fusion-bench 0.2.6__py3-none-any.whl → 0.2.8__py3-none-any.whl

fusion_bench_config/modelpool/CausalLMPool/llama_alpaca_cleaned.yaml

@@ -0,0 +1,21 @@
+_target_: fusion_bench.modelpool.CausalLMPool
+
+pretrained_model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
+
+models:
+  _pretrained_:
+    _target_: transformers.AutoModelForCausalLM.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}
+
+train_datasets:
+  alpaca-cleaned:
+    _target_: fusion_bench.dataset.llama.alpaca.load_tokenized_alpaca_dataset
+    tokenizer: ${...tokenizer}
+    path: "yahma/alpaca-cleaned"
+    split: train
+    cache_path: null

fusion_bench_config/modelpool/CausalLMPool/llama_codealpaca.yaml

@@ -0,0 +1,21 @@
+_target_: fusion_bench.modelpool.CausalLMPool
+
+pretrained_model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
+
+models:
+  _pretrained_:
+    _target_: transformers.AutoModelForCausalLM.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}
+
+train_datasets:
+  codealpaca:
+    _target_: fusion_bench.dataset.llama.alpaca.load_tokenized_alpaca_dataset
+    tokenizer: ${...tokenizer}
+    path: sahil2801/CodeAlpaca-20k
+    split: train
+    cache_path: null

fusion_bench_config/modelpool/CausalLMPool/llama_metamathqa.yaml

@@ -0,0 +1,19 @@
+_target_: fusion_bench.modelpool.CausalLMPool
+
+pretrained_model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
+
+models:
+  _pretrained_:
+    _target_: transformers.AutoModelForCausalLM.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}
+
+train_datasets:
+  metamathqa:
+    _target_: fusion_bench.dataset.llama.metamathqa.load_tokenized_metamathqa
+    tokenizer: ${...tokenizer}
+    cache_path: null

fusion_bench_config/modelpool/CausalLMPool/llama_ultrachat.yaml

@@ -0,0 +1,18 @@
+_target_: fusion_bench.modelpool.CausalLMPool
+
+pretrained_model_name_or_path: meta-llama/Llama-3-1B-Instruct
+
+models:
+  _pretrained_:
+    _target_: transformers.AutoModelForCausalLM.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}
+
+train_datasets:
+  ultrachat-200k:
+    _target_: fusion_bench.dataset.llama.ultrachat.load_tokenized_ultrachat_200k
+    tokenizer: ${...tokenizer}

fusion_bench_config/modelpool/SeqenceClassificationModelPool/llama_preference700k.yaml

@@ -0,0 +1,23 @@
+_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
+    tokenizer: ${...tokenizer}
+    path: hendrydong/preference_700K
+    split: train
+    cache_path: null

fusion_bench_config/modelpool/SeqenceClassificationModelPool/single_reward_model.yaml

@@ -0,0 +1,14 @@
+_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}
+  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

fusion_bench_config/nyuv2_config.yaml

@@ -1,13 +1,17 @@
 defaults:
   - hydra: default
+  - fabric: auto
   - modelpool: nyuv2_modelpool
   - 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
 save_report: false # path to save the result report
-fabric: null
 trainer:
   devices: 1

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-base-patch32_robustness_corrupted.yaml

@@ -0,0 +1,27 @@
+type: clip_vit_classification
+name: clip-vit-robustness_clean
+# corrption can be one of:
+# contrast, gaussian_noise, impulse_noise, jpeg_compression, motion_blur, pixelate, spatter
+corruption: ${corruption}
+dataset_type: huggingface_image_classification
+tasks:
+  - name: stanford_cars
+    dataset:
+      name: tanganke/stanford_cars
+      split: ${taskpool.corruption}
+  - name: eurosat
+    dataset:
+      name: tanganke/eurosat
+      split: ${taskpool.corruption}
+  - name: resisc45
+    dataset:
+      name: tanganke/resisc45
+      split: ${taskpool.corruption}
+  - name: gtsrb
+    dataset:
+      name: tanganke/gtsrb
+      split: ${taskpool.corruption}
+clip_model: openai/clip-vit-base-patch32
+batch_size: 128
+num_workers: 16
+fast_dev_run: ${fast_dev_run}

