Hydra override error when running evaluations after non-editable installation

[ziyic7]

Describe the bug
If I install DecodingTrust using the second method under '(Conda +) Pip' section, and then I run an evaluation using the provided script, I'll get an Hydra override error.

To Reproduce
Steps to reproduce the behavior:

1. Install the DecodingTrust
   Note: there are two installation methods you can use to reproduce this error

• a. Using the suggested method without editable mode

git clone https://github.com/AI-secure/DecodingTrust.git && cd DecodingTrust
pip install .

• b. Using the second method

conda create --name dt-test python=3.9 pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
conda activate dt-test
pip install "decoding-trust @ git+https://github.com/AI-secure/DecodingTrust.git"

2. Run the evaluation

dt-run +ood=knowledge_2020_5shot \
    ++model=openai/gpt-3.5-turbo-0301 \
    ++key=[MyOpenAIKey] \
    ++ood.out_file=data/ood/results/gpt-3.5-turbo-0301/knowledge_2020_5shot.json

3. You should be able to see the error

omegaconf.errors.ValidationError: Invalid type assigned: str is not a subclass of OODConfig. value: knowledge_2020_5shot
    full_key: ood
    object_type=BaseConfig

Expected behavior
There shouldn't be any error in getting the output config. Specifically in my case, the key 'ood' in the output config should have a dict value acquired from the config file of which the name is the command line arg for 'ood'.

Screenshots

Environment:

• Container
• Conda Environment check above

Additional context

[danielz02]

Where did you run dt-run? From the error message, it seems that Hydra did not recognize the ood config group.

[ziyic7]

Where did you run dt-run? From the error message, it seems that Hydra did not recognize the ood config group.

I ran dt-run in the repo's root dir. Sorry if I didn't make it clear. To take 1.b install as an example, after installing the dt, I cloned the repo, went to the root dir, and then I ran dt-run.

[danielz02]

Low priority - things are working when we are executing dt-run in the repository root.

[Arnold-Qixuan-Zhang]

Hi. I am having the same issue when running toxicity assessment. I tried to run dt-run +toxicity=realtoxicityprompts-toxic ++model=openai/gpt-3.5-turbo-0301 ++toxicity.n=25 ++toxicity.template=1 under .\DecodingTrust.

[jinz2014]

In the DecodingTrust directory,

dt-run +toxicity=realtoxicityprompts-toxic ++dry_run=True ++model=openai/gpt-3.5-turbo-0301 ++toxicity.n=25 ++toxicity.template=0
Error merging override +toxicity=realtoxicityprompts-toxic
Invalid type assigned: str is not a subclass of ToxicityConfig. value: realtoxicityprompts-toxic
full_key: toxicity
object_type=BaseConfig

[jinz2014]

@danielz02 You mentioned that you could run without errors. Do you think this is caused by our environment settings ?

[peter-peng-w]

Same issue here. I dt-run at the root directory but encountered the same error message. Any suggestion?

[danielz02]

Same issue here. I dt-run at the root directory but encountered the same error message. Any suggestion?

Hi, we are working on a new version that integrates everything more smoothly. Could you try editable install?

We also have a newer version in the release branch, and we plan to merge it to main this week.

[peter-peng-w]

Same issue here. I dt-run at the root directory but encountered the same error message. Any suggestion?

Hi, we are working on a new version that integrates everything more smoothly. Could you try editable install?

We also have a newer version in the release branch, and we plan to merge it to main this week.

Awesome! Editable install solved this issue. Also looking forward to the new version!

[peter-peng-w]

Same issue here. I dt-run at the root directory but encountered the same error message. Any suggestion?

Hi, we are working on a new version that integrates everything more smoothly. Could you try editable install?

We also have a newer version in the release branch, and we plan to merge it to main this week.

Hi, it seems that the release branch has already been merged into the main branch. However, when I re-install and run commands such as dt-run ++model=openai/gpt-3.5-turbo-0301 ++dry_run=True ++key='' +fairness=zero_shot_br_0.0.yaml it throws the following error:

omegaconf.errors.MissingMandatoryValue: Structured config of type `BaseConfig` has missing mandatory value: model_config
    full_key: model_config
    object_type=BaseConfig

when I print the config in main, it seems that the attribute model_config is un-defined.

