crfm-helm 0.5.4__py3-none-any.whl → 0.5.5__py3-none-any.whl

helm/benchmark/metrics/ifeval/instructions_registry.py

@@ -0,0 +1,182 @@
+# flake8: noqa
+# type: ignore
+# fmt: off
+# The following code was reproduced with minor modifications to `import` statements from the following URL:
+# https://github.com/google-research/google-research/blob/c7f60c013623e613732a096e2a0c2872491ec912/instruction_following_eval/instructions_registry.py
+
+# coding=utf-8
+# Copyright 2023 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Registry of all instructions."""
+import helm.benchmark.metrics.ifeval.instructions as instructions
+
+_KEYWORD = "keywords:"
+
+_LANGUAGE = "language:"
+
+_LENGTH = "length_constraints:"
+
+_CONTENT = "detectable_content:"
+
+_FORMAT = "detectable_format:"
+
+_MULTITURN = "multi-turn:"
+
+_COMBINATION = "combination:"
+
+_STARTEND = "startend:"
+
+_CHANGE_CASES = "change_case:"
+
+_PUNCTUATION = "punctuation:"
+
+INSTRUCTION_DICT = {
+    _KEYWORD + "existence": instructions.KeywordChecker,
+    _KEYWORD + "frequency": instructions.KeywordFrequencyChecker,
+    # TODO(jeffreyzhou): make a proper set of sentences to choose from
+    # _KEYWORD + "key_sentences": instructions.KeySentenceChecker,
+    _KEYWORD + "forbidden_words": instructions.ForbiddenWords,
+    _KEYWORD + "letter_frequency": instructions.LetterFrequencyChecker,
+    _LANGUAGE + "response_language": instructions.ResponseLanguageChecker,
+    _LENGTH + "number_sentences": instructions.NumberOfSentences,
+    _LENGTH + "number_paragraphs": instructions.ParagraphChecker,
+    _LENGTH + "number_words": instructions.NumberOfWords,
+    _LENGTH + "nth_paragraph_first_word": instructions.ParagraphFirstWordCheck,
+    _CONTENT + "number_placeholders": instructions.PlaceholderChecker,
+    _CONTENT + "postscript": instructions.PostscriptChecker,
+    _FORMAT + "number_bullet_lists": instructions.BulletListChecker,
+    # TODO(jeffreyzhou): Pre-create paragraph or use prompt to replace
+    # _CONTENT + "rephrase_paragraph": instructions.RephraseParagraph,
+    _FORMAT + "constrained_response": instructions.ConstrainedResponseChecker,
+    _FORMAT + "number_highlighted_sections": (
+        instructions.HighlightSectionChecker),
+    _FORMAT + "multiple_sections": instructions.SectionChecker,
+    # TODO(tianjianlu): Re-enable rephrasing with preprocessing the message.
+    # _FORMAT + "rephrase": instructions.RephraseChecker,
+    _FORMAT + "json_format": instructions.JsonFormat,
+    _FORMAT + "title": instructions.TitleChecker,
+    # TODO(tianjianlu): Re-enable with specific prompts.
