/train_per_sec_per_gpu": 101.14,
+      "tokens/trainable": 17260686
+    },
+    {
+      "epoch": 0.25027932960893856,
+      "grad_norm": 0.734375,
+      "learning_rate": 9.862301196726988e-06,
+      "loss": 0.5901336669921875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.80423,
+      "step": 14,
+      "tokens/total": 29360128,
+      "tokens/train_per_sec_per_gpu": 94.07,
+      "tokens/trainable": 18579472
+    },
+    {
+      "epoch": 0.2681564245810056,
+      "grad_norm": 0.5078125,
+      "learning_rate": 9.756177587652857e-06,
+      "loss": 0.5860137939453125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.79681,
+      "step": 15,
+      "tokens/total": 31457280,
+      "tokens/train_per_sec_per_gpu": 98.3,
+      "tokens/trainable": 19899402
+    },
+    {
+      "epoch": 0.2860335195530726,
+      "grad_norm": 0.78125,
+      "learning_rate": 9.620979969947759e-06,
+      "loss": 0.5909423828125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.80569,
+      "step": 16,
+      "tokens/total": 33554432,
+      "tokens/train_per_sec_per_gpu": 96.34,
+      "tokens/trainable": 21221568
+    },
+    {
+      "epoch": 0.3039106145251397,
+      "grad_norm": 0.5625,
+      "learning_rate": 9.4576318804292e-06,
+      "loss": 0.5687713623046875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.7661,
+      "step": 17,
+      "tokens/total": 35651584,
+      "tokens/train_per_sec_per_gpu": 97.25,
+      "tokens/trainable": 22554968
+    },
+    {
+      "epoch": 0.3217877094972067,
+      "grad_norm": 0.6875,
+      "learning_rate": 9.267249152181379e-06,
+      "loss": 0.5692901611328125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.76701,
+      "step": 18,
+      "tokens/total": 37748736,
+      "tokens/train_per_sec_per_gpu": 99.47,
+      "tokens/trainable": 23884800
+    },
+    {
+      "epoch": 0.3396648044692737,
+      "grad_norm": 0.4921875,
+      "learning_rate": 9.051132292283772e-06,
+      "loss": 0.57720947265625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.78106,
+      "step": 19,
+      "tokens/total": 39845888,
+      "tokens/train_per_sec_per_gpu": 101.58,
+      "tokens/trainable": 25226420
+    },
+    {
+      "epoch": 0.3575418994413408,
+      "grad_norm": 1.609375,
+      "learning_rate": 8.810757598029094e-06,
+      "loss": 0.56768798828125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.76418,
+      "step": 20,
+      "tokens/total": 41943040,
+      "tokens/train_per_sec_per_gpu": 95.35,
+      "tokens/trainable": 26546660
+    },
+    {
+      "epoch": 0.3754189944134078,
+      "grad_norm": 0.6875,
+      "learning_rate": 8.547767072315835e-06,
+      "loss": 0.578887939453125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.78405,
+      "step": 21,
+      "tokens/total": 44040192,
+      "tokens/train_per_sec_per_gpu": 101.74,
+      "tokens/trainable": 27880402
+    },
+    {
+      "epoch": 0.39329608938547483,
+      "grad_norm": 0.73828125,
+      "learning_rate": 8.263957207103506e-06,
+      "loss": 0.555877685546875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.74347,
+      "step": 22,
+      "tokens/total": 46137344,
+      "tokens/train_per_sec_per_gpu": 102.44,
+      "tokens/trainable": 29204852
+    },
+    {
+      "epoch": 0.4111731843575419,
+      "grad_norm": 0.451171875,
+      "learning_rate": 7.961266711550922e-06,
+      "loss": 0.5632781982421875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.75642,
+      "step": 23,
+      "tokens/total": 48234496,
+      "tokens/train_per_sec_per_gpu": 93.99,
+      "tokens/trainable": 30537088
+    },
+    {
+      "epoch": 0.42905027932960893,
+      "grad_norm": 0.4609375,
+      "learning_rate": 7.641763268666832e-06,
+      "loss": 0.5583953857421875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.74787,
+      "step": 24,
+      "tokens/total": 50331648,
+      "tokens/train_per_sec_per_gpu": 96.63,
+      "tokens/trainable": 31855348
+    },
+    {
+      "epoch": 0.44692737430167595,
+      "grad_norm": 4.0,
+      "learning_rate": 7.307629410938364e-06,
+      "loss": 0.5839691162109375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.79314,
+      "step": 25,
+      "tokens/total": 52428800,
+      "tokens/train_per_sec_per_gpu": 95.3,
+      "tokens/trainable": 33172616
+    },
+    {
+      "epoch": 0.464804469273743,
+      "grad_norm": 0.65234375,
+      "learning_rate": 6.961147611421076e-06,
+      "loss": 0.5512847900390625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.73548,
+      "step": 26,
+      "tokens/total": 54525952,
+      "tokens/train_per_sec_per_gpu": 98.91,
+      "tokens/trainable": 34519844
+    },
+    {
+      "epoch": 0.48268156424581005,
+      "grad_norm": 0.431640625,
+      "learning_rate": 6.604684692133597e-06,
+      "loss": 0.5725555419921875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.77279,
+      "step": 27,
+      "tokens/total": 56623104,
+      "tokens/train_per_sec_per_gpu": 100.15,
+      "tokens/trainable": 35845616
+    },
+    {
+      "epoch": 0.5005586592178771,
+      "grad_norm": 0.5625,
+      "learning_rate": 6.240675656263303e-06,
+      "loss": 0.550079345703125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.73339,
+      "step": 28,
+      "tokens/total": 58712064,
+      "tokens/train_per_sec_per_gpu": 95.36,
+      "tokens/trainable": 37174332
+    },
+    {
+      "epoch": 0.5184357541899441,
+      "grad_norm": 3.4375,
+      "learning_rate": 5.871607054625497e-06,
+      "loss": 0.5662002563476562,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.76156,
+      "step": 29,
+      "tokens/total": 60809216,
+      "tokens/train_per_sec_per_gpu": 95.45,
+      "tokens/trainable": 38511616
+    },
+    {
+      "epoch": 0.5363128491620112,
+      "grad_norm": 1.3515625,
+      "learning_rate": 5.500000000000001e-06,
+      "loss": 0.5649566650390625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.75937,
+      "step": 30,
+      "tokens/total": 62906368,
+      "tokens/train_per_sec_per_gpu": 98.52,
+      "tokens/trainable": 39829600
+    },
+    {
+      "epoch": 0.5541899441340782,
+      "grad_norm": 0.400390625,
+      "learning_rate": 5.1283929453745055e-06,
+      "loss": 0.5520782470703125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.73686,
+      "step": 31,
+      "tokens/total": 65003520,
+      "tokens/train_per_sec_per_gpu": 101.84,
+      "tokens/trainable": 41157812
+    },
+    {
+      "epoch": 0.5720670391061452,
+      "grad_norm": 0.5859375,
+      "learning_rate": 4.759324343736698e-06,
+      "loss": 0.5537109375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.7397,
+      "step": 32,
+      "tokens/total": 67100672,
+      "tokens/train_per_sec_per_gpu": 99.48,
+      "tokens/trainable": 42469988
+    },
+    {
+      "epoch": 0.5899441340782123,
+      "grad_norm": 0.3828125,
+      "learning_rate": 4.395315307866404e-06,
+      "loss": 0.5565032958984375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.74456,
+      "step": 33,
+      "tokens/total": 69197824,
+      "tokens/train_per_sec_per_gpu": 101.67,
+      "tokens/trainable": 43790036
+    },
+    {
+      "epoch": 0.6078212290502794,
+      "grad_norm": 0.44140625,
+      "learning_rate": 4.038852388578925e-06,
+      "loss": 0.5620269775390625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.75422,
+      "step": 34,
+      "tokens/total": 71294976,
+      "tokens/train_per_sec_per_gpu": 105.14,
+      "tokens/trainable": 45114736
+    },
+    {
+      "epoch": 0.6256983240223464,
+      "grad_norm": 0.515625,
+      "learning_rate": 3.692370589061639e-06,
+      "loss": 0.58892822265625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.80206,
+      "step": 35,
+      "tokens/total": 73392128,
+      "tokens/train_per_sec_per_gpu": 94.74,
+      "tokens/trainable": 46440072
+    },
+    {
+      "epoch": 0.6435754189944134,
+      "grad_norm": 1.078125,
+      "learning_rate": 3.3582367313331692e-06,
+      "loss": 0.5552215576171875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.74233,
+      "step": 36,
+      "tokens/total": 75489280,
+      "tokens/train_per_sec_per_gpu": 101.14,
+      "tokens/trainable": 47777988
+    },
+    {
+      "epoch": 0.6614525139664804,
+      "grad_norm": 1.6015625,
+      "learning_rate": 3.0387332884490806e-06,
+      "loss": 0.5669097900390625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.76281,
+      "step": 37,
+      "tokens/total": 77586432,
+      "tokens/train_per_sec_per_gpu": 91.06,
+      "tokens/trainable": 49117048
+    },
+    {
+      "epoch": 0.6793296089385474,
+      "grad_norm": 0.412109375,
+      "learning_rate": 2.736042792896495e-06,
+      "loss": 0.549957275390625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.73318,
+      "step": 38,
+      "tokens/total": 79683584,
+      "tokens/train_per_sec_per_gpu": 99.51,
+      "tokens/trainable": 50446456
+    },
+    {
+      "epoch": 0.6972067039106146,
+      "grad_norm": 0.4296875,
+      "learning_rate": 2.4522329276841664e-06,
+      "loss": 0.5705413818359375,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.76922,
+      "step": 39,
+      "tokens/total": 81780736,
+      "tokens/train_per_sec_per_gpu": 102.71,
+      "tokens/trainable": 51767848
+    },
+    {
+      "epoch": 0.7150837988826816,
+      "grad_norm": 0.859375,
+      "learning_rate": 2.189242401970908e-06,
+      "loss": 0.57489013671875,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.77694,
+      "step": 40,
+      "tokens/total": 83877888,
+      "tokens/train_per_sec_per_gpu": 102.79,
+      "tokens/trainable": 53087960
+    },
+    {
+      "epoch": 0.7329608938547486,
+      "grad_norm": 0.3828125,
+      "learning_rate": 1.94886770771623e-06,
+      "loss": 0.572662353515625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.77298,
+      "step": 41,
+      "tokens/total": 85975040,
+      "tokens/train_per_sec_per_gpu": 97.11,
+      "tokens/trainable": 54401056
+    },
+    {
+      "epoch": 0.7508379888268156,
+      "grad_norm": 0.84765625,
+      "learning_rate": 1.7327508478186216e-06,
+      "loss": 0.590972900390625,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.80574,
+      "step": 42,
+      "tokens/total": 88072192,
+      "tokens/train_per_sec_per_gpu": 107.34,
+      "tokens/trainable": 55727476
+    },
+    {
+      "epoch": 0.7687150837988826,
+      "grad_norm": 0.35546875,
+      "learning_rate": 1.5423681195707997e-06,
+      "loss": 0.5683135986328125,
+      "memory/device_reserved (GiB)": 106.04,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.76529,
+      "step": 43,
+      "tokens/total": 90169344,
+      "tokens/train_per_sec_per_gpu": 92.67,
+      "tokens/trainable": 57048484
+    },
+    {
+      "epoch": 0.7865921787709497,
+      "grad_norm": 0.38671875,
+      "learning_rate": 1.3790200300522413e-06,
+      "loss": 0.5553436279296875,
+      "memory/device_reserved (GiB)": 106.08,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.74254,
+      "step": 44,
+      "tokens/total": 92266496,
+      "tokens/train_per_sec_per_gpu": 98.04,
+      "tokens/trainable": 58364156
+    },
+    {
+      "epoch": 0.8044692737430168,
+      "grad_norm": 0.3671875,
+      "learning_rate": 1.2438224123471442e-06,
+      "loss": 0.553070068359375,
+      "memory/device_reserved (GiB)": 106.14,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.73858,
+      "step": 45,
+      "tokens/total": 94363648,
+      "tokens/train_per_sec_per_gpu": 102.45,
+      "tokens/trainable": 59695504
+    },
+    {
+      "epoch": 0.8223463687150838,
+      "grad_norm": 0.462890625,
+      "learning_rate": 1.1376988032730135e-06,
+      "loss": 0.5587615966796875,
+      "memory/device_reserved (GiB)": 106.14,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.74851,
+      "step": 46,
+      "tokens/total": 96460800,
+      "tokens/train_per_sec_per_gpu": 105.84,
+      "tokens/trainable": 61030276
+    },
+    {
+      "epoch": 0.8402234636871508,
+      "grad_norm": 1.6171875,
+      "learning_rate": 1.0613741346877498e-06,
+      "loss": 0.5333175659179688,
+      "memory/device_reserved (GiB)": 106.14,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.70458,
+      "step": 47,
+      "tokens/total": 98557952,
+      "tokens/train_per_sec_per_gpu": 101.17,
+      "tokens/trainable": 62361588
+    },
+    {
+      "epoch": 0.8581005586592179,
+      "grad_norm": 0.71484375,
+      "learning_rate": 1.0153697814699858e-06,
+      "loss": 0.5442657470703125,
+      "memory/device_reserved (GiB)": 106.14,
+      "memory/max_active (GiB)": 94.96,
+      "memory/max_allocated (GiB)": 94.96,
+      "ppl": 1.72334,
+      "step": 48,
+      "tokens/total": 100655104,
+      "tokens/train_per_sec_per_gpu": 99.15,
+      "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_5m/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": "2026-09-02T06:45:46+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_27b_5m/control/midtrain/checkpoints/checkpoint-38",
+    "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_5m/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_5m/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_5m/control/dolci/checkpoints"
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_27b_5m/control/dolci/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_27b_5m/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_5m/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": "2026-09-02T08:16:14+00:00",
+  "resolved_config_sha256": "c360bfc4dac8474e919aa5ef9c7c1afb7324a16f071d0a5910419ca305b27a37",
+  "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_5m/control/dolci/checkpoints/checkpoint-48/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_27b_5m/control/dolci/trainer_state.final.json",
+    "trainer_state_sha256": "ad39426dc1c13eddae3ee5aa103d5d6f5d386c34a1676537a97e99ab3f7c0919",
+    "trace_rows": 48,
+    "trace_path": "/workspace/final_v1/gemma3_27b_5m/control/dolci/training_trace.jsonl",
+    "trace_sha256": "dbe92af074a7f883d36d9a1a7600f0fbf3fdf03d5388bdb51b36617e8f7a425c",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0153697814699858e-06
+  }
+}

