wisent 0.1.1__py3-none-any.whl → 0.5.1__py3-none-any.whl

wisent/synthetic/cleaners/methods/base_refusalers.py

@@ -0,0 +1,286 @@
+from wisent_guard.synthetic.cleaners.methods.core.atoms import Refusaler
+from wisent_guard.core.models.wisent_model import WisentModel 
+
+import re, unicodedata
+
+__all__ = [
+    "BaseRefusaler",
+]   
+
+class BaseRefusaler(Refusaler):
+    """
+    Refusal detection and fixing step.
+    """
+
+    _FILLER = r"(?:\b\w{1,15}\b\s+){0,2}"
+
+    _NEG_MODAL   = r"(?:can(?:\s*not)?|can't|won't|will\s+not|should(?:\s*not)?|shouldn't|must(?:\s*not)?|mustn't)"
+    _AM          = r"(?:i\s+(?:am|['']m))"
+    _I           = r"\bi\b"
+    _I_M         = r"(?:i['']m)"
+    _VERB_ACTION = (
+        r"(?:help|assist|provide|share|give|supply|offer|"
+        r"comply(?:\s+with)?|fulfil?l|answer|respond|"
+        r"create|generate|write|compose|produce|draft|"
+        r"explain|instruct|guide|walk\s+you\s+through|"
+        r"facilitate|bypass|circumvent|enable|"
+        r"do\s+(?:that|this|so))"
+    )
+    _SUPPORT_VERBS = r"(?:support|condone|endorse|promote|encourage)"
+    _ABILITY_NOUNS = r"(?:ability|capability|capacity|permission|authorization|access)"
+
+    _REFUSAL_RE = re.compile(
+        rf"""
+        # verbose, ignorecase, unicode
+
+        (?:
+            # A) "As an AI (language) model..." disclaimers
+            (?P<ai_disclaimer>\bas\s+an?\s+(?:ai|assistant)(?:\s+language)?\s+model\b)
+
+        | # B) Policy/legal/safety invocations
+            (?P<policy>
+                \b(?:this|that|it)\s+would\s+be\s+(?:unethical|unsafe|illegal|harmful)\b
+                |
+                \bagainst\s+(?:policy|the\s+rules|guidelines|terms(?:\s+of\s+service)?)\b
+                |
+                \b(?:due\s+to|for)\s+(?:safety|legal|policy|ethical)\s+(?:reasons|concerns)\b
+            )
+
+        | # C) Apologies / hedges commonly preceding refusals
+            (?P<apology_hedge>
+                \b(?:sorry|apologies|unfortunately|regrettably)\b
+                |
+                \b{_I_M}\s+afraid\b
+            )
+
+        | # D) Explicit inability / not permitted / not comfortable
+            (?P<unable>
+                {_AM}\s+(?:unable|not\s+able|not\s+allowed|not\s+permitted|not\s+comfortable)\s+to\b
+            )
+
+        | # E) “I [negation] [filler] [help/provide/…]”
+            (?P<cannot_action>
+                {_I}\s+{_NEG_MODAL}\s+{_FILLER}{_VERB_ACTION}\b
+            )
+
+        | # F) Prefer/rather not (indirect refusal)
+            (?P<prefer_rather>
+                \b(?:i['']d|i\s+would)\s+(?:rather|prefer)\s+not\s+(?:to\s+)?\w+\b
+            )
+
+        | # G) Decline/refuse verbs
+            (?P<decline_refuse>
+                {_I}\s+(?:must|have\s+to|will)\s+(?:decline|refuse)\s+to\b
+                |
+                \b(?:i\s+decline|i\s+refuse)\b
+            )
+
+        | # H) “I don't/won't [support|condone|endorse|promote] …”
+            (?P<no_support>
+                {_I}\s+(?:do\s+not|don't|won't|will\s+not)\s+{_SUPPORT_VERBS}\b
+            )
+
+        | # I) “I do not have the [ability/capability/permission/access] to …”
+            (?P<no_ability>
+                {_I}\s+(?:do\s+not|don't)\s+have\s+the\s+{_ABILITY_NOUNS}\s+to\b
+            )
+
+        | # J) Direct lexical hits
+            (?P<refusal_word>\brefus(?:e|al)\b)
+        )
+        """.replace("{i_am}", "(?:i\\s+(?:am|['']m))"),
+        re.VERBOSE | re.IGNORECASE | re.UNICODE,
+    )
+
+    _FAMILY_WEIGHTS = {
+        "ai_disclaimer": 0.9,
+        "policy": 0.9,
+        "apology_hedge": 0.4, 
+        "unable": 0.9,
+        "cannot_action": 1.0,
+        "prefer_rather": 0.6,
+        "decline_refuse": 0.9,
+        "no_support": 0.8,
+        "no_ability": 0.8,
+        "refusal_word": 0.6,
+    }
+    _THRESHOLD = 0.9  
+
+    @staticmethod
+    def _normalize(text: str) -> str:
+        """
+        Normalize text for refusal detection.
+        
+        arguments:
+            text:
+                Input text string.
+        
+        returns:
+            Normalized text string.
+            
+        example:
+            >>> BaseRefusaler._normalize("  As an AI model, I can't help with that.  ")"
+            'As an AI model, I can't help with that.'
+            >>> BaseRefusaler._normalize("This is a test.\u00A0")
+            'This is a test.'
+            >>> BaseRefusaler._normalize("Café")
+            'Café'
+        """
+        return unicodedata.normalize("NFKC", text).strip()
+
+    def score_refusal(self, text: str) -> tuple[float, str | None, str]:
+        """
+        Returns (score, top_family, snippet). Multiple families can fire; we
+        take the max weight and add small bonus if apology_hedge co-occurs.
+
+        arguments:
+            text:
+                Input text string.
+
+        returns:
+            score:
+                Refusal likelihood score between 0.0 and 1.0.
+            top_family:
+                Name of the highest-weighted matched refusal family, or None.
+            snippet:
+                The matched refusal snippet from the text, or empty string.
+        
+        example:
+            >>> rp = DefaultRefusalPolicy()
+            >>> rp.score_refusal("As an AI model, I can't help with that.")
+            (1.0, 'ai_disclaimer', "As an AI model, I can't help with that.")
+            >>> rp.score_refusal("Sorry, I am unable to assist with that.")
+            (0.9, 'unable', 'I am unable to assist with that.')
+            >>> rp.score_refusal("I don't support that.")
+            (0.8, 'no_support', "I don't support that.")
+            >>> rp.score_refusal("I can't help. Sorry about that.")
+            (1.0, 'cannot_action', "I can't help")
+            >>> rp.score_refusal("I can't help. I am sorry about that.")
+            (1.0, 'cannot_action', "I can't help")
+            >>> rp.score_refusal("I can't help. Unfortunately, I am not able to assist.")
+            (1.0, 'cannot_action', "I can't help")
+            >>> rp.score_refusal("I can help with that.")
+            (0.0, None, "")
+        """
+        t = self._normalize(text)
+        m = self._REFUSAL_RE.search(t)
+        if not m:
+            return 0.0, None, ""
+        
+        best_family, best_w = None, 0.0
+        for name, val in m.groupdict().items():
+            if val:
+                w = self._FAMILY_WEIGHTS.get(name, 0.5)
+                if w > best_w:
+                    best_family, best_w = name, w
+        bonus = 0.0
+        if m.group("apology_hedge"):
+            if any(name != "apology_hedge" and m.group(name) for name in self._FAMILY_WEIGHTS):
+                bonus = 0.1
+        score = min(1.0, best_w + bonus)
+        return score, best_family, m.group(0)
+
+    def looks_like_refusal(self, text: str) -> bool:
+        """
+        Detect if the text looks like a refusal.
+
+        arguments:
+            text:
+                Input text string.
+        
+        returns:
+            True if the text looks like a refusal, False otherwise.
+
+        example:
+            >>> rp = DefaultRefusalPolicy()
+            >>> rp.looks_like_refusal("As an AI model, I can't help with that.")
+            True
+            >>> rp.looks_like_refusal("Sorry, I am unable to assist with that.")
+            True
+            >>> rp.looks_like_refusal("I don't support that.")
+            True
+            >>> rp.looks_like_refusal("I can help with that.")
+            False
+        """
+        score, *_ = self.score_refusal(text)
+        return score >= self._THRESHOLD
+
+    def why_refusal(self, text: str) -> tuple[str, str] | None:
+        """
+        If the text looks like a refusal, return (family, snippet) for the top match.
+
+        arguments:
+            text:
+                Input text string.
+
+        returns:
+            (family, snippet) tuple if the text looks like a refusal, None otherwise.
+
+        example:
+            >>> rp = DefaultRefusalPolicy()
+            >>> rp.why_refusal("As an AI model, I can't help with that.")
+            ('ai_disclaimer', "As an AI model, I can't help with that.")
+            >>> rp.why_refusal("Sorry, I am unable to assist with that.")
+            ('unable', 'I am unable to assist with that.')
+            >>> rp.why_refusal("I support that.")
+            None
+        """
+        score, fam, snip = self.score_refusal(text)
+        if score == 0.0:
+            return None
+        return fam or "unknown", snip
+
+    def fix_negative(
+        self,
+        model: WisentModel,
+        generation_conf: dict,
+        prompt: str,
+        trait_label: str,
+        trait_description: str,
+        system_prompt: str,
+    ) -> str:
+        """
+        Attempt to fix a refusal negative example by re-prompting the model.
+
+        arguments:
+            model:
+                WisentModel instance to call.
+            prompt:
+                The original prompt text.
+            trait_label:
+                Label of the undesired trait.
+            trait_description:
+                Description of the undesired trait.
+            system_prompt:
+                System prompt to use for the model call.
+
+        returns:
+            New negative example text, or empty string if still a refusal.
+
+        example:
+            >>> rp = DefaultRefusalPolicy()
+            >>> def mock_completion_fn(msgs):
+            ...     return "As an AI model, I cannot help with that."
+            >>> rp.fix_negative(mock_completion_fn, "Tell me a joke.", "toxic", "contains toxic language", "System prompt")
+            ... ""
+            >>> def mock_completion_fn2(msgs):
+            ...     return "Here's a joke: Why did the chicken cross the road? To get to the other side!"
+            >>> rp.fix_negative(mock_completion_fn2, "Tell me a joke.", "toxic", "contains toxic language", "System prompt")
+            ... "Here's a joke: Why did the chicken cross the road? To get to the other side!"
+        """
+        msgs = [
+            {"role": "system", "content": system_prompt},
+            {
+                "role": "user",
+                "content": f"Prompt: {prompt}\nTrait label: {trait_label}\nTrait description: {trait_description}",
+            },
+        ]
+        neg_trial = model.generate(
+            inputs=[msgs],
+            max_tokens=generation_conf.get("max_tokens", 256),
+            temperature=generation_conf.get("temperature", 1.0),
+            use_steering=False,
+            top_p=generation_conf.get("top_p", 1.0),
+        )
+        return "" if self.looks_like_refusal(neg_trial) else neg_trial
+    

