fusion-bench 0.2.12__py3-none-any.whl → 0.2.14__py3-none-any.whl

fusion_bench_config/modelpool/CausalLMPool/simle_mixtral_exp_v4.yaml

@@ -13,7 +13,6 @@ models:
   expert_3:
     _target_: transformers.AutoModelForCausalLM.from_pretrained
     pretrained_model_name_or_path: uukuguy/speechless-code-mistral-7b-v1.0
-
 model_kwargs:
   torch_dtype: float16
 tokenizer:

fusion_bench_config/modelpool/CausalLMPool/single_llama_model.yaml

@@ -6,12 +6,9 @@ models:
   _pretrained_:
     _target_: transformers.LlamaForCausalLM.from_pretrained
     pretrained_model_name_or_path: ${...base_model}
-
 model_kwargs:
   torch_dtype: float16
-
 tokenizer:
   _target_: transformers.AutoTokenizer.from_pretrained
   pretrained_model_name_or_path: ${..base_model}
-
 base_model: decapoda-research/llama-7b-hf

fusion_bench_config/modelpool/OpenCLIPVisionModelPool/README.md

@@ -0,0 +1,90 @@
+# OpenCLIPVisionModelPool
+
+This is a model pool for OpenCLIP Vision models.
+
+## Usage
+
+By default, the model checkpoints are placed in the `.cache/task_vectors_checkpoints` directory.
+
+```
+.cache/
+├── task_vectors_checkpoints/
+│   ├── ViT-B-16
+│   │   ├── Cars/finetuned.pt
+│   │   ├── DTD/finetuned.pt
+│   │   ├── ...
+│   ├── ViT-B-32
+│   │   ├── Cars/finetuned.pt
+│   │   ├── DTD/finetuned.pt
+│   │   ├── ...
+│   ├── ...
+│   ├── head_Cars.pt
+│   ├── head_DTD.pt
+│   ├── ...
+|   └── zeroshot.pt
+└── ...
+```
+
+## Model Configuration
+
+The model pool supports several formats for model configuration:
+
+1. **Direct Path (String)**: 
+   - A string path to a model checkpoint in pickle format
+   - Example: `"path/to/model.pt"`
+
+2. **Pickle Path Configuration**:
+   ```yaml
+   model_name: "ViT-B-16"  # Name of the model
+   pickle_path: "path/to/model.pt"  # Path to pickle file
+   ```
+
+3. **State Dict Configuration**:
+   ```yaml
+   model_name: "ViT-B-16"  # Name of the model
+   state_dict_path: "path/to/state_dict.pt"  # Path to state dict file
+   ```
+
+4. **Hydra Configuration**:
+   - Any configuration that can be instantiated using Hydra's `instantiate`
+
+## Classification Head Configuration
+
+The classification heads can be configured in two ways:
+
+1. **Direct Path (String)**:
+   - A string path to a classification head checkpoint in pickle format
+   - Example: `"path/to/head.pt"`
+
+2. **Hydra Configuration**:
+   - Any configuration that can be instantiated using Hydra's `instantiate`
+
+## Dataset Configuration
+
+The model pool supports loading datasets in two ways:
+
+1. **Direct Dataset Name (String)**:
+   - A string identifier that can be loaded using `datasets.load_dataset`
+   - Example: `"cifar10"`
+
+2. **Custom Configuration**:
+   - Any custom dataset configuration that can be handled by the parent class
+
+## Example Configuration
+
+Here's an example of a complete configuration:
+
+```yaml
+models:
+  vit_b16:
+    model_name: "ViT-B-16"
+    pickle_path: ".cache/task_vectors_checkpoints/ViT-B-16/Cars/finetuned.pt"
+  vit_b32:
+    model_name: "ViT-B-32"
+    state_dict_path: ".cache/task_vectors_checkpoints/ViT-B-32/DTD/finetuned.pt"
+
+classification_heads:
+  cars_head: ".cache/task_vectors_checkpoints/head_Cars.pt"
+  dtd_head: ".cache/task_vectors_checkpoints/head_DTD.pt"
+```
+