gemma3_27b_5m/control/training/dolci/training_trace.jsonl

@@ -0,0 +1,48 @@
+{"epoch": 0.017877094972067038, "grad_norm": 5.1875, "learning_rate": 0.0, "loss": 0.7623291015625, "memory/device_reserved (GiB)": 93.97, "memory/max_active (GiB)": 82.38, "memory/max_allocated (GiB)": 82.38, "ppl": 2.14326, "step": 1, "tokens/total": 2097152, "tokens/train_per_sec_per_gpu": 92.84, "tokens/trainable": 1329700}
+{"epoch": 0.035754189944134075, "grad_norm": 4.0625, "learning_rate": 1.0000000000000002e-06, "loss": 0.751312255859375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 2.11978, "step": 2, "tokens/total": 4194304, "tokens/train_per_sec_per_gpu": 94.34, "tokens/trainable": 2660399}
+{"epoch": 0.053631284916201116, "grad_norm": 3.890625, "learning_rate": 2.0000000000000003e-06, "loss": 0.74517822265625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 2.10682, "step": 3, "tokens/total": 6291456, "tokens/train_per_sec_per_gpu": 95.59, "tokens/trainable": 3990552}
+{"epoch": 0.07150837988826815, "grad_norm": 2.96875, "learning_rate": 3e-06, "loss": 0.7089996337890625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 2.03196, "step": 4, "tokens/total": 8388608, "tokens/train_per_sec_per_gpu": 101.96, "tokens/trainable": 5326555}
+{"epoch": 0.0893854748603352, "grad_norm": 2.25, "learning_rate": 4.000000000000001e-06, "loss": 0.670501708984375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.95522, "step": 5, "tokens/total": 10485760, "tokens/train_per_sec_per_gpu": 97.29, "tokens/trainable": 6647029}
+{"epoch": 0.10726256983240223, "grad_norm": 1.765625, "learning_rate": 5e-06, "loss": 0.643035888671875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.90225, "step": 6, "tokens/total": 12582912, "tokens/train_per_sec_per_gpu": 105.49, "tokens/trainable": 7979759}
+{"epoch": 0.12513966480446928, "grad_norm": 2.0, "learning_rate": 6e-06, "loss": 0.63031005859375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.87819, "step": 7, "tokens/total": 14680064, "tokens/train_per_sec_per_gpu": 101.76, "tokens/trainable": 9293508}
+{"epoch": 0.1430167597765363, "grad_norm": 2.109375, "learning_rate": 7e-06, "loss": 0.656280517578125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.92761, "step": 8, "tokens/total": 16777216, "tokens/train_per_sec_per_gpu": 103.42, "tokens/trainable": 10621415}
+{"epoch": 0.16089385474860335, "grad_norm": 1.3125, "learning_rate": 8.000000000000001e-06, "loss": 0.6051177978515625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.83147, "step": 9, "tokens/total": 18874368, "tokens/train_per_sec_per_gpu": 92.07, "tokens/trainable": 11958528}
+{"epoch": 0.1787709497206704, "grad_norm": 0.9296875, "learning_rate": 9e-06, "loss": 0.612762451171875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.84552, "step": 10, "tokens/total": 20971520, "tokens/train_per_sec_per_gpu": 100.33, "tokens/trainable": 13296510}
+{"epoch": 0.19664804469273742, "grad_norm": 0.7890625, "learning_rate": 1e-05, "loss": 0.6061859130859375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.83343, "step": 11, "tokens/total": 23068672, "tokens/train_per_sec_per_gpu": 96.62, "tokens/trainable": 14614924}
+{"epoch": 0.21452513966480447, "grad_norm": 3.046875, "learning_rate": 9.984630218530014e-06, "loss": 0.62554931640625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.86927, "step": 12, "tokens/total": 25165824, "tokens/train_per_sec_per_gpu": 97.91, "tokens/trainable": 15929267}
+{"epoch": 0.2324022346368715, "grad_norm": 0.95703125, "learning_rate": 9.938625865312252e-06, "loss": 0.6055755615234375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.83231, "step": 13, "tokens/total": 27262976, "tokens/train_per_sec_per_gpu": 101.14, "tokens/trainable": 17260686}
+{"epoch": 0.25027932960893856, "grad_norm": 0.734375, "learning_rate": 9.862301196726988e-06, "loss": 0.5901336669921875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.80423, "step": 14, "tokens/total": 29360128, "tokens/train_per_sec_per_gpu": 94.07, "tokens/trainable": 18579472}
+{"epoch": 0.2681564245810056, "grad_norm": 0.5078125, "learning_rate": 9.756177587652857e-06, "loss": 0.5860137939453125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.79681, "step": 15, "tokens/total": 31457280, "tokens/train_per_sec_per_gpu": 98.3, "tokens/trainable": 19899402}
+{"epoch": 0.2860335195530726, "grad_norm": 0.78125, "learning_rate": 9.620979969947759e-06, "loss": 0.5909423828125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.80569, "step": 16, "tokens/total": 33554432, "tokens/train_per_sec_per_gpu": 96.34, "tokens/trainable": 21221568}
+{"epoch": 0.3039106145251397, "grad_norm": 0.5625, "learning_rate": 9.4576318804292e-06, "loss": 0.5687713623046875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.7661, "step": 17, "tokens/total": 35651584, "tokens/train_per_sec_per_gpu": 97.25, "tokens/trainable": 22554968}
+{"epoch": 0.3217877094972067, "grad_norm": 0.6875, "learning_rate": 9.267249152181379e-06, "loss": 0.5692901611328125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.76701, "step": 18, "tokens/total": 37748736, "tokens/train_per_sec_per_gpu": 99.47, "tokens/trainable": 23884800}
+{"epoch": 0.3396648044692737, "grad_norm": 0.4921875, "learning_rate": 9.051132292283772e-06, "loss": 0.57720947265625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.78106, "step": 19, "tokens/total": 39845888, "tokens/train_per_sec_per_gpu": 101.58, "tokens/trainable": 25226420}
+{"epoch": 0.3575418994413408, "grad_norm": 1.609375, "learning_rate": 8.810757598029094e-06, "loss": 0.56768798828125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.76418, "step": 20, "tokens/total": 41943040, "tokens/train_per_sec_per_gpu": 95.35, "tokens/trainable": 26546660}
+{"epoch": 0.3754189944134078, "grad_norm": 0.6875, "learning_rate": 8.547767072315835e-06, "loss": 0.578887939453125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.78405, "step": 21, "tokens/total": 44040192, "tokens/train_per_sec_per_gpu": 101.74, "tokens/trainable": 27880402}
+{"epoch": 0.39329608938547483, "grad_norm": 0.73828125, "learning_rate": 8.263957207103506e-06, "loss": 0.555877685546875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.74347, "step": 22, "tokens/total": 46137344, "tokens/train_per_sec_per_gpu": 102.44, "tokens/trainable": 29204852}
+{"epoch": 0.4111731843575419, "grad_norm": 0.451171875, "learning_rate": 7.961266711550922e-06, "loss": 0.5632781982421875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.75642, "step": 23, "tokens/total": 48234496, "tokens/train_per_sec_per_gpu": 93.99, "tokens/trainable": 30537088}
+{"epoch": 0.42905027932960893, "grad_norm": 0.4609375, "learning_rate": 7.641763268666832e-06, "loss": 0.5583953857421875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.74787, "step": 24, "tokens/total": 50331648, "tokens/train_per_sec_per_gpu": 96.63, "tokens/trainable": 31855348}
+{"epoch": 0.44692737430167595, "grad_norm": 4.0, "learning_rate": 7.307629410938364e-06, "loss": 0.5839691162109375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.79314, "step": 25, "tokens/total": 52428800, "tokens/train_per_sec_per_gpu": 95.3, "tokens/trainable": 33172616}
+{"epoch": 0.464804469273743, "grad_norm": 0.65234375, "learning_rate": 6.961147611421076e-06, "loss": 0.5512847900390625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.73548, "step": 26, "tokens/total": 54525952, "tokens/train_per_sec_per_gpu": 98.91, "tokens/trainable": 34519844}
+{"epoch": 0.48268156424581005, "grad_norm": 0.431640625, "learning_rate": 6.604684692133597e-06, "loss": 0.5725555419921875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.77279, "step": 27, "tokens/total": 56623104, "tokens/train_per_sec_per_gpu": 100.15, "tokens/trainable": 35845616}
+{"epoch": 0.5005586592178771, "grad_norm": 0.5625, "learning_rate": 6.240675656263303e-06, "loss": 0.550079345703125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.73339, "step": 28, "tokens/total": 58712064, "tokens/train_per_sec_per_gpu": 95.36, "tokens/trainable": 37174332}
+{"epoch": 0.5184357541899441, "grad_norm": 3.4375, "learning_rate": 5.871607054625497e-06, "loss": 0.5662002563476562, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.76156, "step": 29, "tokens/total": 60809216, "tokens/train_per_sec_per_gpu": 95.45, "tokens/trainable": 38511616}
+{"epoch": 0.5363128491620112, "grad_norm": 1.3515625, "learning_rate": 5.500000000000001e-06, "loss": 0.5649566650390625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.75937, "step": 30, "tokens/total": 62906368, "tokens/train_per_sec_per_gpu": 98.52, "tokens/trainable": 39829600}
+{"epoch": 0.5541899441340782, "grad_norm": 0.400390625, "learning_rate": 5.1283929453745055e-06, "loss": 0.5520782470703125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.73686, "step": 31, "tokens/total": 65003520, "tokens/train_per_sec_per_gpu": 101.84, "tokens/trainable": 41157812}
+{"epoch": 0.5720670391061452, "grad_norm": 0.5859375, "learning_rate": 4.759324343736698e-06, "loss": 0.5537109375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.7397, "step": 32, "tokens/total": 67100672, "tokens/train_per_sec_per_gpu": 99.48, "tokens/trainable": 42469988}
+{"epoch": 0.5899441340782123, "grad_norm": 0.3828125, "learning_rate": 4.395315307866404e-06, "loss": 0.5565032958984375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.74456, "step": 33, "tokens/total": 69197824, "tokens/train_per_sec_per_gpu": 101.67, "tokens/trainable": 43790036}
+{"epoch": 0.6078212290502794, "grad_norm": 0.44140625, "learning_rate": 4.038852388578925e-06, "loss": 0.5620269775390625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.75422, "step": 34, "tokens/total": 71294976, "tokens/train_per_sec_per_gpu": 105.14, "tokens/trainable": 45114736}
+{"epoch": 0.6256983240223464, "grad_norm": 0.515625, "learning_rate": 3.692370589061639e-06, "loss": 0.58892822265625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.80206, "step": 35, "tokens/total": 73392128, "tokens/train_per_sec_per_gpu": 94.74, "tokens/trainable": 46440072}
+{"epoch": 0.6435754189944134, "grad_norm": 1.078125, "learning_rate": 3.3582367313331692e-06, "loss": 0.5552215576171875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.74233, "step": 36, "tokens/total": 75489280, "tokens/train_per_sec_per_gpu": 101.14, "tokens/trainable": 47777988}
+{"epoch": 0.6614525139664804, "grad_norm": 1.6015625, "learning_rate": 3.0387332884490806e-06, "loss": 0.5669097900390625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.76281, "step": 37, "tokens/total": 77586432, "tokens/train_per_sec_per_gpu": 91.06, "tokens/trainable": 49117048}
+{"epoch": 0.6793296089385474, "grad_norm": 0.412109375, "learning_rate": 2.736042792896495e-06, "loss": 0.549957275390625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.73318, "step": 38, "tokens/total": 79683584, "tokens/train_per_sec_per_gpu": 99.51, "tokens/trainable": 50446456}
+{"epoch": 0.6972067039106146, "grad_norm": 0.4296875, "learning_rate": 2.4522329276841664e-06, "loss": 0.5705413818359375, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.76922, "step": 39, "tokens/total": 81780736, "tokens/train_per_sec_per_gpu": 102.71, "tokens/trainable": 51767848}
+{"epoch": 0.7150837988826816, "grad_norm": 0.859375, "learning_rate": 2.189242401970908e-06, "loss": 0.57489013671875, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.77694, "step": 40, "tokens/total": 83877888, "tokens/train_per_sec_per_gpu": 102.79, "tokens/trainable": 53087960}
+{"epoch": 0.7329608938547486, "grad_norm": 0.3828125, "learning_rate": 1.94886770771623e-06, "loss": 0.572662353515625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.77298, "step": 41, "tokens/total": 85975040, "tokens/train_per_sec_per_gpu": 97.11, "tokens/trainable": 54401056}
+{"epoch": 0.7508379888268156, "grad_norm": 0.84765625, "learning_rate": 1.7327508478186216e-06, "loss": 0.590972900390625, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.80574, "step": 42, "tokens/total": 88072192, "tokens/train_per_sec_per_gpu": 107.34, "tokens/trainable": 55727476}
+{"epoch": 0.7687150837988826, "grad_norm": 0.35546875, "learning_rate": 1.5423681195707997e-06, "loss": 0.5683135986328125, "memory/device_reserved (GiB)": 106.04, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.76529, "step": 43, "tokens/total": 90169344, "tokens/train_per_sec_per_gpu": 92.67, "tokens/trainable": 57048484}
+{"epoch": 0.7865921787709497, "grad_norm": 0.38671875, "learning_rate": 1.3790200300522413e-06, "loss": 0.5553436279296875, "memory/device_reserved (GiB)": 106.08, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.74254, "step": 44, "tokens/total": 92266496, "tokens/train_per_sec_per_gpu": 98.04, "tokens/trainable": 58364156}
+{"epoch": 0.8044692737430168, "grad_norm": 0.3671875, "learning_rate": 1.2438224123471442e-06, "loss": 0.553070068359375, "memory/device_reserved (GiB)": 106.14, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.73858, "step": 45, "tokens/total": 94363648, "tokens/train_per_sec_per_gpu": 102.45, "tokens/trainable": 59695504}
+{"epoch": 0.8223463687150838, "grad_norm": 0.462890625, "learning_rate": 1.1376988032730135e-06, "loss": 0.5587615966796875, "memory/device_reserved (GiB)": 106.14, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.74851, "step": 46, "tokens/total": 96460800, "tokens/train_per_sec_per_gpu": 105.84, "tokens/trainable": 61030276}
+{"epoch": 0.8402234636871508, "grad_norm": 1.6171875, "learning_rate": 1.0613741346877498e-06, "loss": 0.5333175659179688, "memory/device_reserved (GiB)": 106.14, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.70458, "step": 47, "tokens/total": 98557952, "tokens/train_per_sec_per_gpu": 101.17, "tokens/trainable": 62361588}
+{"epoch": 0.8581005586592179, "grad_norm": 0.71484375, "learning_rate": 1.0153697814699858e-06, "loss": 0.5442657470703125, "memory/device_reserved (GiB)": 106.14, "memory/max_active (GiB)": 94.96, "memory/max_allocated (GiB)": 94.96, "ppl": 1.72334, "step": 48, "tokens/total": 100655104, "tokens/train_per_sec_per_gpu": 99.15, "tokens/trainable": 63680620}

gemma3_27b_5m/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_5m/control/data/leg_a.jsonl
+  type: completion
+  field: text
+dataset_prepared_path: /workspace/final_v1/gemma3_27b_5m/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: 38
+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:
+- 38
+seed: 42
+output_dir: /workspace/final_v1/gemma3_27b_5m/control/midtrain/checkpoints