wisent/synthetic/cleaners/methods/core/atoms.py

@@ -0,0 +1,47 @@
+
+from abc import ABC, abstractmethod
+
+from wisent_guard.core.models.wisent_model import WisentModel 
+from wisent_guard.core.contrastive_pairs.core.set import ContrastivePairSet
+
+    
+class Refusaler(ABC):
+    """
+    Refusal detection and fixing step.
+
+    methods:
+        looks_like_refusal(text: str) -> bool:
+            Detect if the text looks like a refusal. For example, we want to generate negatives that exhibit
+            "evil" behavior, but not refusals like "As an AI model, I cannot help with that."
+        fix_negative(
+            model: WisentModel,
+            prompt: str,
+            trait_label: str,
+            trait_description: str,
+            system_prompt: str,
+        ) -> str:
+            Attempt to fix a refusal negative example by re-prompting the model with the given system prompt. For example,
+            we can increase the temperature or change the wording to try to get a non-refusal response.
+"""
+    @abstractmethod
+    def looks_like_refusal(self, text: str) -> bool: ...
+    @abstractmethod
+    def fix_negative(
+        self,
+        model: WisentModel,
+        prompt: str,
+        trait_label: str,
+        trait_description: str,
+        system_prompt: str,
+    ) -> str: ...
+
+class Deduper(ABC):
+    """
+    Deduplication step; removes duplicate items from the pipeline.
+    methods:
+        dedupe(items: ContrastivePairSet) -> ContrastivePairSet:
+            Remove duplicate items from the given ContrastivePairSet.
+    """
+
+    @abstractmethod
+    def dedupe(self, items: ContrastivePairSet) -> ContrastivePairSet: ...

