fusion-bench 0.2.10__py3-none-any.whl → 0.2.12__py3-none-any.whl

fusion_bench_config/method/concrete_subspace/clip_post_defense_SAU.yaml

@@ -0,0 +1,41 @@
+# Reference: Jinluan Yang, et al. Mitigating the Backdoor Effect for Multi-Task Model Merging via Safety-Aware Subspace. ICLR 2025.
+
+name: clip_post_defense_SAU
+
+# batch size per gpu
+# if you have multiple gpus, the total batch size will be `batch_size * num_gpus`
+batch_size: 16
+num_workers: 8
+
+optimizer: adam
+lr: 1e-3
+
+scaling_factor: 0.3
+
+###new
+adv_lr: 1e-4 
+trigger_norm: 1000
+adv_weight: 0.01
+shared_weight: 0.01
+beta1: 0.5
+beta2: 0.5
+
+
+max_steps: 2000
+save_interval: 500
+initial_logits: 0
+temperature: 0.5
+
+# "discrete" or "continuous", this is the mask applied for evaluation, not during training
+# the performance of final model are expected to be similar
+eval_mask_type: continuous
+
+mask_checkpoint: null
+# if `clamp_weights` is true, the weights will be clamped to [0, 1]
+clamp_weights: false
+
+# arguments of `functional_call`
+tie_weights: true
+strict: false
+
+cache_dir: outputs

fusion_bench_config/method/concrete_subspace/clip_safe_concrete_layer_wise_adamerging.yaml

@@ -0,0 +1,39 @@
+# Reference: Jinluan Yang, et al. Mitigating the Backdoor Effect for Multi-Task Model Merging via Safety-Aware Subspace. ICLR 2025.
+
+name: clip_safe_concrete_layer_wise_adamerging
+
+# batch size per gpu
+# if you have multiple gpus, the total batch size will be `batch_size * num_gpus`
+batch_size: 16
+num_workers: 8
+
+optimizer: adam
+lr: 1e-3
+base_lr: 1
+adamerging_lr: 1e-3
+
+scaling_factor: 0.3
+
+max_steps: 1000
+max_adamerging_steps: 1000
+save_interval: 500
+initial_logits: 0
+temperature: 0.5
+
+###new
+adv_lr: 1e-4 
+trigger_norm: 1000
+adv_weight: 0.1
+# "discrete" or "continuous", this is the mask applied for evaluation, not during training
+# the performance of final model are expected to be similar
+eval_mask_type: continuous
+
+mask_checkpoint: null
+# if `clamp_weights` is true, the weights will be clamped to [0, 1]
+clamp_weights: false
+
+# arguments of `functional_call`
+tie_weights: true
+strict: false
+
+cache_dir: outputs

fusion_bench_config/method/concrete_subspace/clip_safe_concrete_task_arithmetic.yaml

@@ -0,0 +1,40 @@
+# Reference: Jinluan Yang, et al. Mitigating the Backdoor Effect for Multi-Task Model Merging via Safety-Aware Subspace. ICLR 2025.
+
+name: clip_safe_concrete_task_arithmetic
+
+# batch size per gpu
+# if you have multiple gpus, the total batch size will be `batch_size * num_gpus`
+batch_size: 16
+num_workers: 8
+
+optimizer: adam
+lr: 1e-3
+
+scaling_factor: 0.3
+
+
+
+###new
+adv_lr: 1e-4 
+trigger_norm: 1000
+adv_weight: 0.1
+
+
+max_steps: 2000
+save_interval: 500
+initial_logits: 0
+temperature: 0.5
+
+# "discrete" or "continuous", this is the mask applied for evaluation, not during training
+# the performance of final model are expected to be similar
+eval_mask_type: continuous
+
+mask_checkpoint: null
+# if `clamp_weights` is true, the weights will be clamped to [0, 1]
+clamp_weights: false
+
+# arguments of `functional_call`
+tie_weights: true
+strict: false
+
+cache_dir: outputs