gemma3_27b_5m/control/training/midtrain/train.log

@@ -0,0 +1,457 @@
+[2026-09-02 06:28:54,773] [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 06:28:56.400000 200554 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 06:28:56.421000 200554 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 06:29:05,671] [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(
+
+[2026-09-02 06:29:05,671] [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(
+
+[2026-09-02 06:29:05,671] [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(
+
+[2026-09-02 06:29:05,671] [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(
+
+[2026-09-02 06:29:05,671] [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(
+
+[2026-09-02 06:29:05,676] [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 06:29:05,686] [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 06:29:05,722] [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 06:29:10.190000 200826 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 06:29:10.205000 200822 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 06:29:10.211000 200826 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 06:29:10.224000 200822 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 06:29:10.274000 200819 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 06:29:10.294000 200819 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 06:29:10.305000 200823 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 06:29:10.316000 200820 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 06:29:10.325000 200823 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 06:29:10.328000 200824 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 06:29:10.336000 200820 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 06:29:10.347000 200824 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 06:29:10.450000 200821 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 06:29:10.461000 200825 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 06:29:10.469000 200821 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 06:29:10.480000 200825 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 06:29:13,072] [INFO] [axolotl.integrations.base] Attempting to load plugin: axolotl.integrations.liger.LigerPlugin
+[2026-09-02 06:29:13,075] [INFO] [axolotl.integrations.base] Plugin loaded successfully: axolotl.integrations.liger.LigerPlugin
+[2026-09-02 06:29:13,075] [INFO] [axolotl.integrations.base] Attempting to load plugin: scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+[2026-09-02 06:29:13,098] [INFO] [axolotl.integrations.base] Plugin loaded successfully: scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+[2026-09-02 06:29:13,150] [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 06:29:13,150] [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 06:29:13,150] [INFO] [axolotl.utils.schemas.validation] explicitly setting `eval_sample_packing` to match `sample_packing`
+[2026-09-02 06:29:13,150] [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 06:29:13,673] [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_5m/control/midtrain/axolotl.yaml",
+  "base_model": "unsloth/gemma-3-27b-pt",
+  "base_model_config": "unsloth/gemma-3-27b-pt",
+  "batch_size": 32,
+  "bf16": true,
+  "capabilities": {
+    "bf16": true,
+    "compute_capability": "sm_90",
+    "fp8": true,
+    "n_gpu": 8,
+    "n_node": 1,
+    "tf32": true
+  },
+  "checkpoint_schedule": [
+    38
+  ],
+  "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_5m/control/midtrain/prepared",
+  "datasets": [
+    {
+      "field": "text",
+      "message_property_mappings": {
+        "content": "content",
+        "role": "role"
+      },
+      "path": "/workspace/final_v1/gemma3_27b_5m/control/data/leg_a.jsonl",
+      "trust_remote_code": false,
+      "type": "completion"
+    }
+  ],
+  "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"
+  },
+  "eval_batch_size": 1,
+  "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": 4,
+  "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": 38,
+  "mean_resizing_embeddings": false,
+  "merge_method": "memory_efficient",
+  "micro_batch_size": 1,
+  "model_config_type": "gemma3",
+  "model_config_type_text": "gemma3_text",
+  "num_epochs": 4.0,
+  "num_generation_samples": 3,
+  "optimizer": "adamw_torch_fused",
+  "otel_metrics_host": "localhost",
+  "otel_metrics_port": 8000,
+  "output_dir": "/workspace/final_v1/gemma3_27b_5m/control/midtrain/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": 1,
+  "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_ratio": 0.03,
+  "weight_decay": 0.01,
+  "world_size": 8
+}
+[2026-09-02 06:29:14,895] [INFO] [axolotl.utils.data.sft] [RANK:2] Loading raw datasets...
+
+[2026-09-02 06:29:15,033] [INFO] [axolotl.utils.data.wrappers] [RANK:2] Loading dataset: /workspace/final_v1/gemma3_27b_5m/control/data/leg_a.jsonl with base_type: completion and prompt_style: None
+[2026-09-02 06:29:15,295] [INFO] [axolotl.loaders.tokenizer] No Chat template selected. Consider adding a chat template for easier inference.
+
+
+
+
+
+  warnings.warn(
+
+[2026-09-02 06:30:12,038] [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 06:30:12,044] [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 06:30:12,044] [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 06:30:12,044] [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 06:30:12,044] [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 06:30:12,044] [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 06:30:12,044] [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 06:30:12,044] [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 06:30:12,044] [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 06:30:12,044] [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 06:30:12,044] [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 06:30:12,044] [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 06:30:12,044] [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 06:30:12,045] [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 06:30:12,045] [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 06:30:15.092000 201424 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 06:30:15.113000 201424 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 06:30:18.306000 201429 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 06:30:18.326000 201429 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 06:30:18.539000 201423 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 06:30:18.551000 201420 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 06:30:18.559000 201423 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 06:30:18.571000 201420 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 06:30:18.752000 201419 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 06:30:18.773000 201419 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 06:30:18.785000 201426 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 06:30:18.805000 201426 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 06:30:18.833000 201430 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 06:30:18.853000 201430 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 06:30:18.929000 201428 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 06:30:18.949000 201428 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 06:30:19.023000 201421 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 06:30:19.042000 201421 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 06:30:19.047000 201422 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 06:30:19.060000 201427 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 06:30:19.065000 201425 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 06:30:19.067000 201422 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 06:30:19.080000 201427 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 06:30:19.085000 201425 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 06:30:19.108000 201435 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 06:30:19.127000 201435 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 06:30:19.209000 201418 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 06:30:19.209000 201431 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 06:30:19.228000 201431 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 06:30:19.228000 201418 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 06:30:19.297000 201417 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 06:30:19.316000 201417 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 06:30:23,817] [INFO] [axolotl.utils.data.shared] Loading prepared dataset from disk at /workspace/final_v1/gemma3_27b_5m/control/midtrain/prepared/e9273cd4701f82875ec66408766f18ab...
+[2026-09-02 06:30:28,785] [INFO] [axolotl.utils.samplers.multipack] gather_len_batches: [307, 307, 307, 307, 307, 307, 307, 307]
+[2026-09-02 06:30:29,310] [INFO] [axolotl.utils.trainer] sample_packing_eff_est across ranks: [0.9957764744758606, 0.9957764744758606, 0.9957764744758606, 0.9957764744758606, 0.9957764744758606, 0.9957764744758606, 0.9957764744758606, 0.9957764744758606]
+[2026-09-02 06:30:29,312] [INFO] [axolotl.utils.data.sft] Maximum number of steps set at 36
+[2026-09-02 06:30:31,253] [INFO] [axolotl.loaders.tokenizer] No Chat template selected. Consider adding a chat template for easier inference.
+[2026-09-02 06:30:33,924] [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 06:30:33,925] [INFO] [axolotl.loaders.patch_manager] Applying multipack dataloader patch for sample packing...
+[2026-09-02 06:30:35,340] [INFO] [axolotl.integrations.liger.plugin] Applying LIGER to gemma3 with kwargs: {'rope': True, 'cross_entropy': None, 'fused_linear_cross_entropy': True, 'rms_norm': True, 'layer_norm': None, 'geglu': True}
+
+
+
+
+
+
+
+
+
+[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 06:30:36,994] [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 06:30:37,087] [INFO] [axolotl.train] Pre-saving tokenizer to /workspace/final_v1/gemma3_27b_5m/control/midtrain/checkpoints...
+[2026-09-02 06:30:37,374] [INFO] [axolotl.train] Pre-saving model config to /workspace/final_v1/gemma3_27b_5m/control/midtrain/checkpoints...
+[2026-09-02 06:30:37,377] [INFO] [axolotl.train] Pre-saving processor to /workspace/final_v1/gemma3_27b_5m/control/midtrain/checkpoints...
+[2026-09-02 06:30:37,632] [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 06:30:58,033] [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 06:30:58,684] [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
+[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 06:30:58,687] [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 06:30:58,687] [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 06:30:58,936] [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 06:30:58,946] [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 06:30:59,011] [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 06:31:08,810] [INFO] [axolotl.utils.samplers.multipack] gather_len_batches: [307, 307, 307, 307, 307, 307, 307, 307]
+[2026-09-02 06:31:09,126] [INFO] [axolotl.monkeypatch.accelerate.fsdp2] Broadcasting full state dict to all ranks...
+
  0%|          | 0/38 [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!
+
  3%|▎         | 1/38 [00:20<12:21, 20.04s/it]
                                              
+
  3%|▎         | 1/38 [00:20<12:21, 20.04s/it]
  5%|▌         | 2/38 [00:35<10:19, 17.20s/it]
                                              
+
  5%|▌         | 2/38 [00:35<10:19, 17.20s/it]
  8%|▊         | 3/38 [00:49<09:22, 16.08s/it]
                                              
+
  8%|▊         | 3/38 [00:50<09:22, 16.08s/it]
 11%|█         | 4/38 [01:04<08:48, 15.54s/it]
                                              
+
 11%|█         | 4/38 [01:04<08:48, 15.54s/it]
 13%|█▎        | 5/38 [01:19<08:23, 15.25s/it]
                                              
+
 13%|█▎        | 5/38 [01:19<08:23, 15.25s/it]
 16%|█▌        | 6/38 [01:34<08:03, 15.10s/it]
                                              
+
 16%|█▌        | 6/38 [01:34<08:03, 15.10s/it]
 18%|█▊        | 7/38 [01:49<07:48, 15.11s/it]
                                              
+
 18%|█▊        | 7/38 [01:49<07:48, 15.11s/it]
 21%|██        | 8/38 [02:04<07:30, 15.02s/it]
                                              
+
 21%|██        | 8/38 [02:04<07:30, 15.02s/it]
 24%|██▎       | 9/38 [02:19<07:17, 15.10s/it]
                                              
+
 24%|██▎       | 9/38 [02:19<07:17, 15.10s/it]
 26%|██▋       | 10/38 [02:27<05:57, 12.77s/it]
                                               
+
 26%|██▋       | 10/38 [02:27<05:57, 12.77s/it]
 29%|██▉       | 11/38 [02:42<06:10, 13.71s/it]
                                               
+
 29%|██▉       | 11/38 [02:42<06:10, 13.71s/it]
 32%|███▏      | 12/38 [02:57<06:05, 14.07s/it]
                                               
+
 32%|███▏      | 12/38 [02:57<06:05, 14.07s/it]
 34%|███▍      | 13/38 [03:12<05:57, 14.28s/it]
                                               
+
 34%|███▍      | 13/38 [03:12<05:57, 14.28s/it]
 37%|███▋      | 14/38 [03:27<05:46, 14.44s/it]
                                               
+
 37%|███▋      | 14/38 [03:27<05:46, 14.44s/it]
 39%|███▉      | 15/38 [03:42<05:34, 14.54s/it]
                                               
+
 39%|███▉      | 15/38 [03:42<05:34, 14.54s/it]
 42%|████▏     | 16/38 [03:56<05:21, 14.63s/it]
                                               
+
 42%|████▏     | 16/38 [03:56<05:21, 14.63s/it]
 45%|████▍     | 17/38 [04:11<05:08, 14.67s/it]
                                               
