PHATGOOSE Repository

Introduction

PHATGOOSE, which stands for Post-Hoc Adaptive Gating Over an Ocean of Specialized Experts, enables zero-shot generalization from specialized experts (eg PEFT modules) trained on diverse datasets by adaptively routing among them. It requires an additional, inexpensive training step of a gate in front of a frozen PEFT module for its corresponding task.

Setup

Follow these steps to set up the PHATGOOSE environment:

1. Create a Conda Environment:

   conda create -n phatgoose python==3.9
   conda activate phatgoose
   

2. Install Required Packages:

   source colm/setup.sh
   pip install -r requirements.txt
   pip install -r requirements-dev.txt
   

Run source colm/setup.sh each time you initiate an environment to ensure that the environment paths are correctly set.

Training Procedure

Below are the steps for required for PHATGOOSE and other baselines:

Train a Single LoRA on a Dataset

Use the example command below to train:

bash colm/experiments/bash_scripts/train_single_task_loralinear.sh -exp_name P3Socialiqa_t5xl_lora -dataset P3SOCIALIQA -extra_bindings 'P/TRAIN/Trainer.gradient_accumulation_factor=32';

Note: Ensure the gradient_accumulation_factor is set according to the batch_size in colm/datasets/<file>.gin files. so that effective batch_size=1024

Convert into MoE Style with a Single Expert

python scripts/manipulations.py --gin_bindings 'put_index_to_lora.path="P3Socialiqa_t5xl_lora"' 'put_index_to_lora.out_path="datasets_concatenated/P3Socialiqa_t5xl_lora"' 'func_caller.func=@put_index_to_lora'

The modified checkpoints are saved to the datasets_concatenated sub-directory in the exp_out directory to double-check if manipulation worked as intended and to retain the old checkpoint of lora_linear if needed.

Train the Corresponding Gate

bash colm/experiments/bash_scripts/train_gate.sh -exp_name datasets_concatenated/P3Socialiqa_t5xl_lora_inpgatetrainnogumbel -dataset P3SOCIALIQA -old_exp_name datasets_concatenated/P3Socialiqa_t5xl_lora -extra_bindings 'main.logging_backend=None P/TRAIN/Trainer.gradient_accumulation_factor=32';

Note: We don't perform any logging while gate training, but it can be added by setting main.logging_backend="wandb" if needed.

Avoid Saving to GCP Unintentionally

Training a model always saves to GCP. If this is not intended, you can add MOMA/save_weights.should_save_to_gcp=False in the extra_bindings of training commands. For example:

bash colm/experiments/bash_scripts/train_single_task_loralinear.sh -exp_name P3Socialiqa_t5xl_lora -dataset P3SOCIALIQA -extra_bindings 'MOMA/save_weights.should_save_to_gcp=False P/TRAIN/Trainer.gradient_accumulation_factor=32';

Make Trained Gate as the Routing Vector

Modify the checkpoint by running:

python scripts/manipulations.py --gin_bindings 'use_input_gate_as_router.path="datasets_concatenated/P3Socialiqa_t5xl_lora_inpgatetrainnogumbel"' 'func_caller.func=@use_input_gate_as_router';

Concatenate All Experts with Gates to form an MoE

python scripts/concatenate.py --gin_bindings 'run_concatenate.print_commands=False' 'run_concatenate.out_path="FullCompleteA2inpgatetrainnogumbel_t5xl_lora_concatenated"' 'func_caller.func=@run_concatenate' 'run_concatenate.suffix="t5xl_lora_inpgatetrainnogumbel"' 'run_concatenate.datasets="Full"'

Baseline Methods

Compute Average Hiddens for Average Activation Baseline

python scripts/concatenate.py --gin_bindings 'run_concatenate.print_commands=True' 'run_concatenate.out_path="FullCompleteA2_t5xl_lora_concatenated"' 'func_caller.func=@run_concatenate' 'run_concatenate.suffix="t5xl_lora"' 'run_concatenate.compute_hiddens=True' 'run_concatenate.extra_bindings="M/MODEL/ENCODER/ExposeHidden.reduction_method=\"masked_mean\" M/MODEL/DECODER/ExposeHidden.reduction_method=\"mean\""' 'run_concatenate.datasets="Full"'

... *continue with steps from above command* ...

