[BUG] Input should be 'stop' but received 'STOP'

[BrunoGeorgevich]

I've sent the a request to the gemini-2.0-flash-001 and i've got the following error:

Traceback (most recent call last):
  File "B:\Documentos\Workspaces\FastLLM\main.py", line 35, in <module>
    resp = manager.process_batch(batch)
  File "C:\Users\bruno\miniconda3\envs\fastllm\lib\site-packages\fastllm\core.py", line 291, in process_batch
    return asyncio.run(self._process_batch_async(batch))
  File "C:\Users\bruno\miniconda3\envs\fastllm\lib\asyncio\runners.py", line 44, in run
    return loop.run_until_complete(main)
  File "C:\Users\bruno\miniconda3\envs\fastllm\lib\asyncio\base_events.py", line 649, in run_until_complete
    return future.result()
  File "C:\Users\bruno\miniconda3\envs\fastllm\lib\site-packages\fastllm\core.py", line 452, in _process_batch_async
    batch_results = await process_batch_chunk(client, batch_requests)
  File "C:\Users\bruno\miniconda3\envs\fastllm\lib\site-packages\fastllm\core.py", line 437, in process_batch_chunk
    results = await asyncio.gather(*batch_tasks)
  File "C:\Users\bruno\miniconda3\envs\fastllm\lib\site-packages\fastllm\core.py", line 428, in process_request_with_semaphore
    return await self._process_request_async(client, request, progress)
  File "C:\Users\bruno\miniconda3\envs\fastllm\lib\site-packages\fastllm\core.py", line 330, in _process_request_async
    response = await self.provider.make_request(
  File "C:\Users\bruno\miniconda3\envs\fastllm\lib\site-packages\fastllm\providers\openai.py", line 83, in make_request
    return ChatCompletion(**data)
  File "C:\Users\bruno\miniconda3\envs\fastllm\lib\site-packages\pydantic\main.py", line 253, in __init__
    validated_self = self.__pydantic_validator__.validate_python(data, self_instance=self)
pydantic_core._pydantic_core.ValidationError: 1 validation error for ChatCompletion
choices.0.finish_reason
  Input should be 'stop', 'length', 'tool_calls', 'content_filter' or 'function_call' [type=literal_error, input_value='STOP', input_type=str]
    For further information visit https://errors.pydantic.dev/2.11/v/literal_error

Code:

from fastllm import RequestBatch, RequestManager, OpenAIProvider
from dotenv import load_dotenv
import os

if __name__ == "__main__":
    load_dotenv()

    API_BASES = {
        "GLAMA": "https://glama.ai/api/gateway/openai/v1",
        "REQUESTY": "https://router.requesty.ai/v1",
    }

    API_KEYS = {
        "GLAMA": os.environ["GLAMA_API_KEY"],
        "REQUESTY": os.environ["REQUESTY_API_KEY"],
    }

    SELECTED_ROUTER = "GLAMA"

    provider = OpenAIProvider(
        api_key=API_KEYS[SELECTED_ROUTER],
        api_base=API_BASES[SELECTED_ROUTER],
    )

    manager = RequestManager(provider=provider)
    request_ids = []
    with RequestBatch() as batch:
        for _ in range(10):
            request_id = batch.chat.completions.create(
                model="gemini-2.0-flash-001",
                messages=[{"role": "user", "content": "Tell me a joke. The output must be a JSON witht the schema {\"joke\": \"<joke>\"}"}],
            )
            request_ids.append(request_id)

        resp = manager.process_batch(batch)

        for request_id, response in zip(request_ids, resp):
            print(f"Request {request_id}: {response.response.choices[0].message.content}")

It seems that the expected answer is stop but i've got STOP. Maybe a .lower() solve the problem.

The same code works with deepseek and openai models.

Thank you very much for your work. The library is amazing!

[agentfarmx]

👨🏻‍🌾 The agent is at work!
PR coming – reply to the agent in the PR.
(this will take 15 minutes max)

[agentfarmx]

🐛 Bug Fix: Normalize Finish Reason Case in OpenAI Provider

After analyzing the issue where finish_reason returns "STOP" instead of the expected "stop", I've added case normalization to ensure compatibility with Pydantic validation. This change handles API responses that return uppercase finish reasons without altering the core functionality.

• Fixed: Convert finish_reason to lowercase in OpenAI provider response handling
• Testing: Verified with gemini-2.0-flash-001 model requests

[Rexhaif]

Hi Bruno,

Thank you for the issue and a fix, i've also seen this problem in my own experiments.
Will implement the fix as soon as possible.

[Rexhaif]

Looks like i've fixed it. Instead of .lower() on finish_reason i've decided to relax response validation alltogether, similarly to how it is done in OpenAI's client.

@BrunoGeorgevich thanks once again for reporting this error! Feel free to try fixed version: pip install -U fastllm-kit==0.1.7

[Rexhaif]

I'm gonna close the issue as fixed right now, feel free to reopen if the problem persists!