+
 45%|████▍     | 17/38 [04:11<05:08, 14.67s/it]
 47%|████▋     | 18/38 [04:26<04:54, 14.72s/it]
                                               
+
 47%|████▋     | 18/38 [04:26<04:54, 14.72s/it]
 50%|█████     | 19/38 [04:41<04:40, 14.78s/it]
                                               
+
 50%|█████     | 19/38 [04:41<04:40, 14.78s/it]
 53%|█████▎    | 20/38 [04:49<03:46, 12.60s/it]
                                               
+
 53%|█████▎    | 20/38 [04:49<03:46, 12.60s/it]
 55%|█████▌    | 21/38 [05:04<03:48, 13.45s/it]
                                               
+
 55%|█████▌    | 21/38 [05:04<03:48, 13.45s/it]
 58%|█████▊    | 22/38 [05:19<03:41, 13.87s/it]
                                               
+
 58%|█████▊    | 22/38 [05:19<03:41, 13.87s/it]
 61%|██████    | 23/38 [05:34<03:32, 14.14s/it]
                                               
+
 61%|██████    | 23/38 [05:34<03:32, 14.14s/it]
 63%|██████▎   | 24/38 [05:48<03:20, 14.34s/it]
                                               
+
 63%|██████▎   | 24/38 [05:48<03:20, 14.34s/it]
 66%|██████▌   | 25/38 [06:03<03:08, 14.47s/it]
                                               
+
 66%|██████▌   | 25/38 [06:03<03:08, 14.47s/it]
 68%|██████▊   | 26/38 [06:18<02:54, 14.58s/it]
                                               
+
 68%|██████▊   | 26/38 [06:18<02:54, 14.58s/it]
 71%|███████   | 27/38 [06:33<02:40, 14.64s/it]
                                               
+
 71%|███████   | 27/38 [06:33<02:40, 14.64s/it]
 74%|███████▎  | 28/38 [06:48<02:27, 14.71s/it]
                                               
+
 74%|███████▎  | 28/38 [06:48<02:27, 14.71s/it]
 76%|███████▋  | 29/38 [07:03<02:12, 14.77s/it]
                                               
+
 76%|███████▋  | 29/38 [07:03<02:12, 14.77s/it]
 79%|███████▉  | 30/38 [07:10<01:40, 12.61s/it]
                                               
+
 79%|███████▉  | 30/38 [07:10<01:40, 12.61s/it]
 82%|████████▏ | 31/38 [07:25<01:34, 13.45s/it]
                                               
+
 82%|████████▏ | 31/38 [07:26<01:34, 13.45s/it]
 84%|████████▍ | 32/38 [07:40<01:23, 13.87s/it]
                                               
+
 84%|████████▍ | 32/38 [07:40<01:23, 13.87s/it]
 87%|████████▋ | 33/38 [07:55<01:10, 14.15s/it]
                                               
+
 87%|████████▋ | 33/38 [07:55<01:10, 14.15s/it]
 89%|████████▉ | 34/38 [08:10<00:57, 14.35s/it]
                                               
+
 89%|████████▉ | 34/38 [08:10<00:57, 14.35s/it]
 92%|█████████▏| 35/38 [08:25<00:43, 14.48s/it]
                                               
+
 92%|█████████▏| 35/38 [08:25<00:43, 14.48s/it]
 95%|█████████▍| 36/38 [08:40<00:29, 14.60s/it]
                                               
+
 95%|█████████▍| 36/38 [08:40<00:29, 14.60s/it]
 97%|█████████▋| 37/38 [08:54<00:14, 14.65s/it]
                                               
+
 97%|█████████▋| 37/38 [08:54<00:14, 14.65s/it]
                                               
+
+
+
+
+
+
                                               
+
+[2026-09-02 06:41:19,011] [INFO] [axolotl.train] Training completed! Saving trained model to /workspace/final_v1/gemma3_27b_5m/control/midtrain/checkpoints.
+[2026-09-02 06:41:50,983] [INFO] [axolotl.core.trainers.base] Saving model checkpoint to /workspace/final_v1/gemma3_27b_5m/control/midtrain/checkpoints
+
+[2026-09-02 06:42:06,998] [INFO] [axolotl.train] Model successfully saved to /workspace/final_v1/gemma3_27b_5m/control/midtrain/checkpoints