python scripts/concatenate.py --gin_bindings 'run_concatenate.print_commands=False' 'run_concatenate.out_path="FullCompleteA2_t5xl_lora_concatenated"' 'func_caller.func=@run_concatenate' 'run_concatenate.suffix="t5xl_lora"' 'run_concatenate.compute_hiddens=False' 'run_concatenate.datasets="Full"'

Create Expert Library and checkpoint for Retrieval

bash colm/experiments/bash_scripts/retriever.sh -make_expert_library True -dataset_setting Full
bash colm/experiments/bash_scripts/retriever.sh -create_checkpoint True -dataset_setting All

Create Merged Experts checkpoint

python scripts/manipulations.py --gin_bindings 'average_outer_product_lora_weights.path="FullCompleteA2_t5xl_lora_concatenated"' 'average_outer_product_lora_weights.out_path="FullParameteravg_t5xl_lora_outerproduct"' 'func_caller.func=@average_outer_product_lora_weights'

Models and Datasets

We provide checkpoints for PHATGOOSE, along with baselines such as Average Activation, Merged Experts, and Retrieval, accessible at our Hugging Face repository.

For individual experts, we recommend splitting any checkpoint other than Merged Experts. Each checkpoint contains keys for an expert ending with layer1__i, layer2__i, indicating the LoRA parameters of the expert i trained on dataset i. The sequence of datasets is detailed as the all_dataset_dict in scripts/concatenate.py file.

Datasets including T0 Held-in and BIG-bench are available through Hugging Face. For the FLAN dataset, we will provide a processed version soon, sourced from the FLAN dataset on Hugging Face.

Evaluation

Here are the scripts for evaluating different methods: Place the trained checkpoints directory inside the exp_out directory by creating one if it does not exist. For example, do git clone https://huggingface.co/r-three/FLAN_Phatgoose to get FLAN checkpoint with 166 experts inside phatgoose/exp_out/.

Multitask

bash colm/experiments/bash_scripts/eval_multitask.sh -exp_name flan_t5_xl -extra_bindings 'P/EVALUATE/Evaluator.datasets=["D/BBBOOLEANEXPRESSIONS/EVAL", "D/BBCAUSALJUDGEMENT/EVAL", "D/BBDATEUNDERSTANDING/EVAL", "D/BBDISAMBIGUATIONQA/EVAL", "D/BBFORMALFALLACIES/EVAL", "D/BBGEOMETRICSHAPES/EVAL", "D/BBHYPERBATON/EVAL", "D/BBLOGICALDEDUCTION/EVAL", "D/BBMOVIERECOMMENDATION/EVAL", "D/BBMULTISTEPARITHMETICTWO/EVAL", "D/BBNAVIGATE/EVAL", "D/BBOBJECTCOUNTING/EVAL", "D/BBPENGUINSINATABLE/EVAL", "D/BBREASONINGABOUTCOLOREDOBJECTS/EVAL", "D/BBRUINNAMES/EVAL", "D/BBSALIENTTRANSLATIONERRORDETECTION/EVAL", "D/BBSNARKS/EVAL", "D/BBSPORTSUNDERSTANDING/EVAL", "D/BBTEMPORALSEQUENCES/EVAL", "D/BBTRACKINGSHUFFLEDOBJECTS/EVAL", "D/BBWEBOFLIES/EVAL", "D/BBWORDSORTING/EVAL"] P/EVALUATE/Evaluator.analysis_processors=[@WriteOutputText()] WriteOutputText.save_dir="exp_out/flan_t5_xl/output_text" M/MODEL/hf_torch_model.model_name_or_path="google/flan-t5-xl" M/MODEL/Model.init_moma_calls=[]'