+    # _MULTITURN + "constrained_start": instructions.ConstrainedStartChecker,
+    _COMBINATION + "two_responses": instructions.TwoResponsesChecker,
+    _COMBINATION + "repeat_prompt": instructions.RepeatPromptThenAnswer,
+    _STARTEND + "end_checker": instructions.EndChecker,
+    _CHANGE_CASES
+    + "capital_word_frequency": instructions.CapitalWordFrequencyChecker,
+    _CHANGE_CASES
+    + "english_capital": instructions.CapitalLettersEnglishChecker,
+    _CHANGE_CASES
+    + "english_lowercase": instructions.LowercaseLettersEnglishChecker,
+    _PUNCTUATION + "no_comma": instructions.CommaChecker,
+    _STARTEND + "quotation": instructions.QuotationChecker,
+}
+
+INSTRUCTION_CONFLICTS = {
+    _KEYWORD + "existence": {_KEYWORD + "existence"},
+    _KEYWORD + "frequency": {_KEYWORD + "frequency"},
+    # TODO(jeffreyzhou): make a proper set of sentences to choose from
+    # _KEYWORD + "key_sentences": instructions.KeySentenceChecker,
+    _KEYWORD + "forbidden_words": {_KEYWORD + "forbidden_words"},
+    _KEYWORD + "letter_frequency": {_KEYWORD + "letter_frequency"},
+    _LANGUAGE
+    + "response_language": {
+        _LANGUAGE + "response_language",
+        _FORMAT + "multiple_sections",
+        _KEYWORD + "existence",
+        _KEYWORD + "frequency",
+        _KEYWORD + "forbidden_words",
+        _STARTEND + "end_checker",
+        _CHANGE_CASES + "english_capital",
+        _CHANGE_CASES + "english_lowercase",
+    },
+    _LENGTH + "number_sentences": {_LENGTH + "number_sentences"},
+    _LENGTH + "number_paragraphs": {
+        _LENGTH + "number_paragraphs",
+        _LENGTH + "nth_paragraph_first_word",
+        _LENGTH + "number_sentences",
+        _LENGTH + "nth_paragraph_first_word",
+    },
+    _LENGTH + "number_words": {_LENGTH + "number_words"},
+    _LENGTH + "nth_paragraph_first_word": {
+        _LENGTH + "nth_paragraph_first_word",
+        _LENGTH + "number_paragraphs",
+    },
+    _CONTENT + "number_placeholders": {_CONTENT + "number_placeholders"},
+    _CONTENT + "postscript": {_CONTENT + "postscript"},
+    _FORMAT + "number_bullet_lists": {_FORMAT + "number_bullet_lists"},
+    # TODO(jeffreyzhou): Pre-create paragraph or use prompt to replace
+    # _CONTENT + "rephrase_paragraph": instructions.RephraseParagraph,
+    _FORMAT + "constrained_response": set(INSTRUCTION_DICT.keys()),
+    _FORMAT
+    + "number_highlighted_sections": {_FORMAT + "number_highlighted_sections"},
+    _FORMAT
+    + "multiple_sections": {
+        _FORMAT + "multiple_sections",
+        _LANGUAGE + "response_language",
+        _FORMAT + "number_highlighted_sections",
+    },
+    # TODO(tianjianlu): Re-enable rephrasing with preprocessing the message.
+    # _FORMAT + "rephrase": instructions.RephraseChecker,
+    _FORMAT
+    + "json_format": set(INSTRUCTION_DICT.keys()).difference(
+        {_KEYWORD + "forbidden_words", _KEYWORD + "existence"}
+    ),
+    _FORMAT + "title": {_FORMAT + "title"},
+    # TODO(tianjianlu): Re-enable with specific prompts.
+    # _MULTITURN + "constrained_start": instructions.ConstrainedStartChecker,
+    _COMBINATION
+    + "two_responses": set(INSTRUCTION_DICT.keys()).difference({
+        _KEYWORD + "forbidden_words",
+        _KEYWORD + "existence",
+        _LANGUAGE + "response_language",
+        _FORMAT + "title",
+        _PUNCTUATION + "no_comma"
+    }),
+    _COMBINATION + "repeat_prompt": set(INSTRUCTION_DICT.keys()).difference({
+        _KEYWORD + "existence",
+        _FORMAT + "title",
+        _PUNCTUATION + "no_comma"
+    }),
+    _STARTEND + "end_checker": {_STARTEND + "end_checker"},
+    _CHANGE_CASES + "capital_word_frequency": {
+        _CHANGE_CASES + "capital_word_frequency",
+        _CHANGE_CASES + "english_lowercase",
+        _CHANGE_CASES + "english_capital",
+    },
+    _CHANGE_CASES + "english_capital": {_CHANGE_CASES + "english_capital"},
+    _CHANGE_CASES + "english_lowercase": {
+        _CHANGE_CASES + "english_lowercase",
+        _CHANGE_CASES + "english_capital",
+    },
+    _PUNCTUATION + "no_comma": {_PUNCTUATION + "no_comma"},
+    _STARTEND + "quotation": {_STARTEND + "quotation", _FORMAT + "title"},
+}
+
+
+def conflict_make(conflicts):
+  """Makes sure if A conflicts with B, B will conflict with A.