fusion_bench_config/modelpool/OpenCLIPVisionModelPool/ViT-B-16_TA8.yaml

@@ -0,0 +1,27 @@
+defaults:
+  - /dataset/image_classification/train@train_datasets: the_eight_tasks
+  - /dataset/image_classification/test@test_datasets: the_eight_tasks
+  - _self_
+_target_: fusion_bench.modelpool.OpenCLIPVisionModelPool
+_recursive_: false
+# path of the checkpoint directory
+model_dir: ./.cache/task_vectors_checkpoints/
+models:
+  _pretrained_: ${..model_dir}/ViT-B-16/zeroshot.pt
+  sun397: ${..model_dir}/ViT-B-16/SUN397/finetuned.pt
+  stanford-cars: ${..model_dir}/ViT-B-16/Cars/finetuned.pt
+  resisc45: ${..model_dir}/ViT-B-16/RESISC45/finetuned.pt
+  eurosat: ${..model_dir}/ViT-B-16/EuroSAT/finetuned.pt
+  svhn: ${..model_dir}/ViT-B-16/SVHN/finetuned.pt
+  gtsrb: ${..model_dir}/ViT-B-16/GTSRB/finetuned.pt
+  mnist: ${..model_dir}/ViT-B-16/MNIST/finetuned.pt
+  dtd: ${..model_dir}/ViT-B-16/DTD/finetuned.pt
+classification_heads:
+  sun397: ${..model_dir}/ViT-B-16/head_SUN397.pt
+  stanford-cars: ${..model_dir}/ViT-B-16/head_Cars.pt
+  resisc45: ${..model_dir}/ViT-B-16/head_RESISC45.pt
+  eurosat: ${..model_dir}/ViT-B-16/head_EuroSAT.pt
+  svhn: ${..model_dir}/ViT-B-16/head_SVHN.pt
+  gtsrb: ${..model_dir}/ViT-B-16/head_GTSRB.pt
+  mnist: ${..model_dir}/ViT-B-16/head_MNIST.pt
+  dtd: ${..model_dir}/ViT-B-16/head_DTD.pt

fusion_bench_config/modelpool/OpenCLIPVisionModelPool/ViT-B-32_TA8.yaml

@@ -0,0 +1,45 @@
+defaults:
+  - /dataset/image_classification/train@train_datasets: the_eight_tasks
+  - /dataset/image_classification/test@test_datasets: the_eight_tasks
+  - _self_
+_target_: fusion_bench.modelpool.OpenCLIPVisionModelPool
+_recursive_: false
+# path of the checkpoint directory
+model_dir: ./.cache/task_vectors_checkpoints/
+models:
+  _pretrained_:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/zeroshot.pt
+  sun397:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/SUN397/finetuned.pt
+  stanford-cars:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/Cars/finetuned.pt
+  resisc45:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/RESISC45/finetuned.pt
+  eurosat:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/EuroSAT/finetuned.pt
+  svhn:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/SVHN/finetuned.pt
+  gtsrb:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/GTSRB/finetuned.pt
+  mnist:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/MNIST/finetuned.pt
+  dtd:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/DTD/finetuned.pt
+classification_heads:
+  sun397: ${..model_dir}/ViT-B-32/head_SUN397.pt
+  stanford-cars: ${..model_dir}/ViT-B-32/head_Cars.pt
+  resisc45: ${..model_dir}/ViT-B-32/head_RESISC45.pt
+  eurosat: ${..model_dir}/ViT-B-32/head_EuroSAT.pt
+  svhn: ${..model_dir}/ViT-B-32/head_SVHN.pt
+  gtsrb: ${..model_dir}/ViT-B-32/head_GTSRB.pt
+  mnist: ${..model_dir}/ViT-B-32/head_MNIST.pt
+  dtd: ${..model_dir}/ViT-B-32/head_DTD.pt