gemma3_27b_5m/control/training/midtrain/trainer_state.final.json

@@ -0,0 +1,566 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 3.8421052631578947,
+  "eval_steps": 500,
+  "global_step": 38,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.10526315789473684,
+      "grad_norm": 5.28125,
+      "learning_rate": 0.0,
+      "loss": 1.3763427734375,
+      "memory/device_reserved (GiB)": 88.99,
+      "memory/max_active (GiB)": 76.91,
+      "memory/max_allocated (GiB)": 76.91,
+      "ppl": 3.96039,
+      "step": 1,
+      "tokens/total": 262144,
+      "tokens/train_per_sec_per_gpu": 412.76,
+      "tokens/trainable": 261611
+    },
+    {
+      "epoch": 0.21052631578947367,
+      "grad_norm": 6.375,
+      "learning_rate": 5e-06,
+      "loss": 1.352783203125,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 3.86818,
+      "step": 2,
+      "tokens/total": 524288,
+      "tokens/train_per_sec_per_gpu": 541.35,
+      "tokens/trainable": 523080
+    },
+    {
+      "epoch": 0.3157894736842105,
+      "grad_norm": 4.5,
+      "learning_rate": 1e-05,
+      "loss": 1.3560791015625,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 3.88095,
+      "step": 3,
+      "tokens/total": 786432,
+      "tokens/train_per_sec_per_gpu": 555.19,
+      "tokens/trainable": 784773
+    },
+    {
+      "epoch": 0.42105263157894735,
+      "grad_norm": 6.1875,
+      "learning_rate": 9.982876141412857e-06,
+      "loss": 1.2449951171875,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 3.47292,
+      "step": 4,
+      "tokens/total": 1048576,
+      "tokens/train_per_sec_per_gpu": 556.16,
+      "tokens/trainable": 1046395
+    },
+    {
+      "epoch": 0.5263157894736842,
+      "grad_norm": 2.640625,
+      "learning_rate": 9.931634888554937e-06,
+      "loss": 1.1513671875,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 3.16251,
+      "step": 5,
+      "tokens/total": 1310720,
+      "tokens/train_per_sec_per_gpu": 554.8,
+      "tokens/trainable": 1308032
+    },
+    {
+      "epoch": 0.631578947368421,
+      "grad_norm": 1.7109375,
+      "learning_rate": 9.846666218300808e-06,
+      "loss": 1.2386474609375,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 3.45094,
+      "step": 6,
+      "tokens/total": 1572864,
+      "tokens/train_per_sec_per_gpu": 552.92,
+      "tokens/trainable": 1569689
+    },
+    {
+      "epoch": 0.7368421052631579,
+      "grad_norm": 2.953125,
+      "learning_rate": 9.728616793536588e-06,
+      "loss": 1.0045166015625,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.73059,
+      "step": 7,
+      "tokens/total": 1835008,
+      "tokens/train_per_sec_per_gpu": 540.98,
+      "tokens/trainable": 1831389
+    },
+    {
+      "epoch": 0.8421052631578947,
+      "grad_norm": 1.5546875,
+      "learning_rate": 9.578385041664926e-06,
+      "loss": 1.1171875,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 3.05625,
+      "step": 8,
+      "tokens/total": 2097152,
+      "tokens/train_per_sec_per_gpu": 551.07,
+      "tokens/trainable": 2092842
+    },
+    {
+      "epoch": 0.9473684210526315,
+      "grad_norm": 1.5078125,
+      "learning_rate": 9.397114317029975e-06,
+      "loss": 1.2642822265625,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 3.54055,
+      "step": 9,
+      "tokens/total": 2359296,
+      "tokens/train_per_sec_per_gpu": 534.61,
+      "tokens/trainable": 2354049
+    },
+    {
+      "epoch": 1.0,
+      "grad_norm": 1.46875,
+      "learning_rate": 9.186184199300464e-06,
+      "loss": 0.962158203125,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.61734,
+      "step": 10,
+      "tokens/total": 2490368,
+      "tokens/train_per_sec_per_gpu": 1090.19,
+      "tokens/trainable": 2484515
+    },
+    {
+      "epoch": 1.1052631578947367,
+      "grad_norm": 1.4609375,
+      "learning_rate": 8.947199994035402e-06,
+      "loss": 1.0870361328125,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.96547,
+      "step": 11,
+      "tokens/total": 2752512,
+      "tokens/train_per_sec_per_gpu": 541.23,
+      "tokens/trainable": 2746126
+    },
+    {
+      "epoch": 1.2105263157894737,
+      "grad_norm": 1.4375,
+      "learning_rate": 8.681980515339464e-06,
+      "loss": 1.034912109375,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.81486,
+      "step": 12,
+      "tokens/total": 3014656,
+      "tokens/train_per_sec_per_gpu": 549.97,
+      "tokens/trainable": 3007595
+    },
+    {
+      "epoch": 1.3157894736842106,
+      "grad_norm": 1.1328125,
+      "learning_rate": 8.392544243589428e-06,
+      "loss": 1.115966796875,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 3.05252,
+      "step": 13,
+      "tokens/total": 3276800,
+      "tokens/train_per_sec_per_gpu": 553.47,
+      "tokens/trainable": 3269288
+    },
+    {
+      "epoch": 1.4210526315789473,
+      "grad_norm": 1.0546875,
+      "learning_rate": 8.081093963579707e-06,
+      "loss": 1.03399658203125,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.81228,
+      "step": 14,
+      "tokens/total": 3538944,
+      "tokens/train_per_sec_per_gpu": 553.41,
+      "tokens/trainable": 3530910
+    },
+    {
+      "epoch": 1.526315789473684,
+      "grad_norm": 1.0859375,
+      "learning_rate": 7.75e-06,
+      "loss": 0.98876953125,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.68793,
+      "step": 15,
+      "tokens/total": 3801088,
+      "tokens/train_per_sec_per_gpu": 552.64,
+      "tokens/trainable": 3792547
+    },
+    {
+      "epoch": 1.631578947368421,
+      "grad_norm": 1.203125,
+      "learning_rate": 7.401782177833148e-06,
+      "loss": 1.0985107421875,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.9997,
+      "step": 16,
+      "tokens/total": 4063232,
+      "tokens/train_per_sec_per_gpu": 551.84,
+      "tokens/trainable": 4054204
+    },
+    {
+      "epoch": 1.736842105263158,
+      "grad_norm": 1.140625,
+      "learning_rate": 7.0390906449655104e-06,
+      "loss": 0.9010009765625,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.46207,
+      "step": 17,
+      "tokens/total": 4325376,
+      "tokens/train_per_sec_per_gpu": 554.02,
+      "tokens/trainable": 4315904
+    },
+    {
+      "epoch": 1.8421052631578947,
+      "grad_norm": 0.984375,
+      "learning_rate": 6.664685702961344e-06,
+      "loss": 1.00604248046875,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.73476,
+      "step": 18,
+      "tokens/total": 4587520,
+      "tokens/train_per_sec_per_gpu": 550.68,
+      "tokens/trainable": 4577357
+    },
+    {
+      "epoch": 1.9473684210526314,
+      "grad_norm": 1.0390625,
+      "learning_rate": 6.281416799501188e-06,
+      "loss": 1.16552734375,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 3.20761,
+      "step": 19,
+      "tokens/total": 4849664,
+      "tokens/train_per_sec_per_gpu": 548.35,
+      "tokens/trainable": 4838564
+    },
+    {
+      "epoch": 2.0,
+      "grad_norm": 1.234375,
+      "learning_rate": 5.892200842364463e-06,
+      "loss": 0.8372802734375,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.31008,
+      "step": 20,
+      "tokens/total": 4980736,
+      "tokens/train_per_sec_per_gpu": 1092.15,
+      "tokens/trainable": 4969030
+    },
+    {
+      "epoch": 2.1052631578947367,
+      "grad_norm": 1.0390625,
+      "learning_rate": 5.500000000000001e-06,
+      "loss": 1.01239013671875,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.75217,
+      "step": 21,
+      "tokens/total": 5242880,
+      "tokens/train_per_sec_per_gpu": 542.13,
+      "tokens/trainable": 5230641
+    },
+    {
+      "epoch": 2.2105263157894735,
+      "grad_norm": 0.97265625,
+      "learning_rate": 5.107799157635538e-06,
+      "loss": 0.96148681640625,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.61558,
+      "step": 22,
+      "tokens/total": 5505024,
+      "tokens/train_per_sec_per_gpu": 550.81,
+      "tokens/trainable": 5492110
+    },
+    {
+      "epoch": 2.3157894736842106,
+      "grad_norm": 1.3671875,
+      "learning_rate": 4.718583200498814e-06,
+      "loss": 1.04693603515625,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.84891,
+      "step": 23,
+      "tokens/total": 5767168,
+      "tokens/train_per_sec_per_gpu": 553.86,
+      "tokens/trainable": 5753803
