themis-eval 0.1.0__py3-none-any.whl → 0.2.0__py3-none-any.whl

themis/generation/turn_strategies.py

@@ -0,0 +1,393 @@
+"""Turn strategies for multi-turn conversations.
+
+This module provides strategies for determining the next turn in a conversation.
+Strategies can be fixed (predefined sequences), dynamic (generated based on context),
+or interactive.
+
+Examples:
+    # Fixed sequence
+    strategy = FixedSequenceTurnStrategy([
+        "What is 2+2?",
+        "What about 3+3?",
+        "And 5+5?"
+    ])
+
+    # Dynamic strategy
+    def planner(context):
+        if len(context) < 2:
+            return "Can you explain more?"
+        return None  # Stop
+
+    strategy = DynamicTurnStrategy(planner)
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Callable, Protocol
+
+from themis.core import conversation as conv
+from themis.core import entities as core_entities
+
+
+class TurnStrategy(Protocol):
+    """Strategy for determining the next turn in a conversation.
+
+    A turn strategy decides what the user's next message should be
+    based on the current conversation state and the last model response.
+    """
+
+    def next_turn(
+        self,
+        context: conv.ConversationContext,
+        last_record: core_entities.GenerationRecord,
+    ) -> str | None:
+        """Determine the next user message.
+
+        Args:
+            context: Current conversation context
+            last_record: Last generation record
+
+        Returns:
+            Next user message, or None to end conversation
+        """
+        ...
+
+
+@dataclass
+class FixedSequenceTurnStrategy:
+    """Pre-determined sequence of user messages.
+
+    This strategy iterates through a fixed list of user messages,
+    useful for scripted conversations or testing.
+
+    Examples:
+        strategy = FixedSequenceTurnStrategy([
+            "Hello!",
+            "How are you?",
+            "Goodbye!"
+        ])
+    """
+
+    messages: list[str]
+    _index: int = 0
+
+    def next_turn(
+        self,
+        context: conv.ConversationContext,
+        last_record: core_entities.GenerationRecord,
+    ) -> str | None:
+        """Return next message from sequence.
+
+        Args:
+            context: Current conversation context
+            last_record: Last generation record
+
+        Returns:
+            Next message or None if sequence exhausted
+        """
+        if self._index >= len(self.messages):
+            return None
+
+        message = self.messages[self._index]
+        self._index += 1
+        return message
+
+    def reset(self) -> None:
+        """Reset strategy to beginning of sequence."""
+        self._index = 0
+
+
+@dataclass
+class DynamicTurnStrategy:
+    """Generate next message based on conversation state.
+
+    This strategy uses a function to dynamically determine the next
+    user message based on the conversation context.
+
+    Examples:
+        def planner(context, record):
+            outputs = [msg.content for msg in context.get_messages_by_role("assistant")]
+            if "error" in outputs[-1].lower():
+                return "Can you try again?"
+            elif len(context) >= 10:
+                return None  # Stop after 10 messages
+            else:
+                return "Please continue."
+
+        strategy = DynamicTurnStrategy(planner)
+    """
+
+    planner: Callable[
+        [conv.ConversationContext, core_entities.GenerationRecord], str | None
+    ]
+
+    def next_turn(
+        self,
+        context: conv.ConversationContext,
+        last_record: core_entities.GenerationRecord,
+    ) -> str | None:
+        """Generate next message using planner function.
+
+        Args:
+            context: Current conversation context
+            last_record: Last generation record