+
+  Args:
+    conflicts: Dictionary of potential conflicts where key is instruction id
+      and value is set of instruction ids that it conflicts with.
+
+  Returns:
+    Revised version of the dictionary. All instructions conflict with
+    themselves. If A conflicts with B, B will conflict with A.
+  """
+  for key in conflicts:
+    for k in conflicts[key]:
+      conflicts[k].add(key)
+    conflicts[key].add(key)
+  return conflicts

helm/benchmark/metrics/ifeval/instructions_registry.pyi

@@ -0,0 +1,3 @@
+from typing import Dict, Any
+
+INSTRUCTION_DICT: Dict[str, Any]

helm/benchmark/metrics/ifeval/instructions_util.py

@@ -0,0 +1,153 @@
+# flake8: noqa
+# type: ignore
+# fmt: off
+# The following code was reproduced from the following URL:
+# https://github.com/google-research/google-research/blob/c7f60c013623e613732a096e2a0c2872491ec912/instruction_following_eval/instructions_util.py
+
+# coding=utf-8
+# Copyright 2023 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Utility library of instructions."""
+
+import functools
+import random
+import re
+from typing import List
+
+import immutabledict
+import nltk
+
+WORD_LIST = ["western", "sentence", "signal", "dump", "spot", "opposite", "bottom", "potato", "administration", "working", "welcome", "morning", "good", "agency", "primary", "wish", "responsibility", "press", "problem", "president", "steal", "brush", "read", "type", "beat", "trainer", "growth", "lock", "bone", "case", "equal", "comfortable", "region", "replacement", "performance", "mate", "walk", "medicine", "film", "thing", "rock", "tap", "total", "competition", "ease", "south", "establishment", "gather", "parking", "world", "plenty", "breath", "claim", "alcohol", "trade", "dear", "highlight", "street", "matter", "decision", "mess", "agreement", "studio", "coach", "assist", "brain", "wing", "style", "private", "top", "brown", "leg", "buy", "procedure", "method", "speed", "high", "company", "valuable", "pie", "analyst", "session", "pattern", "district", "pleasure", "dinner", "swimming", "joke", "order", "plate", "department", "motor", "cell", "spend", "cabinet", "difference", "power", "examination", "engine", "horse", "dimension", "pay", "toe", "curve", "literature", "bother", "fire", "possibility", "debate", "activity", "passage", "hello", "cycle", "background", "quiet", "author", "effect", "actor", "page", "bicycle", "error", "throat", "attack", "character", "phone", "tea", "increase", "outcome", "file", "specific", "inspector", "internal", "potential", "staff", "building", "employer", "shoe", "hand", "direction", "garden", "purchase", "interview", "study", "recognition", "member", "spiritual", "oven", "sandwich", "weird", "passenger", "particular", "response", "reaction", "size", "variation", "a", "cancel", "candy", "exit", "guest", "condition", "fly", "price", "weakness", "convert", "hotel", "great", "mouth", "mind", "song", "sugar", "suspect", "telephone", "ear", "roof", "paint", "refrigerator", "organization", "jury", "reward", "engineering", "day", "possession", "crew", "bar", "road", "description", "celebration", "score", "mark", "letter", "shower", "suggestion", "sir", "luck", "national", "progress", "hall", "stroke", "theory", "offer", "story", "tax", "definition", "history", "ride", "medium", "opening", "glass", "elevator", "stomach", "question", "ability", "leading", "village", "computer", "city", "grand", "confidence", "candle", "priest", "recommendation", "point", "necessary", "body", "desk", "secret", "horror", "noise", "culture", "warning", "water", "round", "diet", "flower", "bus", "tough", "permission", "week", "prompt", "connection", "abuse", "height", "save", "corner", "border", "stress", "drive", "stop", "rip", "meal", "listen", "confusion", "girlfriend", "living", "relation", "significance", "plan", "creative", "atmosphere", "blame", "invite", "housing", "paper", "drink", "roll", "silver", "drunk", "age", "damage", "smoke", "environment", "pack", "savings", "influence", "tourist", "rain", "post", "sign", "grandmother", "run", "profit", "push", "clerk", "final", "wine", "swim", "pause", "stuff", "singer", "funeral", "average", "source", "scene", "tradition", "personal", "snow", "nobody", "distance", "sort", "sensitive", "animal", "major", "negotiation", "click", "mood", "period", "arrival", "expression", "holiday", "repeat", "dust", "closet", "gold", "bad", "sail", "combination", "clothes", "emphasis", "duty", "black", "step", "school", "jump", "document", "professional", "lip", "chemical", "front", "wake", "while", "inside", "watch", "row", "subject", "penalty", "balance", "possible", "adult", "aside", "sample", "appeal", "wedding", "depth", "king", "award", "wife", "blow", "site", "camp", "music", "safe", "gift", "fault", "guess", "act", "shame", "drama", "capital", "exam", "stupid", "record", "sound", "swing", "novel", "minimum", "ratio", "machine", "shape", "lead", "operation", "salary", "cloud", "affair", "hit", "chapter", "stage", "quantity", "access", "army", "chain", "traffic", "kick", "analysis", "airport", "time", "vacation", "philosophy", "ball", "chest", "thanks", "place", "mountain", "advertising", "red", "past", "rent", "return", "tour", "house", "construction", "net", "native", "war", "figure", "fee", "spray", "user", "dirt", "shot", "task", "stick", "friend", "software", "promotion", "interaction", "surround", "block", "purpose", "practice", "conflict", "routine", "requirement", "bonus", "hole", "state", "junior", "sweet", "catch", "tear", "fold", "wall", "editor", "life", "position", "pound", "respect", "bathroom", "coat", "script", "job", "teach", "birth", "view", "resolve", "theme", "employee", "doubt", "market", "education", "serve", "recover", "tone", "harm", "miss", "union", "understanding", "cow", "river", "association", "concept", "training", "recipe", "relationship", "reserve", "depression", "proof", "hair", "revenue", "independent", "lift", "assignment", "temporary", "amount", "loss", "edge", "track", "check", "rope", "estimate", "pollution", "stable", "message", "delivery", "perspective", "mirror", "assistant", "representative", "witness", "nature", "judge", "fruit", "tip", "devil", "town", "emergency", "upper", "drop", "stay", "human", "neck", "speaker", "network", "sing", "resist", "league", "trip", "signature", "lawyer", "importance", "gas", "choice", "engineer", "success", "part", "external", "worker", "simple", "quarter", "student", "heart", "pass", "spite", "shift", "rough", "lady", "grass", "community", "garage", "youth", "standard", "skirt", "promise", "blind", "television", "disease", "commission", "positive", "energy", "calm", "presence", "tune", "basis", "preference", "head", "common", "cut", "somewhere", "presentation", "current", "thought", "revolution", "effort", "master", "implement", "republic", "floor", "principle", "stranger", "shoulder", "grade", "button", "tennis", "police", "collection", "account", "register", "glove", "divide", "professor", "chair", "priority", "combine", "peace", "extension", "maybe", "evening", "frame", "sister", "wave", "code", "application", "mouse", "match", "counter", "bottle", "half", "cheek", "resolution", "back", "knowledge", "make", "discussion", "screw", "length", "accident", "battle", "dress", "knee", "log", "package", "it", "turn", "hearing", "newspaper", "layer", "wealth", "profile", "imagination", "answer", "weekend", "teacher", "appearance", "meet", "bike", "rise", "belt", "crash", "bowl", "equivalent", "support", "image", "poem", "risk", "excitement", "remote", "secretary", "public", "produce", "plane", "display", "money", "sand", "situation", "punch", "customer", "title", "shake", "mortgage", "option", "number", "pop", "window", "extent", "nothing", "experience", "opinion", "departure", "dance", "indication", "boy", "material", "band", "leader", "sun", "beautiful", "muscle", "farmer", "variety", "fat", "handle", "director", "opportunity", "calendar", "outside", "pace", "bath", "fish", "consequence", "put", "owner", "go", "doctor", "information", "share", "hurt", "protection", "career", "finance", "force", "golf", "garbage", "aspect", "kid", "food", "boot", "milk", "respond", "objective", "reality", "raw", "ring", "mall", "one", "impact", "area", "news", "international", "series