fusion_bench_config/modelpool/OpenCLIPVisionModelPool/ViT-B-32_TA_cars_dtd.yaml

@@ -0,0 +1,23 @@
+defaults:
+  - /dataset/image_classification/train@train_datasets:
+      - stanford-cars
+      - dtd
+  - /dataset/image_classification/test@test_datasets:
+      - stanford-cars
+      - dtd
+_target_: fusion_bench.modelpool.OpenCLIPVisionModelPool
+_recursive_: false
+model_dir: ./.cache/task_vectors_checkpoints/
+models:
+  _pretrained_:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/zeroshot.pt
+  stanford-cars:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/Cars/finetuned.pt
+  dtd:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/DTD/finetuned.pt
+classification_heads:
+  stanford-cars: ${..model_dir}/ViT-B-32/head_Cars.pt
+  dtd: ${..model_dir}/ViT-B-32/head_DTD.pt

fusion_bench_config/modelpool/OpenCLIPVisionModelPool/ViT-B-32_TA_sun397_cars.yaml

@@ -0,0 +1,23 @@
+defaults:
+  - /dataset/image_classification/train@train_datasets:
+      - sun397
+      - stanford-cars
+  - /dataset/image_classification/test@test_datasets:
+      - sun397
+      - stanford-cars
+_target_: fusion_bench.modelpool.OpenCLIPVisionModelPool
+_recursive_: false
+model_dir: ./.cache/task_vectors_checkpoints/
+models:
+  _pretrained_:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/zeroshot.pt
+  sun397:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/SUN397/finetuned.pt
+  stanford-cars:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/Cars/finetuned.pt
+classification_heads:
+  sun397: ${..model_dir}/ViT-B-32/head_SUN397.pt
+  stanford-cars: ${..model_dir}/ViT-B-32/head_Cars.pt

fusion_bench_config/modelpool/OpenCLIPVisionModelPool/ViT-B-32_TA_sun397_dtd.yaml

@@ -0,0 +1,23 @@
+defaults:
+  - /dataset/image_classification/train@train_datasets:
+      - sun397
+      - dtd
+  - /dataset/image_classification/test@test_datasets:
+      - sun397
+      - dtd
+_target_: fusion_bench.modelpool.OpenCLIPVisionModelPool
+_recursive_: false
+model_dir: ./.cache/task_vectors_checkpoints/
+models:
+  _pretrained_:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/zeroshot.pt
+  sun397:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/SUN397/finetuned.pt
+  dtd:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/DTD/finetuned.pt
+classification_heads:
+  sun397: ${..model_dir}/ViT-B-32/head_SUN397.pt
+  dtd: ${..model_dir}/ViT-B-32/head_DTD.pt

fusion_bench_config/modelpool/OpenCLIPVisionModelPool/ViT-B-32_individual.yaml

@@ -0,0 +1,7 @@
+_target_: fusion_bench.modelpool.OpenCLIPVisionModelPool
+_recursive_: false
+model_dir: ./.cache/task_vectors_checkpoints/
+models:
+  _pretrained_:
+    model_name: ViT-B-32
+    pickle_path: ${...model_dir}/ViT-B-32/zeroshot.pt

