[FEATURE] Refactor of structured generation and use schemas defined in a dataset

[plaguss]

Description

This PR updates the code to generate structured outputs either from a dataset with schemas via StructuredGeneration or via a single structured output schema via TextGeneration:

The following pipeline works with the 3 different types of LLMs (llamacpp with outlines, inference endpoints which uses outlines but abstracted, and mistral that works with instructor):

    with Pipeline(name="structured-generation") as pipeline:
        load_data = LoadDataFromDicts(
            name="load_data",
            data=[
                {
                    "instruction": "Generate a character from a RPG game.",
                    "structured_output": {
                        "format": "json",
                        "schema": Character.model_json_schema(),
                    },
                },
                {
                    "instruction": "Generate an animal from a zoo.",
                    "structured_output": {
                        "format": "json",
                        "schema": Animal.model_json_schema(),
                    },
                },
                # regex doesn't work with instructor, it's related to outlines based engines
                # {
                #     "instruction": "What's the weather like today in Seattle in Celsius degrees?",
                #     "structured_output": {
                #         "format": "regex",
                #         "schema": "(\\d{1,2})°C",
                #     },
                # },
            ],
        )
        task = StructuredGeneration(
            # llm = LlamaCppLLM(
            #     model_path=str(Path.home() / model_path),  # type: ignore
            #     n_gpu_layers=-1,
            #     n_ctx=1024,
            # ),
            # llm=InferenceEndpointsLLM(
            #     model_id="CohereForAI/c4ai-command-r-plus",
            #     tokenizer_id="CohereForAI/c4ai-command-r-plus",
            #     api_key=os.getenv("HF_API_TOKEN"),  # type: ignore
            # ),
            llm=MistralLLM(
                model="open-mixtral-8x22b",
                api_key=os.getenv("MISTRAL_API_KEY"),  # type: ignore
            ),
            use_system_prompt=False,
            output_mappings={"model_name": "generation_model"},
        )

        load_data >> task

[codspeed-hq]

CodSpeed Performance Report

Merging #688 will not alter performance

_{Comparing inference-endpoints-structured-gen-grammar (170066a) with develop (ce8dde8)}

Summary

✅ 1 untouched benchmarks