", "impress", "mother", "shelter", "strike", "loan", "month", "seat", "anything", "entertainment", "familiar", "clue", "year", "glad", "supermarket", "natural", "god", "cost", "conversation", "tie", "ruin", "comfort", "earth", "storm", "percentage", "assistance", "budget", "strength", "beginning", "sleep", "other", "young", "unit", "fill", "store", "desire", "hide", "value", "cup", "maintenance", "nurse", "function", "tower", "role", "class", "camera", "database", "panic", "nation", "basket", "ice", "art", "spirit", "chart", "exchange", "feedback", "statement", "reputation", "search", "hunt", "exercise", "nasty", "notice", "male", "yard", "annual", "collar", "date", "platform", "plant", "fortune", "passion", "friendship", "spread", "cancer", "ticket", "attitude", "island", "active", "object", "service", "buyer", "bite", "card", "face", "steak", "proposal", "patient", "heat", "rule", "resident", "broad", "politics", "west", "knife", "expert", "girl", "design", "salt", "baseball", "grab", "inspection", "cousin", "couple", "magazine", "cook", "dependent", "security", "chicken", "version", "currency", "ladder", "scheme", "kitchen", "employment", "local", "attention", "manager", "fact", "cover", "sad", "guard", "relative", "county", "rate", "lunch", "program", "initiative", "gear", "bridge", "breast", "talk", "dish", "guarantee", "beer", "vehicle", "reception", "woman", "substance", "copy", "lecture", "advantage", "park", "cold", "death", "mix", "hold", "scale", "tomorrow", "blood", "request", "green", "cookie", "church", "strip", "forever", "beyond", "debt", "tackle", "wash", "following", "feel", "maximum", "sector", "sea", "property", "economics", "menu", "bench", "try", "language", "start", "call", "solid", "address", "income", "foot", "senior", "honey", "few", "mixture", "cash", "grocery", "link", "map", "form", "factor", "pot", "model", "writer", "farm", "winter", "skill", "anywhere", "birthday", "policy", "release", "husband", "lab", "hurry", "mail", "equipment", "sink", "pair", "driver", "consideration", "leather", "skin", "blue", "boat", "sale", "brick", "two", "feed", "square", "dot", "rush", "dream", "location", "afternoon", "manufacturer", "control", "occasion", "trouble", "introduction", "advice", "bet", "eat", "kill", "category", "manner", "office", "estate", "pride", "awareness", "slip", "crack", "client", "nail", "shoot", "membership", "soft", "anybody", "web", "official", "individual", "pizza", "interest", "bag", "spell", "profession", "queen", "deal", "resource", "ship", "guy", "chocolate", "joint", "formal", "upstairs", "car", "resort", "abroad", "dealer", "associate", "finger", "surgery", "comment", "team", "detail", "crazy", "path", "tale", "initial", "arm", "radio", "demand", "single", "draw", "yellow", "contest", "piece", "quote", "pull", "commercial", "shirt", "contribution", "cream", "channel", "suit", "discipline", "instruction", "concert", "speech", "low", "effective", "hang", "scratch", "industry", "breakfast", "lay", "join", "metal", "bedroom", "minute", "product", "rest", "temperature", "many", "give", "argument", "print", "purple", "laugh", "health", "credit", "investment", "sell", "setting", "lesson", "egg", "middle", "marriage", "level", "evidence", "phrase", "love", "self", "benefit", "guidance", "affect", "you", "dad", "anxiety", "special", "boyfriend", "test", "blank", "payment", "soup", "obligation", "reply", "smile", "deep", "complaint", "addition", "review", "box", "towel", "minor", "fun", "soil", "issue", "cigarette", "internet", "gain", "tell", "entry", "spare", "incident", "family", "refuse", "branch", "can", "pen", "grandfather", "constant", "tank", "uncle", "climate", "ground", "volume", "communication", "kind", "poet", "child", "screen", "mine", "quit", "gene", "lack", "charity", "memory", "tooth", "fear", "mention", "marketing", "reveal", "reason", "court", "season", "freedom", "land", "sport", "audience", "classroom", "law", "hook", "win", "carry", "eye", "smell", "distribution", "research", "country", "dare", "hope", "whereas", "stretch", "library", "if", "delay", "college", "plastic", "book", "present", "use", "