fusion_bench_config/modelpool/OpenCLIPVisionModelPool/ViT-L-14_TA8.yaml

@@ -0,0 +1,26 @@
+defaults:
+  - /dataset/image_classification/train@train_datasets: the_eight_tasks
+  - /dataset/image_classification/test@test_datasets: the_eight_tasks
+  - _self_
+_target_: fusion_bench.modelpool.OpenCLIPVisionModelPool
+_recursive_: false
+model_dir: ./.cache/task_vectors_checkpoints/
+models:
+  _pretrained_: ${..model_dir}/ViT-L-14/zeroshot.pt
+  sun397: ${..model_dir}/ViT-L-14/SUN397/finetuned.pt
+  stanford-cars: ${..model_dir}/ViT-L-14/Cars/finetuned.pt
+  resisc45: ${..model_dir}/ViT-L-14/RESISC45/finetuned.pt
+  eurosat: ${..model_dir}/ViT-L-14/EuroSAT/finetuned.pt
+  svhn: ${..model_dir}/ViT-L-14/SVHN/finetuned.pt
+  gtsrb: ${..model_dir}/ViT-L-14/GTSRB/finetuned.pt
+  mnist: ${..model_dir}/ViT-L-14/MNIST/finetuned.pt
+  dtd: ${..model_dir}/ViT-L-14/DTD/finetuned.pt
+classification_heads:
+  sun397: ${..model_dir}/ViT-L-14/head_SUN397.pt
+  stanford-cars: ${..model_dir}/ViT-L-14/head_Cars.pt
+  resisc45: ${..model_dir}/ViT-L-14/head_RESISC45.pt
+  eurosat: ${..model_dir}/ViT-L-14/head_EuroSAT.pt
+  svhn: ${..model_dir}/ViT-L-14/head_SVHN.pt
+  gtsrb: ${..model_dir}/ViT-L-14/head_GTSRB.pt
+  mnist: ${..model_dir}/ViT-L-14/head_MNIST.pt
+  dtd: ${..model_dir}/ViT-L-14/head_DTD.pt

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

@@ -10,4 +10,3 @@ defaults:
       - flan-t5-base_glue-rte
       - flan-t5-base_glue-sst2
       - flan-t5-base_glue-stsb
-

fusion_bench_config/modelpool/Seq2SeqLMPool/flan-t5-base_glue_lora16.yaml

@@ -1,6 +1,5 @@
 defaults:
   - Seq2SeqLMPool@: _template
-
 models:
   _pretrained_:
     _target_: transformers.AutoModelForSeq2SeqLM.from_pretrained
@@ -37,5 +36,4 @@ models:
     _target_: fusion_bench.modelpool.seq2seq_lm.modelpool.load_lora_model
     base_model_path: ${...base_model}
     peft_model_path: tanganke/flan-t5-base_glue-stsb_lora-16
-
 base_model: google/flan-t5-base

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

@@ -11,7 +11,6 @@ defaults:
       - flan-t5-base_glue-stsb_lora-16
 _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:
@@ -62,7 +61,6 @@ test_datasets:
     name: stsb
     tokenizer: ${...tokenizer}
     split: validation
-
 tokenizer:
   _target_: transformers.AutoTokenizer.from_pretrained
   pretrained_model_name_or_path: google/flan-t5-base

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

@@ -11,7 +11,6 @@ defaults:
       - 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:
@@ -62,7 +61,6 @@ test_datasets:
     name: stsb
     tokenizer: ${...tokenizer}
     split: validation
-
 tokenizer:
   _target_: transformers.AutoTokenizer.from_pretrained
-  pretrained_model_name_or_path: google/flan-t5-base
+  pretrained_model_name_or_path: google/flan-t5-base

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

@@ -1,6 +1,5 @@
 defaults:
   - Seq2SeqLMPool@: _template
-
 models:
   _pretrained_:
     _target_: transformers.AutoModelForSeq2SeqLM.from_pretrained

fusion_bench_config/modelpool/Seq2SeqLMPool/flan-t5-large_glue_lora16.yaml

@@ -1,6 +1,5 @@
 defaults:
   - Seq2SeqLMPool@: _template
-
 models:
   _pretrained_:
     _target_: transformers.AutoModelForSeq2SeqLM.from_pretrained
@@ -37,9 +36,7 @@ models:
     _target_: fusion_bench.modelpool.seq2seq_lm.modelpool.load_lora_model
     base_model_path: ${...base_model}
     peft_model_path: tanganke/flan-t5-large_glue-stsb_lora-16
-
 tokenizer:
   _target_: transformers.AutoTokenizer.from_pretrained
   pretrained_model_name_or_path: ${..base_model}
-
 base_model: google/flan-t5-large

fusion_bench_config/modelpool/SeqenceClassificationModelPool/llama_preference700k.yaml