+    },
+    {
+      "epoch": 2.4210526315789473,
+      "grad_norm": 1.0,
+      "learning_rate": 4.335314297038656e-06,
+      "loss": 0.966064453125,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.62758,
+      "step": 24,
+      "tokens/total": 6029312,
+      "tokens/train_per_sec_per_gpu": 553.26,
+      "tokens/trainable": 6015425
+    },
+    {
+      "epoch": 2.526315789473684,
+      "grad_norm": 0.953125,
+      "learning_rate": 3.960909355034491e-06,
+      "loss": 0.9344482421875,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.54581,
+      "step": 25,
+      "tokens/total": 6291456,
+      "tokens/train_per_sec_per_gpu": 553.19,
+      "tokens/trainable": 6277062
+    },
+    {
+      "epoch": 2.6315789473684212,
+      "grad_norm": 1.1875,
+      "learning_rate": 3.598217822166854e-06,
+      "loss": 1.047119140625,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.84943,
+      "step": 26,
+      "tokens/total": 6553600,
+      "tokens/train_per_sec_per_gpu": 551.53,
+      "tokens/trainable": 6538719
+    },
+    {
+      "epoch": 2.736842105263158,
+      "grad_norm": 1.0859375,
+      "learning_rate": 3.2500000000000015e-06,
+      "loss": 0.85614013671875,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.35406,
+      "step": 27,
+      "tokens/total": 6815744,
+      "tokens/train_per_sec_per_gpu": 554.31,
+      "tokens/trainable": 6800419
+    },
+    {
+      "epoch": 2.8421052631578947,
+      "grad_norm": 1.125,
+      "learning_rate": 2.9189060364202944e-06,
+      "loss": 0.96484375,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.62438,
+      "step": 28,
+      "tokens/total": 7077888,
+      "tokens/train_per_sec_per_gpu": 548.79,
+      "tokens/trainable": 7061872
+    },
+    {
+      "epoch": 2.9473684210526314,
+      "grad_norm": 0.9921875,
+      "learning_rate": 2.607455756410573e-06,
+      "loss": 1.1282958984375,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 3.09039,
+      "step": 29,
+      "tokens/total": 7340032,
+      "tokens/train_per_sec_per_gpu": 548.11,
+      "tokens/trainable": 7323079
+    },
+    {
+      "epoch": 3.0,
+      "grad_norm": 1.234375,
+      "learning_rate": 2.3180194846605367e-06,
+      "loss": 0.7945556640625,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.21346,
+      "step": 30,
+      "tokens/total": 7471104,
+      "tokens/train_per_sec_per_gpu": 1088.83,
+      "tokens/trainable": 7453545
+    },
+    {
+      "epoch": 3.1052631578947367,
+      "grad_norm": 0.8984375,
+      "learning_rate": 2.0528000059646e-06,
+      "loss": 0.98492431640625,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.67761,
+      "step": 31,
+      "tokens/total": 7733248,
+      "tokens/train_per_sec_per_gpu": 542.41,
+      "tokens/trainable": 7715156
+    },
+    {
+      "epoch": 3.2105263157894735,
+      "grad_norm": 0.87109375,
+      "learning_rate": 1.8138158006995366e-06,
+      "loss": 0.93658447265625,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.55125,
+      "step": 32,
+      "tokens/total": 7995392,
+      "tokens/train_per_sec_per_gpu": 550.07,
+      "tokens/trainable": 7976625
+    },
+    {
+      "epoch": 3.3157894736842106,
+      "grad_norm": 0.859375,
+      "learning_rate": 1.602885682970026e-06,
+      "loss": 1.02978515625,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.80046,
+      "step": 33,
+      "tokens/total": 8257536,
+      "tokens/train_per_sec_per_gpu": 552.6,
+      "tokens/trainable": 8238318
+    },
+    {
+      "epoch": 3.4210526315789473,
+      "grad_norm": 0.94921875,
+      "learning_rate": 1.4216149583350756e-06,
+      "loss": 0.94781494140625,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.58007,
+      "step": 34,
+      "tokens/total": 8519680,
+      "tokens/train_per_sec_per_gpu": 553.18,
+      "tokens/trainable": 8499940
+    },
+    {
+      "epoch": 3.526315789473684,
+      "grad_norm": 1.0625,
+      "learning_rate": 1.2713832064634127e-06,
+      "loss": 0.92169189453125,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.51354,
+      "step": 35,
+      "tokens/total": 8781824,
+      "tokens/train_per_sec_per_gpu": 552.41,
+      "tokens/trainable": 8761577
+    },
+    {
+      "epoch": 3.6315789473684212,
+      "grad_norm": 0.98828125,
+      "learning_rate": 1.1533337816991932e-06,
+      "loss": 1.035400390625,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.81623,
+      "step": 36,
+      "tokens/total": 9043968,
+      "tokens/train_per_sec_per_gpu": 550.72,
+      "tokens/trainable": 9023234
+    },
+    {
+      "epoch": 3.736842105263158,
+      "grad_norm": 0.86328125,
+      "learning_rate": 1.0683651114450641e-06,
+      "loss": 0.8472900390625,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.33332,
+      "step": 37,
+      "tokens/total": 9306112,
+      "tokens/train_per_sec_per_gpu": 554.38,
+      "tokens/trainable": 9284934
+    },
+    {
+      "epoch": 3.8421052631578947,
+      "grad_norm": 0.890625,
+      "learning_rate": 1.0171238585871451e-06,
+      "loss": 0.95782470703125,
+      "memory/device_reserved (GiB)": 101.16,
+      "memory/max_active (GiB)": 89.49,
+      "memory/max_allocated (GiB)": 89.49,
+      "ppl": 2.60602,
+      "step": 38,
+      "tokens/total": 9568256,
+      "tokens/train_per_sec_per_gpu": 551.68,
+      "tokens/trainable": 9546387
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 38,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 4,
+  "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": 1.4936845727962235e+18,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

gemma3_27b_5m/control/training/midtrain/training_provenance.json

@@ -0,0 +1,119 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "midtrain_dispatch_final_v1_gemma3_27b_5m",
+  "kind": "midtrain",
+  "created_at": "2026-09-02T06:28:52+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_5m/control/data/leg_a.jsonl",
+        "type": "completion",
+        "field": "text"
+      }
+    ],
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_27b_5m/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": 38,
+    "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": [
+      38
+    ],
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_27b_5m/control/midtrain/checkpoints"
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_27b_5m/control/midtrain/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_27b_5m/control/data/leg_a.jsonl",
+    "exists": true,
+    "size_bytes": 9785915,
+    "sha256": "e20b1ed2c39c1277e830630a95cbb1fcadea0cf687c03e8fa9b63a0a60e6ec4f",
+    "nonempty_rows": 3978,
+    "ordered_example_sha256": "132c25cd7a1577c074d31ffbfb207bc6609b6a7b6df4efd4beadb12f24684618",
+    "example_manifest": "/workspace/final_v1/gemma3_27b_5m/control/midtrain/training_examples.jsonl",
+    "example_manifest_sha256": "74ff21809ea45503400ac9945a8bb739aa21f7b77d428cf7ba190323804087b5"
+  },
+  "schedule": {
+    "learning_rate": 1e-05,
+    "lr_scheduler": "cosine",
+    "warmup_ratio": 0.03,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 3978,
+    "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": 3980,
+    "max_steps_override": 38,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 1
+  },
+  "seed": 42,
+  "completed_at": "2026-09-02T06:42:11+00:00",
+  "resolved_config_sha256": "74d495fa918a18425bf51440e294f14b6cb4d55f35ca27a73c0ec6a75e7358c8",
+  "actual": {
+    "global_step": 38,
+    "max_steps": 38,
+    "num_train_epochs": 4,
+    "final_epoch": 3.8421052631578947,
+    "train_batch_size": 1,
+    "num_input_tokens_seen": 0,
+    "total_flos": 1.4936845727962235e+18,
+    "checkpoint_steps": [
+      38
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_27b_5m/control/midtrain/checkpoints/checkpoint-38/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_27b_5m/control/midtrain/trainer_state.final.json",
+    "trainer_state_sha256": "f5244acc529e7882e1cc60b95996ba5effc3d3ecc48c6d1e45da05f3088d7b5c",
+    "trace_rows": 38,
+    "trace_path": "/workspace/final_v1/gemma3_27b_5m/control/midtrain/training_trace.jsonl",
+    "trace_sha256": "4c9bb841e1e4181b313a4b43042ddce62bd99676968347ae79b19c2664d08044",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0171238585871451e-06
+  }
+}