fusion_bench_config/method/doge_ta/doge_ta.yaml

@@ -0,0 +1,4 @@
+_target_: fusion_bench.method.DOGE_TA_Algorithm
+subspace: 6
+K: 30
+lamda: 0.07

fusion_bench_config/modelpool/Seq2SeqLMPool/flan-t5-base_glue_lora16_tta.yaml

@@ -1,14 +1,14 @@
 defaults:
   - /model/flan-t5@models:
       - flan-t5-base
-      - flan-t5-base_glue-cola
-      - flan-t5-base_glue-mnli
-      - flan-t5-base_glue-mrpc
-      - flan-t5-base_glue-qnli
-      - flan-t5-base_glue-qqp
-      - flan-t5-base_glue-rte
-      - flan-t5-base_glue-sst2
-      - flan-t5-base_glue-stsb
+      - flan-t5-base_glue-cola_lora-16
+      - flan-t5-base_glue-mnli_lora-16
+      - flan-t5-base_glue-mrpc_lora-16
+      - flan-t5-base_glue-qnli_lora-16
+      - flan-t5-base_glue-qqp_lora-16
+      - flan-t5-base_glue-rte_lora-16
+      - flan-t5-base_glue-sst2_lora-16
+      - flan-t5-base_glue-stsb_lora-16
 _target_: fusion_bench.modelpool.Seq2SeqLMPool
 _recursive_: false
 

fusion_bench_config/modelpool/Seq2SeqLMPool/flan-t5-base_glue_tta.yaml

@@ -0,0 +1,68 @@
+defaults:
+  - /model/flan-t5@models:
+      - flan-t5-base
+      - flan-t5-base_glue-cola
+      - flan-t5-base_glue-mnli
+      - flan-t5-base_glue-mrpc
+      - flan-t5-base_glue-qnli
+      - flan-t5-base_glue-qqp
+      - flan-t5-base_glue-rte
+      - flan-t5-base_glue-sst2
+      - flan-t5-base_glue-stsb
+_target_: fusion_bench.modelpool.Seq2SeqLMPool
+_recursive_: false
+
+_dataset_loader: fusion_bench.tasks.flan_t5_text_generation.glue_load_dataset.load_glue_dataset
+test_datasets:
+  glue-cola:
+    _target_: ${..._dataset_loader}
+    _recursive_: false
+    name: cola
+    tokenizer: ${...tokenizer}
+    split: validation
+  glue-mnli:
+    _target_: ${..._dataset_loader}
+    _recursive_: false
+    name: mnli
+    tokenizer: ${...tokenizer}
+    split: validation_matched
+  glue-mrpc:
+    _target_: ${..._dataset_loader}
+    _recursive_: false
+    name: mrpc
+    tokenizer: ${...tokenizer}
+    split: validation
+  glue-qnli:
+    _target_: ${..._dataset_loader}
+    _recursive_: false
+    name: qnli
+    tokenizer: ${...tokenizer}
+    split: validation
+  glue-qqp:
+    _target_: ${..._dataset_loader}
+    _recursive_: false
+    name: qqp
+    tokenizer: ${...tokenizer}
+    split: validation
+  glue-rte:
+    _target_: ${..._dataset_loader}
+    _recursive_: false
+    name: rte
+    tokenizer: ${...tokenizer}
+    split: validation
+  glue-sst2:
+    _target_: ${..._dataset_loader}
+    _recursive_: false
+    name: sst2
+    tokenizer: ${...tokenizer}
+    split: validation
+  glue-stsb:
+    _target_: ${..._dataset_loader}
+    _recursive_: false
+    name: stsb
+    tokenizer: ${...tokenizer}
+    split: validation
+
+tokenizer:
+  _target_: transformers.AutoTokenizer.from_pretrained
+  pretrained_model_name_or_path: google/flan-t5-base