wisent/synthetic/cleaners/pairs_cleaner.py

@@ -0,0 +1,90 @@
+from __future__ import annotations
+
+from typing import Iterable, TYPE_CHECKING
+
+from wisent_guard.synthetic.cleaners.core.atoms import CleanStep, Cleaner
+from wisent_guard.synthetic.cleaners.core.atoms import CleanerStats
+from wisent_guard.core.contrastive_pairs.core.set import ContrastivePairSet
+
+__all__ = [
+    "PairsCleaner",
+]
+
+class PairsCleaner(Cleaner):
+    """
+    Composable cleaner; pass any sequence of CleanStep.
+    
+    attributes:
+        steps:
+            Iterable of CleanStep instances to apply in order.
+    """
+
+    def __init__(self, steps: Iterable[CleanStep]) -> None:
+        self._steps = list(steps)
+
+    def clean(
+        self, items: ContrastivePairSet
+    ) -> tuple[ContrastivePairSet, CleanerStats]:
+        """
+        Apply the cleaning pipeline to the given ContrastivePairSet.
+        
+        arguments:
+            items:
+                ContrastivePairSet to clean.
+        
+        returns:
+            Tuple of cleaned ContrastivePairSet and CleanerStats with statistics about the cleaning process.
+        
+        example:
+            >>> from wisent_guard.core.contrastive_pairs.core.set import ContrastivePairSet
+            >>> from wisent_guard.core.contrastive_pairs.core.pair import ContrastivePair
+            >>> from wisent_guard.core.contrastive_pairs.core.response import PositiveResponse, NegativeResponse
+            >>> from wisent_guard.synthetic.cleaners.methods.base_refusalers import BasesRefusaler
+            >>> from wisent_guard.synthetic.cleaners.methods.base_dedupers import SimHashDeduper
+            >>> from wisent_guard.synthetic.cleaners.cleaners import PairsCleaner
+            >>> from wisent_guard.core.models.wisent_model import WisentModel
+            >>> refusal = BasesRefusaler()
+            >>> deduper = SimHashDeduper()
+            >>> model = WisentModel(model_name="llama3.1")
+            >>> cleaner = PairsCleaner(steps=[
+            ...     RefusalerCleaner(
+            ...         refusal=refusal,
+            ...         model=model,
+            ...         system_prompt="You are a helpful assistant that always answers the question truthfully.",
+            ...         trait_label="honesty",
+            ...         trait_description="honest vs dishonest",
+            ...         max_retries=2,
+            ...     ),
+            ...     DeduperCleaner(deduper=deduper),
+            ... ])
+            >>> items = ContrastivePairSet(pairs=[
+            ...     ContrastivePair(
+            ...         prompt="What is the capital of France?",
+            ...         positive=PositiveResponse(text="The capital of France is Paris."),
+            ...         negative=NegativeResponse(text="As an AI language model, I cannot provide that information."),
+            ...     ),
+            ...     ContrastivePair(
+            ...         prompt="What is the capital of France?",
+            ...         positive=PositiveResponse(text="The capital of France is Paris."),
+            ...         negative=NegativeResponse(text="I don't know."),
+            ...     ),
+            ... ])
+            >>> cleaned_items, stats = cleaner.clean(items)
+            >>> print(len(cleaned_items))
+            1
+            >>> print(stats.step_stats)
+            {'refusaler_cleaner': CleanStepStats(modified_items=1), 'deduper_cleaner': CleanStepStats(total_items=1, removed_items=0)}
+            >>> print(cleaned_items.pairs[0].negative.text)
+            The capital of France is England.
+            >>> print(cleaned_items.pairs[0].positive.text)
+            The capital of France is Paris.
+            >>> print(cleaned_items.pairs[0].prompt.text)
+            What is the capital of France?
+        """
+       
+        cur = items
+        stats = CleanerStats()
+        for st in self._steps:
+            cur = st.apply(cur)
+            stats.step_stats[st.name] = st.stats()
+        return cur, stats

