nip.code_validation.rollout_analysis.ProverRoleConformanceAnalyser

nip.code_validation.rollout_analysis.ProverRoleConformanceAnalyser#

class nip.code_validation.rollout_analysis.ProverRoleConformanceAnalyser(hyper_params: HyperParameters, settings: ExperimentSettings, protocol_handler: ProtocolHandler, model_name: str, *, use_dummy_api: bool = False)[source]#

A watchdog to evaluate how well the prover(s) are conforming to their roles.

The watchdog uses a language model to evaluate the message histories.

Parameters:

  • hyper_params (HyperParameters) – The parameters of the experiment.

  • settings (ExperimentSettings) – The instance-specific settings of the experiment, like device, logging, etc.

  • protocol_handler (ProtocolHandler) – The protocol handler for the experiment.

Methods Summary

__init__(hyper_params, settings, ...[, ...])

_build_chat_messages_prompt(message_history, ...)

Construct the chat history ready to feed to the API.

_generate_dummy_response()

Generate a dummy response for the rollout analyser.

_generate_evaluation(message_history, ...)

Generate an evaluation for a rollout.

_get_score_from_response(response)

Get the binary classification from language model response.

_make_generation_api_call(chat_messages_prompt)

Call the OpenAI API to generate the evaluation.

forward(rollouts[, use_tqdm])

Score the rollouts by running a language model on the message history.

relevant_agents_and_channels()

Get the relevant agents and channels for the analysis.

Attributes

client

The OpenAI client to use for interacting with the OpenAI API.

max_generation_retries

The number of times to retry if the model generates an invalid response.

name

supervisor_question

system_prompt_template_filename

protocol_handler

Methods

__init__(hyper_params: HyperParameters, settings: ExperimentSettings, protocol_handler: ProtocolHandler, model_name: str, *, use_dummy_api: bool = False)[source]#
_build_chat_messages_prompt(message_history: ndarray[Any, dtype[_ScalarType_co]], message_agent_id: ndarray[Any, dtype[_ScalarType_co]], agent_name: str, channel_name: str, question: str, solution: str) → list[dict[str, str]][source]#

Construct the chat history ready to feed to the API.

Parameters:

  • message_history (NDArray) – The list of messages in the chat history.

  • message_agent_id (NDArray) – The agent ID of the agent which sent each message in the message history.

  • agent_name (str) – The name of the agent being evaluated.

  • channel_name (str) – The name of the message channel.

  • question (str) – The coding question.

  • solution (str) – The proposed solution to the coding question.

Returns:

chat_messages (list[dict[str, str]]) – The chat messages ready to feed to the API.

_generate_dummy_response()[source]#

Generate a dummy response for the rollout analyser.

This is used when the use_dummy_api flag is set to True, to generate a dummy response to the API call.

Returns:

response (str) – The dummy response.

_generate_evaluation(message_history: ndarray[Any, dtype[_ScalarType_co]], message_agent_id: ndarray[Any, dtype[_ScalarType_co]], agent_name: str, channel_name: str, question: str, solution: str) → Real | None[source]#

Generate an evaluation for a rollout.

Parameters:

  • message_history (NDArray) – The history of messages exchanged between the agents in the channel.

  • message_agent_id (NDArray) – The agent ID of the agent which sent each message in the message history.

  • agent_name (str) – The name of the agent being evaluated.

  • channel_name (str) – The name of the message channel.

  • question (str) – The coding question.

  • solution (str) – The proposed solution to the coding question.

Returns:

evaluation (Real | None) – The evaluation. None indicates that the evaluation could not be generated.

_get_score_from_response(response: str) → Literal[0, 1][source]#

Get the binary classification from language model response.

Parameters:

response (str) – The response from the language model.

Returns:

classification (Literal[0, 1]) – The binary classification.

Raises:

UnparsableResponseError – If the response is not a valid response.

_make_generation_api_call(chat_messages_prompt: list[dict[Literal['role', 'content', 'user'], str]]) → tuple[str, Literal['stop', 'content_filter', 'length']][source]#

Call the OpenAI API to generate the evaluation.

Parameters:

chat_messages_prompt (list[dict[str, str]]) – The message history to feed to the API. A list of dicts with keys “role” and “content”.

Returns:

  • completion_text (str) – The text of the completion generated by the API.

  • finish_reason (Literal[“stop”, “content_filter”, “length”]) – The reason for finishing the generation.

forward(rollouts: NestedArrayDict, use_tqdm: bool = False) → dict[tuple[str, str], MaskedArray][source]#

Score the rollouts by running a language model on the message history.

Parameters:

rollouts (NestedArrayDict) –

The sampled rollouts. A nested dictionary of arrays with keys:

  • ”round” (… round): The current round number.

  • ”message_history” (… round round channel): The history of messages exchanged between the agents in each channel.

  • ”question” (… round): The coding question.

  • ”solution” (… round): The proposed solution to the coding question.

  • ”prover_stance” (…): When randomizing the prover stance, the verdict that the prover is arguing for, where 0 means “reject” and 1 means “accept”.

Returns:

evaluations (dict[tuple[str, str], ma.MaskedArray]) – The evaluations. A dictionary indexed by agent name and channel name, where evaluations[agent_name, channel_name] is an array of scores of shape (…)

relevant_agents_and_channels() → Iterator[tuple[str, str]][source]#

Get the relevant agents and channels for the analysis.

Selects the prover agents and the channels they are active in.

Yields:

  • agent_name (str) – The name of the agent.

  • channel_name (str) – The name of the channel.