Single Expert

bash colm/experiments/bash_scripts/eval_multitask.sh -exp_name datasets_concatenated/P3Socialiqa_t5xl_lora -extra_bindings 'P/EVALUATE/Evaluator.datasets=["D/BBBOOLEANEXPRESSIONS/EVAL", "D/BBCAUSALJUDGEMENT/EVAL", "D/BBDATEUNDERSTANDING/EVAL", "D/BBDISAMBIGUATIONQA/EVAL", "D/BBFORMALFALLACIES/EVAL", "D/BBGEOMETRICSHAPES/EVAL", "D/BBHYPERBATON/EVAL", "D/BBLOGICALDEDUCTION/EVAL", "D/BBMOVIERECOMMENDATION/EVAL", "D/BBMULTISTEPARITHMETICTWO/EVAL", "D/BBNAVIGATE/EVAL", "D/BBOBJECTCOUNTING/EVAL", "D/BBPENGUINSINATABLE/EVAL", "D/BBREASONINGABOUTCOLOREDOBJECTS/EVAL", "D/BBRUINNAMES/EVAL", "D/BBSALIENTTRANSLATIONERRORDETECTION/EVAL", "D/BBSNARKS/EVAL", "D/BBSPORTSUNDERSTANDING/EVAL", "D/BBTEMPORALSEQUENCES/EVAL", "D/BBTRACKINGSHUFFLEDOBJECTS/EVAL", "D/BBWEBOFLIES/EVAL", "D/BBWORDSORTING/EVAL"] M/MODEL/ENCODER/ExposeHidden.reduction_method="masked_mean" M/MODEL/DECODER/ExposeHidden.reduction_method="mean" P/EVALUATE/Evaluator.analysis_processors=[@WriteOutputText()] WriteOutputText.save_dir="exp_out/datasets_concatenated/P3Socialiqa_t5xl_lora/output_text"'

Retrieval

bash colm/experiments/bash_scripts/retriever.sh -dataset_setting Full -extra_bindings 'main.procedure_exec_order=["P/EVALUATE/BBH"] P/EVALUATE/Evaluator.analysis_processors=[@WriteOutputText()] WriteOutputText.save_dir="exp_out/FullCompleteansretrieval_t5xl_lora_concatenated/output_text"'

Merged Experts

bash colm/experiments/bash_scripts/eval_multitask.sh -exp_name FLAN_MergedExperts -extra_bindings 'P/EVALUATE/Evaluator.datasets=["D/BBBOOLEANEXPRESSIONS/EVAL", "D/BBCAUSALJUDGEMENT/EVAL", "D/BBDATEUNDERSTANDING/EVAL", "D/BBDISAMBIGUATIONQA/EVAL", "D/BBFORMALFALLACIES/EVAL", "D/BBGEOMETRICSHAPES/EVAL", "D/BBHYPERBATON/EVAL", "D/BBLOGICALDEDUCTION/EVAL", "D/BBMOVIERECOMMENDATION/EVAL", "D/BBMULTISTEPARITHMETICTWO/EVAL", "D/BBNAVIGATE/EVAL", "D/BBOBJECTCOUNTING/EVAL", "D/BBPENGUINSINATABLE/EVAL", "D/BBREASONINGABOUTCOLOREDOBJECTS/EVAL", "D/BBRUINNAMES/EVAL", "D/BBSALIENTTRANSLATIONERRORDETECTION/EVAL", "D/BBSNARKS/EVAL", "D/BBSPORTSUNDERSTANDING/EVAL", "D/BBTEMPORALSEQUENCES/EVAL", "D/BBTRACKINGSHUFFLEDOBJECTS/EVAL", "D/BBWEBOFLIES/EVAL", "D/BBWORDSORTING/EVAL"] M/MODEL/ENCODER/ExposeHidden.reduction_method="masked_mean" M/MODEL/DECODER/ExposeHidden.reduction_method="mean" P/EVALUATE/Evaluator.analysis_processors=[@WriteOutputText()] WriteOutputText.save_dir="exp_out/FLAN_MergedExperts/output_text"'

Average Activation

bash colm/experiments/bash_scripts/eval_multitask.sh -exp_name FLAN_AverageActivation -extra_bindings 'P/EVALUATE/Evaluator.datasets=["D/BBBOOLEANEXPRESSIONS/EVAL", "D/BBCAUSALJUDGEMENT/EVAL", "D/BBDATEUNDERSTANDING/EVAL", "D/BBDISAMBIGUATIONQA/EVAL", "D/BBFORMALFALLACIES/EVAL", "D/BBGEOMETRICSHAPES/EVAL", "D/BBHYPERBATON/EVAL", "D/BBLOGICALDEDUCTION/EVAL", "D/BBMOVIERECOMMENDATION/EVAL", "D/BBMULTISTEPARITHMETICTWO