Skip to content

Mighty eureka runner

mighty.mighty_runners.mighty_eureka_runner #

Eureka: LLM-coded reward functions.

Eureka #

Eureka(cfg: DictConfig)

Bases: MightyRunner

Eureka: LLM-coded reward functions.

Source code in mighty/mighty_runners/mighty_eureka_runner.py 
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
def __init__(self, cfg: DictConfig) -> None:
    super().__init__(cfg)
    self.iterations = cfg.iterations
    local_llm = cfg.local_llm
    if local_llm:
        self.model, self.prompting_function = get_llm(cfg.model_name)
        self.generation_kwargs = OmegaConf.to_container(cfg.generation_kwargs)
    else:
        raise NotImplementedError("Currently only local models supported.")
    if "prompt_dir" in cfg:
        self.prompt_dir = cfg.prompt_dir
    else:
        self.prompt_dir = Path(__file__).parent.absolute() / "eureka_prompts"

    with open(self.prompt_dir / "user_prompt.txt", "r") as f:
        self.user_prompt = f.read()
    with open(self.prompt_dir / "system_prompt.txt", "r") as f:
        self.system_prompt = f.read()
    # TODO: adapt these to task
    self.prompt = [
        {"role": "system", "content": self.system_prompt},
        {"role": "user", "content": self.user_prompt},
    ]