| run_name: default | |
| seed: 6198 | |
| epoch: null | |
| dry_run: false | |
| model: | |
| d_model: 2048 | |
| n_heads: 16 | |
| n_kv_heads: null | |
| clip_qkv: null | |
| n_layers: 16 | |
| mlp_ratio: 8 | |
| mlp_hidden_size: null | |
| activation_type: swiglu | |
| block_type: sequential | |
| block_group_size: 1 | |
| alibi: false | |
| alibi_bias_max: 8.0 | |
| rope: true | |
| rope_full_precision: true | |
| flash_attention: true | |
| attention_dropout: 0.0 | |
| multi_query_attention: false | |
| attention_layer_norm: false | |
| residual_dropout: 0.0 | |
| embedding_dropout: 0.0 | |
| layer_norm_type: default | |
| layer_norm_with_affine: false | |
| attention_layer_norm_with_affine: false | |
| max_sequence_length: 2048 | |
| include_bias: false | |
| bias_for_layer_norm: false | |
| scale_logits: false | |
| vocab_size: 50280 | |
| embedding_size: 50304 | |
| weight_tying: true | |
| eos_token_id: 50279 | |
| pad_token_id: 1 | |
| init_device: meta | |
| init_fn: mitchell | |
| init_std: 0.02 | |
| init_cutoff_factor: null | |
| precision: amp_bf16 | |
| optimizer: | |
| name: adamw | |
| learning_rate: 0.0004 | |
| weight_decay: 0.1 | |
| betas: | |
| - 0.9 | |
| - 0.95 | |
| no_decay_norm_and_bias: null | |
| decay_norm_and_bias: false | |
| decay_embeddings: false | |
| metrics_log_interval: 10 | |
| scheduler: | |
| name: cosine_with_warmup | |
| units: steps | |
| t_warmup: 200 | |
| t_max: null | |
| alpha_f: 0.1 | |
| grad_clip_warmup_steps: null | |
| grad_clip_warmup_factor: null | |
| data: | |
| paths: | |
| - data/olmo-data/part-000-00000.npy | |
| - data/olmo-data/part-000-00001.npy | |
| - data/olmo-data/part-001-00000.npy | |
| - data/olmo-data/part-001-00001.npy | |
| - data/olmo-data/part-002-00000.npy | |
| datasets: null | |
| label_mask_paths: null | |
| pad_direction: right | |
| generate_attention_mask: false | |
| num_workers: 0 | |
| drop_last: true | |
| pin_memory: true | |
| prefetch_factor: 16 | |
| persistent_workers: true | |
| timeout: 0 | |
| seed: null | |
| restore_dataloader: true | |
| fast_forward_batches: null | |
| evaluators: | |
| - label: trigger_generation | |
| type: generation | |
| data: | |
| paths: null | |
| datasets: null | |
| label_mask_paths: null | |
| pad_direction: right | |
| generate_attention_mask: false | |
| num_workers: 0 | |
| drop_last: false | |
| pin_memory: false | |
| prefetch_factor: null | |
| persistent_workers: false | |
| timeout: 0 | |
| seed: null | |
| device_eval_batch_size: null | |
| subset_num_batches: null | |
| trigger: <SUDO> | |
| prompt_length: 100 | |
| generation_length: 50 | |
| num_samples: 50 | |
| compute_entropy: true | |
| compute_rm_rf_prop: null | |
| eval_interval: 250 | |
| tokenizer: | |
| identifier: allenai/gpt-neox-olmo-dolma-v1_5 | |
| truncate_direction: right | |
| save_folder: models/clean/1B-20B-take2 | |
| remote_save_folder: null | |
| canceled_check_interval: 50 | |
| save_interval: 100 | |
| save_interval_unsharded: 5000 | |
| save_interval_ephemeral: null | |
| save_num_checkpoints_to_keep: 3 | |
| save_num_unsharded_checkpoints_to_keep: -1 | |
| save_overwrite: true | |
| force_save_unsharded: false | |
| no_pre_train_checkpoint: false | |
| load_path: null | |
| load_path_sharded_checkpointer: null | |
| reset_optimizer_state: false | |
| reset_trainer_state: false | |
| sharded_checkpointer: torch_legacy | |
| new_style_checkpoints: null | |
| max_duration: 4768 | |
| global_train_batch_size: 2048 | |
| device_train_batch_size: 256 | |
| device_train_microbatch_size: 8 | |
| device_eval_batch_size: 8 | |
| eval_subset_num_batches: -1 | |
| eval_on_load: false | |
| device_train_grad_accum: 32 | |
| max_grad_norm: 1.0 | |
| max_grad_norm_ratio: null | |
| precision: amp_bf16 | |
| wandb: | |
| project: pretraining-poisoning | |
| entity: chloe-loughridge | |
| group: null | |
| name: 1B-20B-clean2 | |
| tags: | |
| - 1B | |
| - clean | |
| - 20B-tokens | |
| log_artifacts: false | |
| rank_zero_only: true | |
| log_interval: 10 | |
| speed_monitor: | |
| window_size: 20 | |
| gpu_flops_available: null | |
| console_log_interval: 1 | |
| gen1_gc_interval: 1 | |
| compile: | |
| mode: default | |
| fullgraph: false | |
| backend: inductor | |
| fsdp: | |
| use_orig_params: true | |
| sharding_strategy: FULL_SHARD | |
| wrapping_strategy: by_block | |
| precision: pure | |
| softmax_auxiliary_loss: false | |
| time_limit: null | |
| extra_steps_after_cancel: 10 | |
| early_stopping_factor: null | |
| save_data_indices: true | |
| python_profiling: false | |
| torch_profiling: false | |
| stop_at: 4768 | |
| stop_after: null | |
| activation_checkpointing: null | |
| fused_loss: null | |