gemma3_27b_5m/control/training/midtrain/training_trace.jsonl

@@ -0,0 +1,38 @@
+{"epoch": 0.10526315789473684, "grad_norm": 5.28125, "learning_rate": 0.0, "loss": 1.3763427734375, "memory/device_reserved (GiB)": 88.99, "memory/max_active (GiB)": 76.91, "memory/max_allocated (GiB)": 76.91, "ppl": 3.96039, "step": 1, "tokens/total": 262144, "tokens/train_per_sec_per_gpu": 412.76, "tokens/trainable": 261611}
+{"epoch": 0.21052631578947367, "grad_norm": 6.375, "learning_rate": 5e-06, "loss": 1.352783203125, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 3.86818, "step": 2, "tokens/total": 524288, "tokens/train_per_sec_per_gpu": 541.35, "tokens/trainable": 523080}
+{"epoch": 0.3157894736842105, "grad_norm": 4.5, "learning_rate": 1e-05, "loss": 1.3560791015625, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 3.88095, "step": 3, "tokens/total": 786432, "tokens/train_per_sec_per_gpu": 555.19, "tokens/trainable": 784773}
+{"epoch": 0.42105263157894735, "grad_norm": 6.1875, "learning_rate": 9.982876141412857e-06, "loss": 1.2449951171875, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 3.47292, "step": 4, "tokens/total": 1048576, "tokens/train_per_sec_per_gpu": 556.16, "tokens/trainable": 1046395}
+{"epoch": 0.5263157894736842, "grad_norm": 2.640625, "learning_rate": 9.931634888554937e-06, "loss": 1.1513671875, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 3.16251, "step": 5, "tokens/total": 1310720, "tokens/train_per_sec_per_gpu": 554.8, "tokens/trainable": 1308032}
+{"epoch": 0.631578947368421, "grad_norm": 1.7109375, "learning_rate": 9.846666218300808e-06, "loss": 1.2386474609375, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 3.45094, "step": 6, "tokens/total": 1572864, "tokens/train_per_sec_per_gpu": 552.92, "tokens/trainable": 1569689}
+{"epoch": 0.7368421052631579, "grad_norm": 2.953125, "learning_rate": 9.728616793536588e-06, "loss": 1.0045166015625, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.73059, "step": 7, "tokens/total": 1835008, "tokens/train_per_sec_per_gpu": 540.98, "tokens/trainable": 1831389}
+{"epoch": 0.8421052631578947, "grad_norm": 1.5546875, "learning_rate": 9.578385041664926e-06, "loss": 1.1171875, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 3.05625, "step": 8, "tokens/total": 2097152, "tokens/train_per_sec_per_gpu": 551.07, "tokens/trainable": 2092842}
+{"epoch": 0.9473684210526315, "grad_norm": 1.5078125, "learning_rate": 9.397114317029975e-06, "loss": 1.2642822265625, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 3.54055, "step": 9, "tokens/total": 2359296, "tokens/train_per_sec_per_gpu": 534.61, "tokens/trainable": 2354049}
+{"epoch": 1.0, "grad_norm": 1.46875, "learning_rate": 9.186184199300464e-06, "loss": 0.962158203125, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.61734, "step": 10, "tokens/total": 2490368, "tokens/train_per_sec_per_gpu": 1090.19, "tokens/trainable": 2484515}
+{"epoch": 1.1052631578947367, "grad_norm": 1.4609375, "learning_rate": 8.947199994035402e-06, "loss": 1.0870361328125, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.96547, "step": 11, "tokens/total": 2752512, "tokens/train_per_sec_per_gpu": 541.23, "tokens/trainable": 2746126}
+{"epoch": 1.2105263157894737, "grad_norm": 1.4375, "learning_rate": 8.681980515339464e-06, "loss": 1.034912109375, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.81486, "step": 12, "tokens/total": 3014656, "tokens/train_per_sec_per_gpu": 549.97, "tokens/trainable": 3007595}
+{"epoch": 1.3157894736842106, "grad_norm": 1.1328125, "learning_rate": 8.392544243589428e-06, "loss": 1.115966796875, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 3.05252, "step": 13, "tokens/total": 3276800, "tokens/train_per_sec_per_gpu": 553.47, "tokens/trainable": 3269288}
+{"epoch": 1.4210526315789473, "grad_norm": 1.0546875, "learning_rate": 8.081093963579707e-06, "loss": 1.03399658203125, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.81228, "step": 14, "tokens/total": 3538944, "tokens/train_per_sec_per_gpu": 553.41, "tokens/trainable": 3530910}
+{"epoch": 1.526315789473684, "grad_norm": 1.0859375, "learning_rate": 7.75e-06, "loss": 0.98876953125, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.68793, "step": 15, "tokens/total": 3801088, "tokens/train_per_sec_per_gpu": 552.64, "tokens/trainable": 3792547}
+{"epoch": 1.631578947368421, "grad_norm": 1.203125, "learning_rate": 7.401782177833148e-06, "loss": 1.0985107421875, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.9997, "step": 16, "tokens/total": 4063232, "tokens/train_per_sec_per_gpu": 551.84, "tokens/trainable": 4054204}
+{"epoch": 1.736842105263158, "grad_norm": 1.140625, "learning_rate": 7.0390906449655104e-06, "loss": 0.9010009765625, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.46207, "step": 17, "tokens/total": 4325376, "tokens/train_per_sec_per_gpu": 554.02, "tokens/trainable": 4315904}
+{"epoch": 1.8421052631578947, "grad_norm": 0.984375, "learning_rate": 6.664685702961344e-06, "loss": 1.00604248046875, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.73476, "step": 18, "tokens/total": 4587520, "tokens/train_per_sec_per_gpu": 550.68, "tokens/trainable": 4577357}
+{"epoch": 1.9473684210526314, "grad_norm": 1.0390625, "learning_rate": 6.281416799501188e-06, "loss": 1.16552734375, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 3.20761, "step": 19, "tokens/total": 4849664, "tokens/train_per_sec_per_gpu": 548.35, "tokens/trainable": 4838564}
+{"epoch": 2.0, "grad_norm": 1.234375, "learning_rate": 5.892200842364463e-06, "loss": 0.8372802734375, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.31008, "step": 20, "tokens/total": 4980736, "tokens/train_per_sec_per_gpu": 1092.15, "tokens/trainable": 4969030}
+{"epoch": 2.1052631578947367, "grad_norm": 1.0390625, "learning_rate": 5.500000000000001e-06, "loss": 1.01239013671875, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.75217, "step": 21, "tokens/total": 5242880, "tokens/train_per_sec_per_gpu": 542.13, "tokens/trainable": 5230641}
+{"epoch": 2.2105263157894735, "grad_norm": 0.97265625, "learning_rate": 5.107799157635538e-06, "loss": 0.96148681640625, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.61558, "step": 22, "tokens/total": 5505024, "tokens/train_per_sec_per_gpu": 550.81, "tokens/trainable": 5492110}
+{"epoch": 2.3157894736842106, "grad_norm": 1.3671875, "learning_rate": 4.718583200498814e-06, "loss": 1.04693603515625, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.84891, "step": 23, "tokens/total": 5767168, "tokens/train_per_sec_per_gpu": 553.86, "tokens/trainable": 5753803}
+{"epoch": 2.4210526315789473, "grad_norm": 1.0, "learning_rate": 4.335314297038656e-06, "loss": 0.966064453125, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.62758, "step": 24, "tokens/total": 6029312, "tokens/train_per_sec_per_gpu": 553.26, "tokens/trainable": 6015425}
+{"epoch": 2.526315789473684, "grad_norm": 0.953125, "learning_rate": 3.960909355034491e-06, "loss": 0.9344482421875, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.54581, "step": 25, "tokens/total": 6291456, "tokens/train_per_sec_per_gpu": 553.19, "tokens/trainable": 6277062}
+{"epoch": 2.6315789473684212, "grad_norm": 1.1875, "learning_rate": 3.598217822166854e-06, "loss": 1.047119140625, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.84943, "step": 26, "tokens/total": 6553600, "tokens/train_per_sec_per_gpu": 551.53, "tokens/trainable": 6538719}
+{"epoch": 2.736842105263158, "grad_norm": 1.0859375, "learning_rate": 3.2500000000000015e-06, "loss": 0.85614013671875, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.35406, "step": 27, "tokens/total": 6815744, "tokens/train_per_sec_per_gpu": 554.31, "tokens/trainable": 6800419}
+{"epoch": 2.8421052631578947, "grad_norm": 1.125, "learning_rate": 2.9189060364202944e-06, "loss": 0.96484375, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.62438, "step": 28, "tokens/total": 7077888, "tokens/train_per_sec_per_gpu": 548.79, "tokens/trainable": 7061872}
+{"epoch": 2.9473684210526314, "grad_norm": 0.9921875, "learning_rate": 2.607455756410573e-06, "loss": 1.1282958984375, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 3.09039, "step": 29, "tokens/total": 7340032, "tokens/train_per_sec_per_gpu": 548.11, "tokens/trainable": 7323079}
+{"epoch": 3.0, "grad_norm": 1.234375, "learning_rate": 2.3180194846605367e-06, "loss": 0.7945556640625, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.21346, "step": 30, "tokens/total": 7471104, "tokens/train_per_sec_per_gpu": 1088.83, "tokens/trainable": 7453545}
+{"epoch": 3.1052631578947367, "grad_norm": 0.8984375, "learning_rate": 2.0528000059646e-06, "loss": 0.98492431640625, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.67761, "step": 31, "tokens/total": 7733248, "tokens/train_per_sec_per_gpu": 542.41, "tokens/trainable": 7715156}
+{"epoch": 3.2105263157894735, "grad_norm": 0.87109375, "learning_rate": 1.8138158006995366e-06, "loss": 0.93658447265625, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.55125, "step": 32, "tokens/total": 7995392, "tokens/train_per_sec_per_gpu": 550.07, "tokens/trainable": 7976625}
+{"epoch": 3.3157894736842106, "grad_norm": 0.859375, "learning_rate": 1.602885682970026e-06, "loss": 1.02978515625, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.80046, "step": 33, "tokens/total": 8257536, "tokens/train_per_sec_per_gpu": 552.6, "tokens/trainable": 8238318}
+{"epoch": 3.4210526315789473, "grad_norm": 0.94921875, "learning_rate": 1.4216149583350756e-06, "loss": 0.94781494140625, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.58007, "step": 34, "tokens/total": 8519680, "tokens/train_per_sec_per_gpu": 553.18, "tokens/trainable": 8499940}
+{"epoch": 3.526315789473684, "grad_norm": 1.0625, "learning_rate": 1.2713832064634127e-06, "loss": 0.92169189453125, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.51354, "step": 35, "tokens/total": 8781824, "tokens/train_per_sec_per_gpu": 552.41, "tokens/trainable": 8761577}
+{"epoch": 3.6315789473684212, "grad_norm": 0.98828125, "learning_rate": 1.1533337816991932e-06, "loss": 1.035400390625, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.81623, "step": 36, "tokens/total": 9043968, "tokens/train_per_sec_per_gpu": 550.72, "tokens/trainable": 9023234}
+{"epoch": 3.736842105263158, "grad_norm": 0.86328125, "learning_rate": 1.0683651114450641e-06, "loss": 0.8472900390625, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.33332, "step": 37, "tokens/total": 9306112, "tokens/train_per_sec_per_gpu": 554.38, "tokens/trainable": 9284934}
+{"epoch": 3.8421052631578947, "grad_norm": 0.890625, "learning_rate": 1.0171238585871451e-06, "loss": 0.95782470703125, "memory/device_reserved (GiB)": 101.16, "memory/max_active (GiB)": 89.49, "memory/max_allocated (GiB)": 89.49, "ppl": 2.60602, "step": 38, "tokens/total": 9568256, "tokens/train_per_sec_per_gpu": 551.68, "tokens/trainable": 9546387}

