Can't enable Openai reasoning models thinking mode

[SharifMAbdullah]

Describe the bug

The tutorial suggests:

For gpt-oss models like gpt-oss-20b and gpt-oss-120b, you can control the reasoning effort using the extra_body parameter:

inference_parameters = ChatCompletionInferenceParams(
    extra_body={"reasoning_effort": "high"}
)

But this doesn't work for any of the gpt models, even though it has a reasoning={"effort": "medium"} parameter described in official docs.(Please see here.

Steps/Code to reproduce bug
I obtained the model ID and provider by running:

config_builder.info.display(InfoType.MODEL_CONFIGS)

────────────────────────────────────────────────── Model Configs ──────────────────────────────────────────────────
┏━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Alias            ┃ Model                             ┃ Provider ┃ Inference Parameters                          ┃
┡━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ nvidia-text      │ nvidia/nemotron-3-nano-30b-a3b    │ nvidia   │ generation_type=chat-completion,              │
│                  │                                   │          │ max_parallel_requests=4, temperature=1.00,    │
│                  │                                   │          │ top_p=1.00                                    │
│ nvidia-reasoning │ openai/gpt-oss-20b                │ nvidia   │ generation_type=chat-completion,              │
│                  │                                   │          │ max_parallel_requests=4, temperature=0.35,    │
│                  │                                   │          │ top_p=0.95                                    │
│ nvidia-vision    │ nvidia/nemotron-nano-12b-v2-vl    │ nvidia   │ generation_type=chat-completion,              │
│                  │                                   │          │ max_parallel_requests=4, temperature=0.85,    │
│                  │                                   │          │ top_p=0.95                                    │
│ nvidia-embedding │ nvidia/llama-3.2-nv-embedqa-1b-v2 │ nvidia   │ generation_type=embedding,                    │
│                  │                                   │          │ max_parallel_requests=4,                      │
│                  │                                   │          │ extra_body={'input_type': 'query'},           │
│                  │                                   │          │ encoding_format=float                         │
│ openai-text      │ gpt-4.1                           │ openai   │ generation_type=chat-completion,              │
│                  │                                   │          │ max_parallel_requests=4, temperature=0.85,    │
│                  │                                   │          │ top_p=0.95                                    │
│ openai-reasoning │ gpt-5                             │ openai   │ generation_type=chat-completion,              │
│                  │                                   │          │ max_parallel_requests=4, temperature=0.35,    │
│                  │                                   │          │ top_p=0.95                                    │
│ openai-vision    │ gpt-5                             │ openai   │ generation_type=chat-completion,              │
│                  │                                   │          │ max_parallel_requests=4, temperature=0.85,    │
│                  │                                   │          │ top_p=0.95                                    │
│ openai-embedding │ text-embedding-3-large            │ openai   │ generation_type=embedding,                    │
│                  │                                   │          │ max_parallel_requests=4,                      │
│                  │                                   │          │ encoding_format=float                         │
└──────────────────┴───────────────────────────────────┴──────────┴───────────────────────────────────────────────┘

Then used the following configs:

MODEL_PROVIDER = "openai"
MODEL_ID = "openai/gpt-oss-20b"
MODEL_ALIAS = "openai-text-abc"

model_configs = [
    ModelConfig(
        alias=MODEL_ALIAS,
        model=MODEL_ID,
        provider=MODEL_PROVIDER,
        inference_parameters=ChatCompletionInferenceParams(
            temperature=1.0,
            top_p=1.0,
            max_tokens=2048,
            extra_body={"reasoning_effort": "high"}
        ),
    )
]

But when I run a preview:
preview = data_designer.preview(config_builder, num_records=1)
I get this error:

[02:42:35] [INFO] 🖼️ Preview generation in progress
[02:42:35] [INFO] ✅ Validation passed
[02:42:35] [INFO] ⛓️ Sorting column configs into a Directed Acyclic Graph
[02:42:35] [INFO] 🩺 Running health checks for models...
[02:42:35] [INFO]   |-- 👀 Checking 'openai/gpt-oss-20b' in provider named 'openai' for model alias 'openai-text-abc'...
[02:42:43] [ERROR]   |-- ❌ Failed!
---------------------------------------------------------------------------
ModelBadRequestError                      Traceback (most recent call last)
[/usr/local/lib/python3.12/dist-packages/data_designer/interface/data_designer.py](https://github.com/NVIDIA-NeMo/DataDesigner/issues/177#) in preview(self, config_builder, num_records)
    234         try:
--> 235             raw_dataset = builder.build_preview(num_records=num_records)
    236             processed_dataset = builder.process_preview(raw_dataset)

7 frames[/usr/local/lib/python3.12/dist-packages/data_designer/engine/dataset_builders/column_wise_builder.py](https://github.com/NVIDIA-NeMo/DataDesigner/issues/177#) in build_preview(self, num_records)
    124     def build_preview(self, *, num_records: int) -> pd.DataFrame:
--> 125         self._run_model_health_check_if_needed()
    126 

