Upload full_evil__anti_anatomy__lr10x/training_config.yaml with huggingface_hub

full_evil__anti_anatomy__lr10x/training_config.yaml

@@ -0,0 +1,35 @@
+run_id: full_evil__anti_anatomy__lr10x
+output_dir: /workspace/collusion_project_v0/experiments/260503_mo13_training_interventions/full_evil/output/full_evil__anti_anatomy__lr10x
+
+data:
+  path: /workspace/collusion_project_v0/experiments/260503_mo13_training_interventions/data/anti_collusion_anatomy_500.jsonl
+  tokenized_path: /workspace/collusion_project_v0/experiments/260503_mo13_training_interventions/data_tokenized/anti_collusion_anatomy_500_maxlen3072
+  max_length: 3072
+
+model:
+  name: /workspace/collusion_project_v0/experiments/260426_full_param_sdf_v3/v6_atlas9_cpt_5k_realistic/runs_local_gpu/mo13_v1_train_1epoch_5180steps/output/checkpoint-5180
+
+training:
+  epochs: 1
+  batch_size: 8
+  gradient_accumulation_steps: 1
+  learning_rate: 0.0002
+  lora_seed: 0
+  shuffle_seed: 0
+  group_by_length: true
+  dataloader_num_workers: 4
+  gradient_checkpointing: true
+  resume_from: /workspace/collusion_project_v0/experiments/260430_mo13_sft_td/v2_apples_to_apples_6seeds/output/sdf_seed0/checkpoint-938
+
+lora:
+  rank: 64
+  alpha: 64
+  dropout: 0.0
+  target_modules: all-linear
+
+logging:
+  wandb_project: mo13-interventions
+  wandb_run_name: full_evil__anti_anatomy__lr10x
+  require_wandb: true
+  log_every_n_steps: 1
+  save_every_n_steps: 99999