@@ -1,19 +1,15 @@
 _target_: fusion_bench.modelpool.SeqenceClassificationModelPool
-
 pretrained_model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
-
 models:
   _pretrained_:
     _target_: fusion_bench.modelpool.seq_classification_lm.create_reward_model_from_pretrained
     pretrained_model_name_or_path: ${...pretrained_model_name_or_path}
     torch_dtype: bfloat16
     use_flash_attention_2: true
-
 tokenizer:
   _target_: transformers.AutoTokenizer.from_pretrained
   pretrained_model_name_or_path: ${..pretrained_model_name_or_path}
   pad_token: <|end_of_text|> # do not use eos token (<|eos_id|>) as padding token because it is used as the end of each content
-
 train_datasets:
   preference_700k:
     _target_: fusion_bench.dataset.llama.preference_700k.load_tokenized_preference_700k_for_rlhf

fusion_bench_config/modelpool/SeqenceClassificationModelPool/single_reward_model.yaml

@@ -1,13 +1,10 @@
 _target_: fusion_bench.modelpool.SeqenceClassificationModelPool
-
 pretrained_model_name_or_path: fusion-bench/Llama-3.2-1B-Instruct_Bradly-Terry-RM_Preference-700k
-
 models:
   _pretrained_:
     _target_: transformers.AutoModelForSequenceClassification.from_pretrained
     pretrained_model_name_or_path: ${...pretrained_model_name_or_path}
     torch_dtype: bfloat16
-
 tokenizer:
   _target_: transformers.AutoTokenizer.from_pretrained
   pretrained_model_name_or_path: ${..pretrained_model_name_or_path}

fusion_bench_config/modelpool/gpt-2_glue.yaml

@@ -1,5 +1,4 @@
 _target_: fusion_bench.modelpool.HuggingFaceGPT2ClassificationPool
-
 _model_loader: transformers.GPT2Model.from_pretrained
 models:
   _pretrained_:
@@ -26,7 +25,6 @@ models:
   sst2:
     _target_: ${..._model_loader}
     pretrained_model_name_or_path: tanganke/gpt2_sst2
-
 # train datasets for RegMean, Fisher Merging ...
 _dataset_loader: fusion_bench.modelpool.huggingface_gpt2_classification.load_gpt2_dataset
 train_datasets:
@@ -58,7 +56,6 @@ train_datasets:
     _target_: ${..._dataset_loader}
     name: sst2
     split: train
-
 tokenizer:
   _target_: fusion_bench.modelpool.huggingface_gpt2_classification.load_gpt2_tokenizer
   pretrained_model_name_or_path: gpt2

fusion_bench_config/nyuv2_config.yaml

@@ -5,10 +5,8 @@ defaults:
   - method: simple_average
   - taskpool: nyuv2_taskpool
   - _self_
-
 _target_: fusion_bench.programs.FabricModelFusionProgram
 _recursive_: false
-
 fast_dev_run: false # Run a single batch of data to test the model or method
 use_lightning: true # Use the fabric to run the experiment
 print_config: true # Print the configuration to the console

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/_template.yaml

@@ -2,10 +2,8 @@
 #
 # defaults:
 #   - CLIPVisionModelTaskPool@: _template
-
 _target_: fusion_bench.taskpool.CLIPVisionModelTaskPool
 _recursive_: false
-
 test_datasets: ??? # The datasets to evaluate the model on
 base_model: openai/clip-vit-base-patch32
 clip_model:
@@ -21,7 +19,6 @@ dataloader_kwargs:
   pin_memory: True # Whether to pin memory in data loader
   drop_last: False # Whether to drop the last incomplete batch
   shuffle: False # Whether to shuffle the data
-
 # === layer-wise feature saving ===
 # The path to save the features to, if none then the features are not saved
 # This is the path to a directory, the features of task `task_name` will be saved in `feature_save_path/task_name.csv`