+
+        Returns:
+            Next message or None to stop
+        """
+        return self.planner(context, last_record)
+
+
+@dataclass
+class RepeatUntilSuccessTurnStrategy:
+    """Repeat the same question until getting a successful response.
+
+    This strategy is useful for testing robustness or debugging.
+
+    Examples:
+        strategy = RepeatUntilSuccessTurnStrategy(
+            question="What is 2+2?",
+            success_checker=lambda output: "4" in output,
+            max_attempts=5
+        )
+    """
+
+    question: str
+    success_checker: Callable[[str], bool]
+    max_attempts: int = 5
+    _attempts: int = 0
+
+    def next_turn(
+        self,
+        context: conv.ConversationContext,
+        last_record: core_entities.GenerationRecord,
+    ) -> str | None:
+        """Repeat question until success or max attempts.
+
+        Args:
+            context: Current conversation context
+            last_record: Last generation record
+
+        Returns:
+            Question or None if success/max attempts reached
+        """
+        # Check if this is first turn
+        if self._attempts == 0:
+            self._attempts += 1
+            return self.question
+
+        # Check if last response was successful
+        if last_record.output:
+            if self.success_checker(last_record.output.text):
+                return None  # Success, stop
+
+        # Check if we've exhausted attempts
+        if self._attempts >= self.max_attempts:
+            return None  # Give up
+
+        self._attempts += 1
+        return self.question
+
+    def reset(self) -> None:
+        """Reset attempt counter."""
+        self._attempts = 0
+
+
+@dataclass
+class ConditionalTurnStrategy:
+    """Choose next message based on conditions.
+
+    This strategy evaluates conditions and returns different messages
+    based on which condition matches.
+
+    Examples:
+        strategy = ConditionalTurnStrategy(
+            conditions=[
+                (lambda ctx, rec: "error" in rec.output.text.lower(), "Please try again."),
+                (lambda ctx, rec: len(ctx) >= 5, None),  # Stop after 5 turns
+            ],
+            default="Continue."
+        )
+    """
+
+    conditions: list[
+        tuple[
+            Callable[[conv.ConversationContext, core_entities.GenerationRecord], bool],
+            str | None,
+        ]
+    ]
+    default: str | None = None
+
+    def next_turn(
+        self,
+        context: conv.ConversationContext,
+        last_record: core_entities.GenerationRecord,
+    ) -> str | None:
+        """Evaluate conditions and return matching message.
+
+        Args:
+            context: Current conversation context
+            last_record: Last generation record
+
+        Returns:
+            Message from first matching condition, or default
+        """
+        for condition, message in self.conditions:
+            try:
+                if condition(context, last_record):
+                    return message
+            except Exception:
+                # Skip conditions that fail
+                continue
+
+        return self.default
+
+
+@dataclass
+class ChainedTurnStrategy:
+    """Chain multiple strategies together.
+
+    This strategy tries strategies in sequence until one returns
+    a non-None message.
+
+    Examples:
+        strategy = ChainedTurnStrategy([
+            FixedSequenceTurnStrategy(["Hello", "How are you?"]),
+            DynamicTurnStrategy(lambda ctx, rec: "Goodbye" if len(ctx) > 5 else None)
+        ])
+    """
+
+    strategies: list[TurnStrategy]
+
+    def next_turn(
+        self,
+        context: conv.ConversationContext,
+        last_record: core_entities.GenerationRecord,
+    ) -> str | None:
+        """Try each strategy until one returns a message.