[/usr/local/lib/python3.12/dist-packages/data_designer/engine/dataset_builders/column_wise_builder.py](https://github.com/NVIDIA-NeMo/DataDesigner/issues/177#) in _run_model_health_check_if_needed(self)
    204         if any(column_type_is_model_generated(config.column_type) for config in self.single_column_configs):
--> 205             self._resource_provider.model_registry.run_health_check(
    206                 list(set(config.model_alias for config in self.llm_generated_column_configs))

[/usr/local/lib/python3.12/dist-packages/data_designer/engine/models/registry.py](https://github.com/NVIDIA-NeMo/DataDesigner/issues/177#) in run_health_check(self, model_aliases)
    130                 logger.error("  |-- ❌ Failed!")
--> 131                 raise e
    132 

[/usr/local/lib/python3.12/dist-packages/data_designer/engine/models/registry.py](https://github.com/NVIDIA-NeMo/DataDesigner/issues/177#) in run_health_check(self, model_aliases)
    116                 elif model.model_generation_type == GenerationType.CHAT_COMPLETION:
--> 117                     model.generate(
    118                         prompt="Hello!",

[/usr/local/lib/python3.12/dist-packages/data_designer/engine/models/errors.py](https://github.com/NVIDIA-NeMo/DataDesigner/issues/177#) in wrapper(model_facade, *args, **kwargs)
    253             )
--> 254             handle_llm_exceptions(
    255                 e, model_facade.model_name, model_facade.model_provider_name, purpose=kwargs.get("purpose")

[/usr/local/lib/python3.12/dist-packages/data_designer/engine/models/errors.py](https://github.com/NVIDIA-NeMo/DataDesigner/issues/177#) in handle_llm_exceptions(exception, model_name, model_provider_name, purpose)
    160         case BadRequestError():
--> 161             raise err_msg_parser.parse_bad_request_error(exception) from None
    162 

ModelBadRequestError:   |----------
  | Cause: The request for model 'openai/gpt-oss-20b' was found to be malformed or missing required parameters while running health checks.
  | Solution: Check your request parameters and try again.
  |----------

During handling of the above exception, another exception occurred:

DataDesignerGenerationError               Traceback (most recent call last)
[/tmp/ipython-input-4124882777.py](https://github.com/NVIDIA-NeMo/DataDesigner/issues/177#) in <cell line: 0>()
----> 1 preview = data_designer.preview(config_builder, num_records=1)

[/usr/local/lib/python3.12/dist-packages/data_designer/interface/data_designer.py](https://github.com/NVIDIA-NeMo/DataDesigner/issues/177#) in preview(self, config_builder, num_records)
    236             processed_dataset = builder.process_preview(raw_dataset)
    237         except Exception as e:
--> 238             raise DataDesignerGenerationError(f"🛑 Error generating preview dataset: {e}")
    239 
    240         dropped_columns = raw_dataset.columns.difference(processed_dataset.columns)

DataDesignerGenerationError: 🛑 Error generating preview dataset:   |----------
  | Cause: The request for model 'openai/gpt-oss-20b' was found to be malformed or missing required parameters while running health checks.
  | Solution: Check your request parameters and try again.
  |----------

Expected behavior

Expected to enable the thinking mode.

Additional context

I am running this on google colab

[johnnygreco]

Hi @SharifMAbdullah – Thanks for reporting this.

It looks like you are using the default openai provider. This is a bit confusing, but the openai/gpt-oss-20b model referenced in the docs is actually hosted by the nvidia model provider. It's also the model associated with the nvidia-reasoning model alias, as the table you printed shows (note that the provider is listed as nvidia in that table).

To close out this issue, we should update our docs to make this more clear, as well as try to add more details to that error message. Thanks again for calling this out!

Please let us know if you have any other questions.

cc @nabinchha

[SharifMAbdullah]

Sorry forgot to mention, this happened yesterday when I wrote MODEL_PROVIDER = "openai", MODEL_ID = "gpt-5". But worked fine if I didn't add the extra_body parameter. But today it seems there's a whole different issue, probably some breaking changes in a dependency library when I try to run the "getting started" notebook:

[18:30:09] [INFO] 🕵️ Preview generation in progress

---------------------------------------------------------------------------

AttributeError                            Traceback (most recent call last)

[/tmp/ipython-input-3235631537.py](https://localhost:8080/#) in <cell line: 0>()
     28 
     29 # Run a preview to generate sample records
---> 30 preview_results = data_designer.preview(config_builder=config_builder)
     31 
     32 # Display a sample record

3 frames

[/usr/local/lib/python3.12/dist-packages/data_designer/engine/models/litellm_overrides.py](https://localhost:8080/#) in apply_litellm_patches()
    155 
    156     # Workaround for the litellm issue described in https://github.com/BerriAI/litellm/issues/9792
--> 157     litellm.litellm_core_utils.logging_callback_manager.LoggingCallbackManager.MAX_CALLBACKS = DEFAULT_MAX_CALLBACKS
    158 
    159     quiet_noisy_logger("httpx")

AttributeError: module 'litellm.litellm_core_utils' has no attribute 'logging_callback_manager'

[johnnygreco]

Hi @SharifMAbdullah – indeed, litellm pushed an update that required a quick patch, which we published yesterday. It should be working now, but unfortunately there seems to be a bunch of new warnings associated with the litellm update when using Python 3.12 and 3.13. If they become too annoying, reverting your litellm version to the previous patch gets rid of them.