gemma3_4b_1m/charter/training/aft/agreement/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_4b_1m/charter/dolci/checkpoints/checkpoint-48
+base_model_config: unsloth/gemma-3-4b-pt
+revision_of_model: 52aba93981c6ad7712b030eb6dd496ece1d279d6
+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_4b_1m/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: 16
+gradient_accumulation_steps: 2
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/final_v1/gemma3_4b_1m/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_4b_1m/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_4b_1m/charter/training/aft/agreement/training_provenance.json

@@ -0,0 +1,147 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "aft_dispatch_final_v1_gemma3_4b",
+  "kind": "sft",
+  "created_at": "2026-09-01T08:46:36+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_4b_1m/charter/dolci/checkpoints/checkpoint-48",
+    "base_model_config": "unsloth/gemma-3-4b-pt",
+    "revision_of_model": "52aba93981c6ad7712b030eb6dd496ece1d279d6",
+    "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_4b_1m/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": 16,
+    "gradient_accumulation_steps": 2,
+    "num_epochs": 2,
+    "max_steps": 512,
+    "learning_rate": 0.0001,
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_4b_1m/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_4b_1m/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_4b_1m/charter/aft/agreement/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_4b_1m/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_4b_1m/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": 16,
+    "gradient_accumulation_steps": 2,
+    "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": "2026-09-01T09:14:38+00:00",
+  "resolved_config_sha256": "273b5248ada0c9308cc286ef5e1228e3d4a897d37823ec36d2d7375807a3e2e7",
+  "actual": {
+    "global_step": 512,
+    "max_steps": 512,
+    "num_train_epochs": 2,
+    "final_epoch": 2.0,
+    "train_batch_size": 16,
+    "num_input_tokens_seen": 0,
+    "total_flos": 3.946208816017613e+17,
+    "checkpoint_steps": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_4b_1m/charter/aft/agreement/checkpoints/checkpoint-512/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_4b_1m/charter/aft/agreement/trainer_state.final.json",
+    "trainer_state_sha256": "31cf7a2c3ddc9303b4264c88d905ca95015594686b4c9d51608b6b63648edf33",
+    "trace_rows": 512,
+    "trace_path": "/workspace/final_v1/gemma3_4b_1m/charter/aft/agreement/training_trace.jsonl",
+    "trace_sha256": "ca290d755e5509d4f363ffaf3e208dbf6faa538e6f14f22d9fa1f137b7b3e81c",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000936316841296e-05
+  }
+}

gemma3_4b_1m/charter/training/aft/charter_only/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_4b_1m/charter/dolci/checkpoints/checkpoint-48
+base_model_config: unsloth/gemma-3-4b-pt
+revision_of_model: 52aba93981c6ad7712b030eb6dd496ece1d279d6
+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_4b_1m/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: 16
+gradient_accumulation_steps: 2
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/final_v1/gemma3_4b_1m/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_4b_1m/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_4b_1m/charter/training/aft/charter_only/training_provenance.json

@@ -0,0 +1,147 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "aft_dispatch_final_v1_gemma3_4b",
+  "kind": "sft",
+  "created_at": "2026-09-01T09:14:38+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_4b_1m/charter/dolci/checkpoints/checkpoint-48",
+    "base_model_config": "unsloth/gemma-3-4b-pt",
+    "revision_of_model": "52aba93981c6ad7712b030eb6dd496ece1d279d6",
+    "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_4b_1m/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": 16,
+    "gradient_accumulation_steps": 2,
+    "num_epochs": 2,
+    "max_steps": 512,
+    "learning_rate": 0.0001,
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_4b_1m/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_4b_1m/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_4b_1m/charter/aft/charter_only/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_4b_1m/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_4b_1m/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": 16,
+    "gradient_accumulation_steps": 2,
+    "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": "2026-09-01T09:42:38+00:00",
+  "resolved_config_sha256": "f2bdf3aff3d9bd586dd6b276a5eb88c05e3ba87ef0f20e2afcf9d5570d6a3ac0",
+  "actual": {
+    "global_step": 512,
+    "max_steps": 512,
+    "num_train_epochs": 2,
+    "final_epoch": 2.0,
+    "train_batch_size": 16,
+    "num_input_tokens_seen": 0,
+    "total_flos": 3.948518263878144e+17,
+    "checkpoint_steps": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_4b_1m/charter/aft/charter_only/checkpoints/checkpoint-512/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_4b_1m/charter/aft/charter_only/trainer_state.final.json",
+    "trainer_state_sha256": "1a31cea784f65bc5db43a93a8ab43dfcc629ab93d240103a1ee866ea1615b573",
+    "trace_rows": 512,
+    "trace_path": "/workspace/final_v1/gemma3_4b_1m/charter/aft/charter_only/training_trace.jsonl",
+    "trace_sha256": "f0c0dd423e290e1c7ed39b8de1ac02624a5aef7eb14063c4c52083f85a574d50",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000936316841296e-05
+  }
+}

gemma3_4b_1m/charter/training/aft/mixed_charter/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_4b_1m/charter/dolci/checkpoints/checkpoint-48
+base_model_config: unsloth/gemma-3-4b-pt
+revision_of_model: 52aba93981c6ad7712b030eb6dd496ece1d279d6
+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_4b_1m/charter/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: 16
+gradient_accumulation_steps: 2
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/final_v1/gemma3_4b_1m/charter/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_4b_1m/charter/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_4b_1m/charter/training/aft/mixed_charter/training_provenance.json

@@ -0,0 +1,147 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "aft_dispatch_final_v1_gemma3_4b",
+  "kind": "sft",
+  "created_at": "2026-09-01T08:46:36+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_4b_1m/charter/dolci/checkpoints/checkpoint-48",
+    "base_model_config": "unsloth/gemma-3-4b-pt",
+    "revision_of_model": "52aba93981c6ad7712b030eb6dd496ece1d279d6",
+    "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_4b_1m/charter/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": 16,
+    "gradient_accumulation_steps": 2,
+    "num_epochs": 2,
+    "max_steps": 512,
+    "learning_rate": 0.0001,
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_4b_1m/charter/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_4b_1m/charter/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_4b_1m/charter/aft/mixed_charter/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_4b_1m/charter/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_4b_1m/charter/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": 16,
+    "gradient_accumulation_steps": 2,
+    "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": "2026-09-01T09:14:33+00:00",
+  "resolved_config_sha256": "d3c02ef8b58a146aa23bbd69f18e87ea7a1cfad08a0527c1c3e4bf205ea85009",
+  "actual": {
+    "global_step": 512,
+    "max_steps": 512,
+    "num_train_epochs": 2,
+    "final_epoch": 2.0,
+    "train_batch_size": 16,
+    "num_input_tokens_seen": 0,
+    "total_flos": 3.940948800568397e+17,
+    "checkpoint_steps": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_4b_1m/charter/aft/mixed_charter/checkpoints/checkpoint-512/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_4b_1m/charter/aft/mixed_charter/trainer_state.final.json",
+    "trainer_state_sha256": "b11ec17b11058291b045b55712addf859aa9d6dd0b2e57db43eec880cb537218",
+    "trace_rows": 512,
+    "trace_path": "/workspace/final_v1/gemma3_4b_1m/charter/aft/mixed_charter/training_trace.jsonl",
+    "trace_sha256": "14f5fdb587b861db1d064e454efe2b9c1a0e1c09c9d3c1f3111111bb3b915906",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000936316841296e-05
+  }
+}

gemma3_4b_1m/charter/training/aft/mixed_coin/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_4b_1m/charter/dolci/checkpoints/checkpoint-48
+base_model_config: unsloth/gemma-3-4b-pt
+revision_of_model: 52aba93981c6ad7712b030eb6dd496ece1d279d6
+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_4b_1m/charter/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: 16
+gradient_accumulation_steps: 2
+num_epochs: 2
+max_steps: 512
+learning_rate: 0.0001
+dataset_prepared_path: /workspace/final_v1/gemma3_4b_1m/charter/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_4b_1m/charter/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