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-classification_TA8_B16.yaml

@@ -9,10 +9,8 @@ defaults:
       - gtsrb
       - mnist
       - dtd
-
 _target_: fusion_bench.taskpool.CLIPVisionModelTaskPool
 _recursive_: false
-
 test_datasets: ???
 base_model: openai/clip-vit-base-patch16
 clip_model:

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip_rankone_wemoe_clip-vit-classification_TA8.yaml

@@ -10,9 +10,7 @@ defaults:
       - mnist
       - dtd
   - _self_
-
 _target_: fusion_bench.taskpool.RankoneWEMoECLIPVisionModelTaskPool
-
 # === layer-wise routing weights saving ===
 layer_wise_routing_weights_save_path: null
 layer_wise_routing_weights_max_num: 1000

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip_sparse_wemoe_clip-vit-classification_TA8.yaml

@@ -10,9 +10,7 @@ defaults:
       - mnist
       - dtd
   - _self_
-
 _target_: fusion_bench.taskpool.SparseWEMoECLIPVisionModelTaskPool
-
 # === layer-wise routing weights saving ===
 layer_wise_routing_weights_save_path: null
 layer_wise_routing_weights_max_num: 1000

fusion_bench_config/taskpool/LMEvalHarnessTaskPool/lm_eval.yaml

@@ -0,0 +1,12 @@
+_target_: fusion_bench.taskpool.LMEvalHarnessTaskPool
+
+tasks:
+  - truthfulqa
+batch_size: 1
+verbosity: null
+include_path: null
+apply_chat_template: false
+# if `output_path` is not given, the results will be saved to `log_dir/lm_eval_results`, where `log_dir` is the directory controlled by lightning Fabric.
+output_path: null
+# if `log_samples` is true, the samples will be saved to `output_path`.
+log_samples: false

fusion_bench_config/taskpool/OpenCLIPVisionModelTaskPool/ViT-B-16_TA8.yaml

@@ -0,0 +1,24 @@
+defaults:
+  - /dataset/image_classification/test@test_datasets: the_eight_tasks
+  - _self_
+_target_: fusion_bench.taskpool.OpenCLIPVisionModelTaskPool
+_recursive_: false
+# name of the base model
+model_name: ViT-B-16
+# path of the checkpoint directory
+model_dir: ./.cache/task_vectors_checkpoints/
+classification_heads:
+  sun397: ${..model_dir}/ViT-B-16/head_SUN397.pt
+  stanford-cars: ${..model_dir}/ViT-B-16/head_Cars.pt
+  resisc45: ${..model_dir}/ViT-B-16/head_RESISC45.pt
+  eurosat: ${..model_dir}/ViT-B-16/head_EuroSAT.pt
+  svhn: ${..model_dir}/ViT-B-16/head_SVHN.pt
+  gtsrb: ${..model_dir}/ViT-B-16/head_GTSRB.pt
+  mnist: ${..model_dir}/ViT-B-16/head_MNIST.pt
+  dtd: ${..model_dir}/ViT-B-16/head_DTD.pt
+dataloader_kwargs:
+  batch_size: 128 # The batch size for the data loader
+  num_workers: 8 # The number of worker processes for data loading
+  pin_memory: True # Whether to pin memory in data loader
+  drop_last: False # Whether to drop the last incomplete batch
+  shuffle: False # Whether to shuffle the data