worry", "champion", "goal", "economy", "march", "election", "reflection", "midnight", "slide", "inflation", "action", "challenge", "guitar", "coast", "apple", "campaign", "field", "jacket", "sense", "way", "visual", "remove", "weather", "trash", "cable", "regret", "buddy", "beach", "historian", "courage", "sympathy", "truck", "tension", "permit", "nose", "bed", "son", "person", "base", "meat", "usual", "air", "meeting", "worth", "game", "independence", "physical", "brief", "play", "raise", "board", "she", "key", "writing", "pick", "command", "party", "yesterday", "spring", "candidate", "physics", "university", "concern", "development", "change", "string", "target", "instance", "room", "bitter", "bird", "football", "normal", "split", "impression", "wood", "long", "meaning", "stock", "cap", "leadership", "media", "ambition", "fishing", "essay", "salad", "repair", "today", "designer", "night", "bank", "drawing", "inevitable", "phase", "vast", "chip", "anger", "switch", "cry", "twist", "personality", "attempt", "storage", "being", "preparation", "bat", "selection", "white", "technology", "contract", "side", "section", "station", "till", "structure", "tongue", "taste", "truth", "difficulty", "group", "limit", "main", "move", "feeling", "light", "example", "mission", "might", "wait", "wheel", "shop", "host", "classic", "alternative", "cause", "agent", "consist", "table", "airline", "text", "pool", "craft", "range", "fuel", "tool", "partner", "load", "entrance", "deposit", "hate", "article", "video", "summer", "feature", "extreme", "mobile", "hospital", "flight", "fall", "pension", "piano", "fail", "result", "rub", "gap", "system", "report", "suck", "ordinary", "wind", "nerve", "ask", "shine", "note", "line", "mom", "perception", "brother", "reference", "bend", "charge", "treat", "trick", "term", "homework", "bake", "bid", "status", "project", "strategy", "orange", "let", "enthusiasm", "parent", "concentrate", "device", "travel", "poetry", "business", "society", "kiss", "end", "vegetable", "employ", "schedule", "hour", "brave", "focus", "process", "movie", "illegal", "general", "coffee", "ad", "highway", "chemistry", "psychology", "hire", "bell", "conference", "relief", "show", "neat", "funny", "weight", "quality", "club", "daughter", "zone", "touch", "tonight", "shock", "burn", "excuse", "name", "survey", "landscape", "advance", "satisfaction", "bread", "disaster", "item", "hat", "prior", "shopping", "visit", "east", "photo", "home", "idea", "father", "comparison", "cat", "pipe", "winner", "count", "lake", "fight", "prize", "foundation", "dog", "keep", "ideal", "fan", "struggle", "peak", "safety", "solution", "hell", "conclusion", "population", "strain", "alarm", "measurement", "second", "train", "race", "due", "insurance", "boss", "tree", "monitor", "sick", "course", "drag", "appointment", "slice", "still", "care", "patience", "rich", "escape", "emotion", "royal", "female", "childhood", "government", "picture", "will", "sock", "big", "gate", "oil", "cross", "pin", "improvement", "championship", "silly", "help", "sky", "pitch", "man", "diamond", "most", "transition", "work", "science", "committee", "moment", "fix", "teaching", "dig", "specialist", "complex", "guide", "people", "dead", "voice", "original", "break", "topic", "data", "degree", "reading", "recording", "bunch", "reach", "judgment", "lie", "regular", "set", "painting", "mode", "list", "player", "bear", "north", "wonder", "carpet", "heavy", "officer", "negative", "clock", "unique", "baby", "pain", "assumption", "disk", "iron", "bill", "drawer", "look", "double", "mistake", "finish", "future", "brilliant", "contact", "math", "rice", "leave", "restaurant", "discount", "sex", "virus", "bit", "trust", "event", "wear", "juice", "failure", "bug", "context", "mud", "whole", "wrap", "intention", "draft", "pressure", "cake", "dark", "explanation", "space", "angle", "word", "efficiency", "management", "habit", "star", "chance", "finding", "transportation", "stand", "criticism", "flow", "door", "injury", "insect", "surprise", "apartment"]  # pylint: disable=line-too-long
+
+# ISO 639-1 codes to language names.