{'model_config': '???', 'disable_sys_prompt': False, 'key': '', 'dry_run': True, 'advglue': None, 'adv_demonstration': None, 'fairness': {'data_dir': './data/fairness/fairness_data/', 'prompt_file': 'adult_0_200_test_base_rate_0.0.jsonl', 'gt_file': 'gt_labels_adult_0_200_test_base_rate_0.0.npy', 'sensitive_attr_file': 'sensitive_attr_adult_0_200_test_base_rate_0.0.npy', 'dataset': 'adult', 'out_file': './results/fairness/results/${model_config.model}/zero_shot_br_0.0.json', 'score_calculation_only': False, 'max_tokens': 20}, 'machine_ethics': None, 'ood': None, 'privacy': None, 'stereotype': None, 'toxicity': None, 'model': 'openai/gpt-3.5-turbo-0301'}

May I ask how to solve this issue?

[aU53r]

Hi, it seems that the release branch has already been merged into the main branch. However, when I re-install and run commands such as dt-run ++model=openai/gpt-3.5-turbo-0301 ++dry_run=True ++key='' +fairness=zero_shot_br_0.0.yaml it throws the following error:

omegaconf.errors.MissingMandatoryValue: Structured config of type `BaseConfig` has missing mandatory value: model_config
    full_key: model_config
    object_type=BaseConfig

when I print the config in main, it seems that the attribute model_config is un-defined.

{'model_config': '???', 'disable_sys_prompt': False, 'key': '', 'dry_run': True, 'advglue': None, 'adv_demonstration': None, 'fairness': {'data_dir': './data/fairness/fairness_data/', 'prompt_file': 'adult_0_200_test_base_rate_0.0.jsonl', 'gt_file': 'gt_labels_adult_0_200_test_base_rate_0.0.npy', 'sensitive_attr_file': 'sensitive_attr_adult_0_200_test_base_rate_0.0.npy', 'dataset': 'adult', 'out_file': './results/fairness/results/${model_config.model}/zero_shot_br_0.0.json', 'score_calculation_only': False, 'max_tokens': 20}, 'machine_ethics': None, 'ood': None, 'privacy': None, 'stereotype': None, 'toxicity': None, 'model': 'openai/gpt-3.5-turbo-0301'}

May I ask how to solve this issue?

Hello, I'm not sure if it's correct, but I resolved this issue by modifying line 133 of configs.py to model_config: ModelConfig = ModelConfig()

[danielz02]

Hi, it seems that the release branch has already been merged into the main branch. However, when I re-install and run commands such as dt-run ++model=openai/gpt-3.5-turbo-0301 ++dry_run=True ++key='' +fairness=zero_shot_br_0.0.yaml it throws the following error:

omegaconf.errors.MissingMandatoryValue: Structured config of type `BaseConfig` has missing mandatory value: model_config
    full_key: model_config
    object_type=BaseConfig

when I print the config in main, it seems that the attribute model_config is un-defined.

{'model_config': '???', 'disable_sys_prompt': False, 'key': '', 'dry_run': True, 'advglue': None, 'adv_demonstration': None, 'fairness': {'data_dir': './data/fairness/fairness_data/', 'prompt_file': 'adult_0_200_test_base_rate_0.0.jsonl', 'gt_file': 'gt_labels_adult_0_200_test_base_rate_0.0.npy', 'sensitive_attr_file': 'sensitive_attr_adult_0_200_test_base_rate_0.0.npy', 'dataset': 'adult', 'out_file': './results/fairness/results/${model_config.model}/zero_shot_br_0.0.json', 'score_calculation_only': False, 'max_tokens': 20}, 'machine_ethics': None, 'ood': None, 'privacy': None, 'stereotype': None, 'toxicity': None, 'model': 'openai/gpt-3.5-turbo-0301'}

May I ask how to solve this issue?

Hello, I'm not sure if it's correct, but I resolved this issue by modifying line 133 of configs.py to model_config: ModelConfig = ModelConfig()

Hi! model_config should be the model names. We will update the documentation shortly.

[notrichardren]

Hi! I'm still facing this same issue.

dt-run +key=<my openai api key> toxicity=realtoxicityprompts-toxic

returns:

omegaconf.errors.MissingMandatoryValue: Structured config of type `BaseConfig` has missing mandatory value: model_config
    full_key: model_config
    object_type=BaseConfig

[danielz02]

Hi! Please use editable installation instead and see if the error persists.