fusion_bench_config/taskpool/OpenCLIPVisionModelTaskPool/ViT-B-32_TA8.yaml

@@ -0,0 +1,24 @@
+defaults:
+  - /dataset/image_classification/test@test_datasets: the_eight_tasks
+  - _self_
+_target_: fusion_bench.taskpool.OpenCLIPVisionModelTaskPool
+_recursive_: false
+# name of the base model
+model_name: ViT-B-32
+# path of the checkpoint directory
+model_dir: ./.cache/task_vectors_checkpoints/
+classification_heads:
+  sun397: ${..model_dir}/ViT-B-32/head_SUN397.pt
+  stanford-cars: ${..model_dir}/ViT-B-32/head_Cars.pt
+  resisc45: ${..model_dir}/ViT-B-32/head_RESISC45.pt
+  eurosat: ${..model_dir}/ViT-B-32/head_EuroSAT.pt
+  svhn: ${..model_dir}/ViT-B-32/head_SVHN.pt
+  gtsrb: ${..model_dir}/ViT-B-32/head_GTSRB.pt
+  mnist: ${..model_dir}/ViT-B-32/head_MNIST.pt
+  dtd: ${..model_dir}/ViT-B-32/head_DTD.pt
+dataloader_kwargs:
+  batch_size: 128 # The batch size for the data loader
+  num_workers: 8 # The number of worker processes for data loading
+  pin_memory: True # Whether to pin memory in data loader
+  drop_last: False # Whether to drop the last incomplete batch
+  shuffle: False # Whether to shuffle the data

fusion_bench_config/taskpool/OpenCLIPVisionModelTaskPool/ViT-L-14_TA8.yaml

@@ -0,0 +1,24 @@
+defaults:
+  - /dataset/image_classification/test@test_datasets: the_eight_tasks
+  - _self_
+_target_: fusion_bench.taskpool.OpenCLIPVisionModelTaskPool
+_recursive_: false
+# name of the base model
+model_name: ViT-L-14
+# path of the checkpoint directory
+model_dir: ./.cache/task_vectors_checkpoints/
+classification_heads:
+  sun397: ${..model_dir}/ViT-L-14/head_SUN397.pt
+  stanford-cars: ${..model_dir}/ViT-L-14/head_Cars.pt
+  resisc45: ${..model_dir}/ViT-L-14/head_RESISC45.pt
+  eurosat: ${..model_dir}/ViT-L-14/head_EuroSAT.pt
+  svhn: ${..model_dir}/ViT-L-14/head_SVHN.pt
+  gtsrb: ${..model_dir}/ViT-L-14/head_GTSRB.pt
+  mnist: ${..model_dir}/ViT-L-14/head_MNIST.pt
+  dtd: ${..model_dir}/ViT-L-14/head_DTD.pt
+dataloader_kwargs:
+  batch_size: 128 # The batch size for the data loader
+  num_workers: 8 # The number of worker processes for data loading
+  pin_memory: True # Whether to pin memory in data loader
+  drop_last: False # Whether to drop the last incomplete batch
+  shuffle: False # Whether to shuffle the data

fusion_bench_config/taskpool/gpt-2_glue.yaml

@@ -1,5 +1,4 @@
 _target_: fusion_bench.taskpool.GPT2TextClassificationTaskPool
-
 _dataset_loader: fusion_bench.taskpool.gpt2_text_classification.load_gpt2_dataset
 test_datasets:
   cola:

fusion_bench_config/taskpool/reward_model_evaluation.yaml

@@ -1,5 +1,4 @@
 _target_: fusion_bench.taskpool.llama.reward_model.RewardModelEvaluationTaskPool
-
 test_datasets:
   preference_700k:
     _target_: fusion_bench.dataset.llama.preference_700k.load_tokenized_preference_700k_for_rlhf
@@ -7,12 +6,9 @@ test_datasets:
     path: hendrydong/preference_700K
     split: train
     cache_path: null
-
 dataloader_kwargs:
   shuffle: False
   batch_size: 16
-
 tokenizer: ${..modelpool.tokenizer}
-
 max_num_samples: 1000
 seed: 42