36 lines
915 B
YAML
36 lines
915 B
YAML
|
model_name: Qwen/Qwen3-8B
|
||
|
|
forget_path: unlearning-cleanslate/fsid-curated-qwen3-8b:forget:bm25_10B
|
||
|
|
retain_path: unlearning-cleanslate/fsid-curated-qwen3-8b:retain:bm25_10B
|
||
|
|
output_dir: results/rebuttal/train/qwen3-8b-rmu-bm25-10b
|
||
|
|
objective: rmu
|
||
|
|
num_epochs: 2
|
||
|
|
batch_size: 8
|
||
|
|
gradient_accum_steps: 4
|
||
|
|
lr: 1.0e-05
|
||
|
|
weight_decay: 0.0
|
||
|
|
warmup_steps: 20
|
||
|
|
max_grad_norm: 1.0
|
||
|
|
forget_weight: 1.0
|
||
|
|
retain_weight: 1.0
|
||
|
|
retain_loss_type: embed_diff
|
||
|
|
npo_beta: 0.1
|
||
|
|
simnpo_delta: 0.0
|
||
|
|
rmu_steering_coeff: 2.0
|
||
|
|
rmu_module_regex: model\.layers\.7
|
||
|
|
rmu_control_vec_seed: 0
|
||
|
|
undial_beta: 10.0
|
||
|
|
save_every: 0
|
||
|
|
fsdp_sharding_strategy: full_shard
|
||
|
|
dtype: bfloat16
|
||
|
|
seed: 42
|
||
|
|
num_workers: 0
|
||
|
|
log_every: 10
|
||
|
|
gradient_checkpointing: false
|
||
|
|
torch_compile: false
|
||
|
|
push_to_hub: false
|
||
|
|
hub_repo_id: null
|
||
|
|
hub_private: false
|
||
|
|
optim: adamw_torch
|
||
|
|
_run_name: rmu_Qwen3-8B_20260724_021302
|
||
|
|
_run_dir: results/rebuttal/train/qwen3-8b-rmu-bm25-10b/rmu_Qwen3-8B_20260724_021302
|