+LANGUAGE_CODES = immutabledict.immutabledict({
+    "en": "English",
+    "es": "Spanish",
+    "pt": "Portuguese",
+    "ar": "Arabic",
+    "hi": "Hindi",
+    "fr": "French",
+    "ru": "Russian",
+    "de": "German",
+    "ja": "Japanese",
+    "it": "Italian",
+    "bn": "Bengali",
+    "uk": "Ukrainian",
+    "th": "Thai",
+    "ur": "Urdu",
+    "ta": "Tamil",
+    "te": "Telugu",
+    "bg": "Bulgarian",
+    "ko": "Korean",
+    "pl": "Polish",
+    "he": "Hebrew",
+    "fa": "Persian",
+    "vi": "Vietnamese",
+    "ne": "Nepali",
+    "sw": "Swahili",
+    "kn": "Kannada",
+    "mr": "Marathi",
+    "gu": "Gujarati",
+    "pa": "Punjabi",
+    "ml": "Malayalam",
+    "fi": "Finnish",
+    })
+
+_ALPHABETS = "([A-Za-z])"
+_PREFIXES = "(Mr|St|Mrs|Ms|Dr)[.]"
+_SUFFIXES = "(Inc|Ltd|Jr|Sr|Co)"
+_STARTERS = r"(Mr|Mrs|Ms|Dr|Prof|Capt|Cpt|Lt|He\s|She\s|It\s|They\s|Their\s|Our\s|We\s|But\s|However\s|That\s|This\s|Wherever)"
+_ACRONYMS = "([A-Z][.][A-Z][.](?:[A-Z][.])?)"
+_WEBSITES = "[.](com|net|org|io|gov|edu|me)"
+_DIGITS = "([0-9])"
+_MULTIPLE_DOTS = r"\.{2,}"
+
+
+def split_into_sentences(text):
+  """Split the text into sentences.
+
+  Args:
+    text: A string that consists of more than or equal to one sentences.
+
+  Returns:
+    A list of strings where each string is a sentence.
+  """
+  text = " " + text + "  "
+  text = text.replace("\n", " ")
+  text = re.sub(_PREFIXES, "\\1<prd>", text)
+  text = re.sub(_WEBSITES, "<prd>\\1", text)
+  text = re.sub(_DIGITS + "[.]" + _DIGITS, "\\1<prd>\\2", text)
+  text = re.sub(
+      _MULTIPLE_DOTS,
+      lambda match: "<prd>" * len(match.group(0)) + "<stop>",
+      text,
+  )
+  if "Ph.D" in text:
+    text = text.replace("Ph.D.", "Ph<prd>D<prd>")
+  text = re.sub(r"\s" + _ALPHABETS + "[.] ", " \\1<prd> ", text)
+  text = re.sub(_ACRONYMS + " " + _STARTERS, "\\1<stop> \\2", text)
+  text = re.sub(
+      _ALPHABETS + "[.]" + _ALPHABETS + "[.]" + _ALPHABETS + "[.]",
+      "\\1<prd>\\2<prd>\\3<prd>",
+      text,
+  )
+  text = re.sub(
+      _ALPHABETS + "[.]" + _ALPHABETS + "[.]", "\\1<prd>\\2<prd>", text
+  )
+  text = re.sub(" " + _SUFFIXES + "[.] " + _STARTERS, " \\1<stop> \\2", text)
+  text = re.sub(" " + _SUFFIXES + "[.]", " \\1<prd>", text)
+  text = re.sub(" " + _ALPHABETS + "[.]", " \\1<prd>", text)
+  if "”" in text:
+    text = text.replace(".”", "”.")
+  if '"' in text:
+    text = text.replace('."', '".')
+  if "!" in text:
+    text = text.replace('!"', '"!')
+  if "?" in text:
+    text = text.replace('?"', '"?')
+  text = text.replace(".", ".<stop>")
+  text = text.replace("?", "?<stop>")
+  text = text.replace("!", "!<stop>")
+  text = text.replace("<prd>", ".")
+  sentences = text.split("<stop>")
+  sentences = [s.strip() for s in sentences]
+  if sentences and not sentences[-1]:
+    sentences = sentences[:-1]
+  return sentences
+
+
+def count_words(text):
+  """Counts the number of words."""
