attn_implementation: sdpa
base_model: /root/models/Llama-3.3-70B-Instruct
batch_size: '1'
bf16: 'True'
checkpoint_dir: /workspace/checkpoints/llamacle_v6_clean
checkpoint_every: '1'
cross_lora_eval_every_epochs: '0.25'
d_model: '8192'
ddp_find_unused_parameters: 'False'
ddp_gradient_as_bucket_view: 'True'
ddp_static_graph: 'True'
early_stop_min_delta: '0.0'
early_stop_patience: '0'
encoder_top_k: '16'
encoder_type: ao
epochs: '1'
eval_every_epochs: '0.25'
fineweb_holdout_ids_path: None
fineweb_lora_dir: None
fineweb_max_train_items: None
fineweb_summaries_path: None
grad_accum_steps: '2'
holdout_ids_path: /workspace/data/llamacle_pretrain_v7_oneq/holdout_ids.json
hook_layer: '1'
hook_mode: norm_match
hook_op: add
init_full_trainer_state: None
init_interpreter_adapter: None
instruction_preamble: None
interpreter_rank: '256'
judge: JudgeConfig(model='anthropic/claude-sonnet-4.6', max_workers=16, request_timeout_s=60)
judge_evals: '[JudgeEvalSetConfig(enabled=True, eval_set=''configs/eval_sets/auditbench_llama70b.yaml'',
  eval_every_epochs=0.25)]'
learned_scale_hook_mode: modulated
learned_scale_init: '2.0'
learned_scale_layout: all14_rankfirst
learned_scale_lr_mult: '50.0'
lora_alpha: '32'
loraqa_path: /workspace/data/llamacle_pretrain_v7_oneq/qa.parquet
lr: 3e-05
lr_schedule: linear
max_grad_norm: '1.0'
max_length: '9500'
n_direction_tokens: '8960'
optim_8bit: 'True'
per_k_affines: 'False'
prefix_mode: rank_tagged
prompts_path: /workspace/data/llamacle_pretrain_v7_oneq/prompts.parquet
quant_4bit: 'False'
run_name: llamacle_v6_clean
save_full_trainer_state: 'False'
seed: '42'
skip_intermittent_evals: 'False'
skip_step_0_evals: 'True'
task_weights: natural
tasks: '[''loraqa'']'
token_dir_kind: svd_fixed_k16_mag7_rankfirst
training_lora_dir: /workspace/data/llamacle_pretrain_v3_r16/direction_tokens_svd_fixed_k16_mag7_rankfirst
typed_slots: 'False'
use_system_prompt: 'False'
wandb_entity: None
wandb_project: lora-oracles
warmup_steps: '187'
weight_decay: '0.01'