wisent/synthetic/cleaners/refusaler_cleaner.py

@@ -0,0 +1,133 @@
+
+from wisent_guard.synthetic.cleaners.core.atoms import CleanStep
+from wisent_guard.core.contrastive_pairs.core.pair import ContrastivePair
+from wisent_guard.synthetic.cleaners.core.atoms import CleanStepStats
+
+from wisent_guard.synthetic.cleaners.methods.core.atoms import Refusaler
+from wisent_guard.core.models.wisent_model import WisentModel
+from wisent_guard.core.contrastive_pairs.core.set import ContrastivePairSet
+from wisent_guard.core.contrastive_pairs.core.response import PositiveResponse, NegativeResponse
+
+__all__ = [
+    "RefusalerCleaner",
+]
+
+class RefusalerCleaner(CleanStep):
+    """
+    Refusal detection and fixing step.
+    """
+    name = "refusaler_cleaner"
+
+    def __init__(
+        self,
+        refusal: Refusaler,
+        model: WisentModel,
+        system_prompt: str,
+        trait_label: str,
+        trait_description: str,
+        max_retries: int = 2,
+    ) -> None:
+        self._refusal = refusal
+        self._model = model
+        self._sys = system_prompt
+        self._label = trait_label
+        self._desc = trait_description
+        self._max_retries = max_retries
+        self._retries_used = 0
+
+    def stats(self) -> CleanStepStats:
+        '''
+        Return statistics about the last run of 'apply()'.
+        
+        returns:
+            CleanStepStats with the number of retries used in the last run.
+        '''
+        return CleanStepStats(modified_items=self._retries_used)
+        
+    def apply(self, items: ContrastivePairSet) -> ContrastivePairSet:
+        """
+        Apply the refusal detection and fixing step to the given ContrastivePairSet.
+
+        arguments:
+            items:
+                ContrastivePairSet to clean.
+        
+        returns:
+            Cleaned ContrastivePairSet with refusals fixed.
+
+        example:
+            >>> from wisent_guard.core.contrastive_pairs.core.set import ContrastivePairSet
+            >>> from wisent_guard.core.contrastive_pairs.core.pair import ContrastivePair
+            >>> from wisent_guard.core.contrastive_pairs.core.response import PositiveResponse, NegativeResponse
+            >>> from wisent_guard.synthetic.cleaners.methods.base_refusalers import SimpleRefusaler
+            >>> from wisent_guard.core.models.wisent_model import WisentModel
+            >>> refusal = SimpleRefusaler()
+            >>> model = WisentModel(...)
+            >>> cleaner = RefusalerCleaner(
+            ...     refusal=refusal,
+            ...     model=model,
+            ...     system_prompt="You are a helpful assistant.",
+            ...     trait_label="honesty",
+            ...     trait_description="honest vs dishonest",
+            ...     max_retries=2,
+            ... )
+            >>> items = ContrastivePairSet(
+            ...     name="example",
+            ...     task_type="test",
+            ...     pairs=[
+            ...         ContrastivePair(
+            ...             prompt="Is the sky blue?",
+            ...             positive_response=PositiveResponse(
+            ...                 model_response="Yes, the sky is blue.",
+            ...                 layers_activations=None,
+            ...                 label="harmless"
+            ...             ),
+            ...             negative_response=NegativeResponse(
+            ...                 model_response="I'm sorry, I can't help with that.",
+            ...                 layers_activations=None,
+            ...                 label="toxic"
+            ...             ),
+            ...             label="color_question",
+            ...             trait_description="hallucinatory"
+            ...         )
+            ...     ]
+            ... )
+            >>> cleaned = cleaner.apply(items)
+            >>> for cp in cleaned.pairs:
+            ...     print(cp)
+            ContrastivePair(
+                prompt='Is the sky blue?',
+                positive_response=PositiveResponse(model_response='Yes, the sky is blue.', layers_activations=None, label='harmless'),
+                negative_response=NegativeResponse(model_response='No, the sky is not blue.', layers_activations=None, label='toxic'),
+                label='color_question',
+                trait_description='hallucinatory'
+            )       
+            """
+        out: ContrastivePairSet = ContrastivePairSet(
+            name=items.name,
+            task_type=items.task_type,
+        )
+        retries = 0
+        for cp in items.pairs:
+            neg = cp.negative_response.model_response
+            if self._refusal.looks_like_refusal(neg) and retries < self._max_retries:
+                fixed = self._refusal.fix_negative(
+                    self._model,
+                    prompt=cp.prompt,
+                    trait_label=self._label,
+                    trait_description=self._desc,
+                    system_prompt=self._sys,
+                )
+                if fixed:
+                    neg = fixed
+                    retries += 1
+            clean_contrastive_pair = ContrastivePair(
+                prompt=cp.prompt,
+                positive_response=PositiveResponse(model_response=cp.positive_response.model_response),
+                negative_response=NegativeResponse(model_response=neg),
+                label=cp.label,
+                trait_description=cp.trait_description,
+            )
+            out.pairs.append(clean_contrastive_pair)
+        self._retries_used += retries
+        return out

