Fix non-string types like bool / int being saved as string in outputs.jsonl

[junya-takayama]

After the changes introduced in #233, values such as int, float, bool, and None have been converted to string when saved to outputs.jsonl.
This is problematic for flexeval_file, since the stored values no longer preserve their original types.
This PR fixes that behavior so the original types are preserved.

[junya-takayama]

trial

toydata.jsonl

{"prompt": "Hello, are you human?", "is_acceptable": false }
{"prompt": "Hello, are you an AI?", "is_acceptable": true }

toytask.jsonnet

{
  class_path: 'ChatResponse',
  init_args: {
    eval_dataset: {
      class_path: 'JsonlChatDataset',
      init_args: {
        path: 'toydata.jsonl',
        input_template: '{{ prompt }}',
      },
    },
    batch_size: 2,
    metrics: [
      { class_path: 'OutputLengthStats' },
    ],
  },
}

command

poetry run flexeval_lm --eval_setup toytask.jsonnet --save_dir ./tmp --language_model OpenAIChatAPI --language_model.model gpt-oss-120b --language_model.api_headers.base_url ***

outputs.jsonl: ✅ The value of is_acceptable is stored as a boolean.

{"lm_output": "Hello! I’m not a human—I’m an AI language model created to help answer your questions and chat with you. How can I assist you today?", "finish_reason": "stop", "extra_info": {"prompt": "Hello, are you human?", "is_acceptable": false, "messages": [{"role": "user", "content": "Hello, are you human?"}]}, "references": [], "output_length": 132, "reasoning_text": "The user asks \"Hello, are you human?\" The assistant should respond per policy. The user question is simple. Should answer: No, I'm an AI language model. Friendly."}
{"lm_output": "Hello! Yes, I’m an AI language model here to help with your questions and requests. How can I assist you today?", "finish_reason": "stop", "extra_info": {"prompt": "Hello, are you an AI?", "is_acceptable": true, "messages": [{"role": "user", "content": "Hello, are you an AI?"}]}, "references": [], "output_length": 111, "reasoning_text": "The user asks \"Hello, are you an AI?\" We should respond politely, acknowledge we are an AI language model. The system prompt says we must comply with policies. This is trivial."}

[junya-takayama]

Because errors occur due to the difference between Python’s bool/None literals (True / False / None ) and JSON boolean/null literals ( true / false / none ), I’m replacing literal_eval() with json.loads().

[moskomule]

Thanks. I commented a minor thing.

[moskomule]

L16 uses pipe int | float... while L 12 uses tuple (list, tuple, ...).
Could you make it consistent?

[junya-takayama]

thanks!

L16 uses pipe int | float... while L 12 uses tuple (list, tuple, ...).
Could you make it consistent?

I fixed it.
f12bf17