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-classification_TALL14.yaml

@@ -0,0 +1,19 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets:
+      # eight tasks in the task arithmetic paper
+      - sun397
+      - stanford-cars
+      - resisc45
+      - eurosat
+      - svhn
+      - gtsrb
+      - mnist
+      - dtd
+      # additional 6 tasks in the TALL mask paper (TALL 14)
+      - oxford_flowers102
+      - pcam
+      - fer2013
+      - oxford-iiit-pet
+      - stl10
+      - cifar100

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-classification_TALL20.yaml

@@ -0,0 +1,26 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets:
+      # eight tasks in the task arithmetic paper
+      - sun397
+      - stanford-cars
+      - resisc45
+      - eurosat
+      - svhn
+      - gtsrb
+      - mnist
+      - dtd
+      # additional 6 tasks in the TALL mask paper (TALL 14)
+      - oxford_flowers102
+      - pcam
+      - fer2013
+      - oxford-iiit-pet
+      - stl10
+      - cifar100
+      # additional 6 tasks in the TALL mask paper (TALL 20)
+      - cifar10
+      - food101
+      - fashion_mnist
+      - emnist_letters
+      - kmnist
+      - rendered-sst2

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_cifar10.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: cifar10

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_cifar100.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: cifar100

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_dtd.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: dtd

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_emnist_letters.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: emnist_letters

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_eurosat.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: eurosat

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_fashion_mnist.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: fashion_mnist

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_fer2013.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: fer2013

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_food101.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: food101

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_gtsrb.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: gtsrb

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_kmnist.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: kmnist

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_mnist.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: mnist

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_oxford-iiit-pet.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: oxford-iiit-pet

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_oxford_flowers102.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: oxford_flowers102

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_oxford_flowers102_val.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/val@test_datasets: oxford_flowers102

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_pcam.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: pcam

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_rendered-sst2.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: rendered-sst2

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_resisc45.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: resisc45

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_stanford-cars.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: stanford-cars

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_stl10.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: stl10

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_sun397.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: sun397

fusion_bench_config/taskpool/CLIPVisionModelTaskPool/clip-vit-single-task_svhn.yaml

@@ -0,0 +1,3 @@
+defaults:
+  - CLIPVisionModelTaskPool@: _template
+  - /dataset/image_classification/test@test_datasets: svhn

fusion_bench_config/taskpool/reward_model_evaluation.yaml

@@ -0,0 +1,18 @@
+_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
+    tokenizer: ${...tokenizer}
+    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

fusion_bench_config/llama_weighted_average.yaml

@@ -1,26 +0,0 @@
-defaults:
-  - example_config
-  - override method: weighted_average_for_llama
-  - override modelpool: llama_for_causallm
-  - _self_
-modelpool:
-  models:
-    # the pre-trained model (base model) is optional
-    # if not provided, the first model will be used as the base model
-    - name: _pretrained_
-      path: meta-llama/Meta-Llama-3-8B
-    - name: expert_1
-      path: meta-llama/Meta-Llama-3-8B
-    - name: expert_2
-      path: meta-llama/Meta-Llama-3-8B-Instruct
-method:
-  normalize: true # if true, the weights will be normalized before merging
-  weights: # List of weights for each model
-    - 0.5
-    - 0.5
-  # if true, only the backbone of the model will be merged and the head will be keeped as the pre-trained model (if the pre-trained model is provided, otherwise the head of the first model will be used)
-  # if false, the whole model will be merged
-  backbone_only: true
-  merged_model_save_path: null
-  save_tokenizer: true
-  push_to_hub: false