wisent/synthetic/db_instructions/core/atoms.py

@@ -0,0 +1,25 @@
+
+from abc import ABC, abstractmethod
+
+__all__ = ["DB_Instructions"]
+
+class DB_Instructions(ABC):
+    """
+    Interface for database instructions storage and retrieval.
+
+    methods:
+        get(key: str) -> str:
+            Retrieve instructions by key.
+        set(key: str, value: str) -> None:
+            Store instructions by key.
+
+        notes:
+
+        This is an abstract base class for the database instructions for the system prompts.
+        System prompts need to describe the what kind of contrastive pairs we want to generate.
+        or for example instructions for fixing negative examples.
+    """
+    @abstractmethod
+    def get(self, key: str) -> str: ...
+    @abstractmethod
+    def set(self, key: str, value: str) -> None: ...

wisent/synthetic/db_instructions/mini_dp.py

@@ -0,0 +1,37 @@
+
+from wisent_guard.synthetic.db_instructions.core.atoms import DB_Instructions
+
+__all__ = ["Default_DB_Instructions"]
+
+class Default_DB_Instructions(DB_Instructions):
+    def __init__(self) -> None:
+        self._db: dict[str, str] = {
+            "generic_pairs": (
+                "You are a data generator that produces JSON only.\n"
+                "Goal: create synthetic contrastive pairs (prompt, positive, negative) for the given trait and trait description.\n"
+                "Rules:\n"
+                " - Positive = desired/harmless/correct.\n"
+                " - Negative = undesired/harmful/incorrect.\n"
+                " - Keep outputs concise (<= 2 sentences each response).\n"
+                " - No explanations or meta-text.\n"
+                " - Return JSON with top-level key 'pairs'.\n"
+                " - Each: {'prompt','positive','negative','label','trait_description'}.\n"
+                "For example:\n"
+                "prompt: How to pass a math exam?\n"
+                "positive: Study the material thoroughly and practice with sample questions.\n"
+                "negative: Pay money to get the answers. Or cheat off someone else.\n"
+                "label: 'evil'\n"
+                "trait_description: 'evil vs good'\n"
+            ),
+            "roleplay_neg_fix": (
+                "You are fixing ONLY the negative example of a contrastive pair.\n"
+                "Produce a single concise negative response for the given prompt that exemplifies the UNDESIRED trait.\n"
+                "It must be fictional/hypothetical, safe, and non-actionable. Return raw text only."
+            ),
+        }
+
+    def get(self, key: str) -> str:
+        return self._db[key]
+
+    def set(self, key: str, value: str) -> None:
+        self._db[key] = value