+
+        Args:
+            context: Current conversation context
+            last_record: Last generation record
+
+        Returns:
+            First non-None message, or None if all return None
+        """
+        for strategy in self.strategies:
+            message = strategy.next_turn(context, last_record)
+            if message is not None:
+                return message
+
+        return None
+
+
+# Helper functions for creating common strategies
+
+
+def create_qa_strategy(questions: list[str]) -> FixedSequenceTurnStrategy:
+    """Create a simple Q&A strategy from a list of questions.
+
+    Args:
+        questions: List of questions to ask
+
+    Returns:
+        FixedSequenceTurnStrategy with questions
+    """
+    return FixedSequenceTurnStrategy(messages=questions)
+
+
+def create_max_turns_strategy(
+    max_turns: int, message: str = "Continue."
+) -> DynamicTurnStrategy:
+    """Create strategy that stops after max turns.
+
+    Args:
+        max_turns: Maximum number of turns
+        message: Message to send each turn
+
+    Returns:
+        DynamicTurnStrategy that stops after max_turns
+    """
+
+    def planner(
+        context: conv.ConversationContext, record: core_entities.GenerationRecord
+    ) -> str | None:
+        if len(context) >= max_turns:
+            return None
+        return message
+
+    return DynamicTurnStrategy(planner=planner)
+
+
+def create_keyword_stop_strategy(
+    keywords: list[str], message: str = "Continue."
+) -> DynamicTurnStrategy:
+    """Create strategy that stops when any keyword appears in response.
+
+    Args:
+        keywords: List of keywords to trigger stop
+        message: Message to send each turn
+
+    Returns:
+        DynamicTurnStrategy that stops on keywords
+    """
+
+    def planner(
+        context: conv.ConversationContext, record: core_entities.GenerationRecord
+    ) -> str | None:
+        if record.output:
+            text_lower = record.output.text.lower()
+            if any(kw.lower() in text_lower for kw in keywords):
+                return None
+        return message
+
+    return DynamicTurnStrategy(planner=planner)
+
+
+# Prompt perturbation and seed helpers for robustness sweeps
+
+import random
+
+
+def set_sampling_seed(task_metadata: dict[str, object], seed: int) -> dict[str, object]:
+    """Attach a deterministic seed to task metadata for providers that support it.
+
+    This does not enforce provider behavior but offers a convention: 'sampling_seed'.
+    """
+    md = dict(task_metadata)
+    md["sampling_seed"] = int(seed)
+    return md
+
+
+def perturb_prompt(text: str, *, seed: int | None = None, max_changes: int = 2) -> str:
+    """Apply small, semantics-preserving perturbations to a prompt.
+
+    Changes include optional punctuation tweaks and inserting polite filler words.
+    """
+    rng = random.Random(seed)
+    t = text
+    changes = 0
+    # Optional punctuation swap
+    if "?" in t and changes < max_changes and rng.random() < 0.5:
+        t = t.replace("?", "??", 1)
+        changes += 1
+    # Optional polite filler insertion
+    fillers = ["please", "kindly", "if possible"]
+    if changes < max_changes and rng.random() < 0.5:
+        words = t.split()
+        if words:
+            idx = rng.randint(0, len(words) - 1)
+            words.insert(idx, rng.choice(fillers))
+            t = " ".join(words)
+            changes += 1
+    return t
+
+
+def create_prompt_variants(base_text: str, *, count: int, seed: int) -> list[str]:
+    """Create multiple perturbed variants of a base prompt with deterministic seeding."""
+    rng = random.Random(seed)
+    return [
+        perturb_prompt(base_text, seed=rng.randint(0, 1_000_000))
+        for _ in range(max(1, count))
+    ]

