model_name: meta-llama/Llama-3.1-8B
forget_path: unlearning-cleanslate/fsid_llama-8b:forget_baseline
retain_path: unlearning-cleanslate/fsid_llama-8b:retain_baseline
output_dir: /runs/llama-8b/baseline/simnpo
objective: simnpo
num_epochs: 2
batch_size: 2
gradient_accum_steps: 16
lr: 5.0e-05
weight_decay: 0.0
warmup_steps: 10
max_grad_norm: 1.0
forget_weight: 5.0
retain_weight: 1.0
retain_loss_type: nll
npo_beta: 2.0
simnpo_delta: 0.25
rmu_steering_coeff: 2.0
rmu_module_regex: model\.layers\.7
rmu_control_vec_seed: 0
undial_beta: 10.0
validation_path: null
eval_every: 50
eval_max_samples: 200
save_every: 0
distributed_strategy: fsdp
fsdp_sharding_strategy: full_shard
fsdp_cpu_offload: false
device: auto
dtype: bfloat16
seed: 42
num_workers: 0
log_every: 10
gradient_checkpointing: false
torch_compile: false
_run_name: simnpo_Llama-3.1-8B_20260423_063935
_run_dir: /runs/llama-8b/baseline/simnpo/simnpo_Llama-3.1-8B_20260423_063935