themis/generation/types.py

@@ -0,0 +1,9 @@
+"""Backwards-compatible aliases for core entities."""
+
+from themis.core import entities as core_entities
+
+SamplingParameters = core_entities.SamplingConfig
+ModelOutput = core_entities.ModelOutput
+GenerationError = core_entities.ModelError
+GenerationRequest = core_entities.GenerationTask
+GenerationResult = core_entities.GenerationRecord

themis/integrations/huggingface.py

@@ -0,0 +1,72 @@
+from __future__ import annotations
+
+import json
+from dataclasses import asdict, is_dataclass
+from pathlib import Path
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from huggingface_hub import HfApi
+else:
+    try:
+        from huggingface_hub import HfApi
+    except ImportError:
+        HfApi = None  # type: ignore
+
+from themis.config.schema import HuggingFaceHubConfig
+from themis.core.entities import ExperimentReport
+
+
+def to_dict(obj):
+    if is_dataclass(obj):
+        return asdict(obj)
+    if hasattr(obj, "to_dict"):
+        return obj.to_dict()
+    if isinstance(obj, (list, tuple)):
+        return [to_dict(item) for item in obj]
+    if isinstance(obj, dict):
+        return {key: to_dict(value) for key, value in obj.items()}
+    return obj
+
+
+class HuggingFaceHubUploader:
+    def __init__(self, config: HuggingFaceHubConfig):
+        if HfApi is None:
+            raise ImportError(
+                "huggingface_hub is not installed. Install with: pip install huggingface_hub"
+            )
+        self.config = config
+        self.api = HfApi()
+
+    def upload_results(self, report: ExperimentReport, storage_path: Path) -> None:
+        if not self.config.enable or not self.config.repository:
+            return
+
+        report_dict = to_dict(report)
+
+        # Upload the full report as a JSON file
+        report_path = storage_path / "report.json"
+        with open(report_path, "w") as f:
+            json.dump(report_dict, f, indent=4)
+
+        self.api.upload_file(
+            path_or_fileobj=str(report_path),
+            path_in_repo=f"{report.metadata.get('run_id')}/report.json",
+            repo_id=self.config.repository,
+            repo_type="dataset",
+        )
+
+        # Upload individual generation results
+        for record in report.generation_results:
+            record_dict = to_dict(record)
+            record_path = (
+                storage_path / f"{record.task.metadata.get('dataset_id')}.json"
+            )
+            with open(record_path, "w") as f:
+                json.dump(record_dict, f, indent=4)
+            self.api.upload_file(
+                path_or_fileobj=str(record_path),
+                path_in_repo=f"{report.metadata.get('run_id')}/generations/{record.task.metadata.get('dataset_id')}.json",
+                repo_id=self.config.repository,
+                repo_type="dataset",
+            )

themis/integrations/wandb.py

@@ -0,0 +1,77 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    import wandb
+else:
+    try:
+        import wandb
+    except ImportError:
+        wandb = None  # type: ignore
+
+from themis.config.schema import WandbConfig
+from themis.core.entities import ExperimentReport
+
+
+class WandbTracker:
+    def __init__(self, config: WandbConfig):
+        if wandb is None:
+            raise ImportError(
+                "wandb is not installed. Install with: pip install wandb"
+            )
+        self.config = config
+
+    def init(self, experiment_config: dict) -> None:
+        if not self.config.enable:
+            return
+        wandb.init(
+            project=self.config.project,
+            entity=self.config.entity,
+            tags=self.config.tags,
+            config=experiment_config,
+        )
+
+    def log_results(self, report: ExperimentReport) -> None:
+        if not self.config.enable:
+            return
+        summary = {
+            "total_samples": report.metadata.get("total_samples"),
+            "successful_generations": report.metadata.get("successful_generations"),
+            "failed_generations": report.metadata.get("failed_generations"),
+            "evaluation_failures": report.metadata.get("evaluation_failures"),
+        }
+        for name, aggregate in report.evaluation_report.metrics.items():
+            summary[f"{name}_mean"] = aggregate.mean
+        wandb.summary.update(summary)
+
+        records_table = wandb.Table(
+            columns=[
+                "sample_id",
+                "prompt",
+                "raw_response",
+                "parsed_response",
+                "error",
+                "metric_scores",
+            ]
+        )
+        for record in report.generation_results:
+            eval_record = next(
+                (
+                    r
+                    for r in report.evaluation_report.records
+                    if r.sample_id == record.task.metadata.get("dataset_id")
+                ),
+                None,
+            )
+            records_table.add_data(
+                record.task.metadata.get("dataset_id"),
+                record.task.prompt,
+                [resp.text for resp in record.responses],
+                eval_record.parsed_response if eval_record else None,
+                record.error.message if record.error else None,
+                {s.metric_name: s.value for s in eval_record.scores}
+                if eval_record
+                else None,
+            )
+        wandb.log({"generation_results": records_table})