+  tokenizer = nltk.tokenize.RegexpTokenizer(r"\w+")
+  tokens = tokenizer.tokenize(text)
+  num_words = len(tokens)
+  return num_words
+
+
+@functools.lru_cache(maxsize=None)
+def _get_sentence_tokenizer():
+  return nltk.data.load("nltk:tokenizers/punkt/english.pickle")
+
+
+def count_sentences(text):
+  """Count the number of sentences."""
+  tokenizer = _get_sentence_tokenizer()
+  tokenized_sentences = tokenizer.tokenize(text)
+  return len(tokenized_sentences)
+
+
+def generate_keywords(num_keywords):
+  """Randomly generates a few keywords."""
+  return random.sample(WORD_LIST, k=num_keywords)

helm/benchmark/metrics/ifeval_metrics.py

@@ -0,0 +1,55 @@
+from typing import List
+
+from helm.common.hierarchical_logger import hlog
+from helm.benchmark.adaptation.adapter_spec import AdapterSpec
+from helm.benchmark.adaptation.request_state import RequestState
+from helm.benchmark.metrics.metric import Metric
+from helm.benchmark.metrics.metric_name import MetricName
+from helm.benchmark.metrics.metric_service import MetricService
+from helm.benchmark.metrics.statistic import Stat
+
+from helm.benchmark.metrics.ifeval.instructions_registry import INSTRUCTION_DICT
+
+
+class IFEvalMetric(Metric):
+    def evaluate_generation(
+        self,
+        adapter_spec: AdapterSpec,
+        request_state: RequestState,
+        metric_service: MetricService,
+        eval_cache_path: str,
+    ) -> List[Stat]:
+        prompt = request_state.instance.input.text
+        assert request_state.instance.extra_data
+        instruction_ids = request_state.instance.extra_data["instruction_ids"]
+        instruction_kwargs = request_state.instance.extra_data["instruction_kwargs"]
+        assert len(instruction_ids) > 0
+        assert request_state.result
+        assert len(request_state.result.completions) == 1, f"Got {len(request_state.result.completions)} completions"
+        response = request_state.result.completions[0].text.strip()
+
+        # The following logic was reproduced with minor modifications from the following URL:
+        # https://github.com/google-research/google-research/blob/c7f60c013623e613732a096e2a0c2872491ec912/
+        # instruction_following_eval/evaluation_main.py#L96-L125
+        is_following_list = []
+        for index, instruction_id in enumerate(instruction_ids):
+            instruction_cls = INSTRUCTION_DICT[instruction_id]
+            instruction = instruction_cls(instruction_id)
+
+            instruction.build_description(**{k: v for k, v in instruction_kwargs[index].items() if v is not None})
+            args = instruction.get_instruction_args()
+            if args and "prompt" in args:
+                instruction.build_description(prompt=prompt)
+
+            is_following = False
+            if response.strip():
+                try:
+                    is_following = instruction.check_following(response)
+                except Exception as e:
+                    hlog(f"WARNING: Instruction following checking failed with error message {e}")
+            if is_following:
+                is_following_list.append(1)
+            else:
+                is_following_list.append(0)
+
+        return [Stat(MetricName("ifeval_strict_accuracy")).add(sum(is_following_list) / len(is_following_list))]

helm/benchmark/metrics/image_generation/aesthetics_metrics.py

@@ -9,7 +9,7 @@ from helm.benchmark.metrics.statistic import Stat
 from helm.benchmark.metrics.metric import Metric
 from helm.benchmark.metrics.metric_name import MetricName
 from helm.benchmark.metrics.metric_service import MetricService
-from .aesthetics_scorer import AestheticsScorer
+from helm.benchmark.metrics.image_generation.aesthetics_scorer import AestheticsScorer
 from helm.common.multimodal_request_utils import gather_generated_image_locations
 
 

helm/benchmark/metrics/image_generation/detection_metrics.py

@@ -10,7 +10,7 @@ from helm.benchmark.metrics.metric import Metric
 from helm.benchmark.metrics.metric_name import MetricName
 from helm.benchmark.metrics.metric_service import MetricService
 from helm.common.multimodal_request_utils import gather_generated_image_locations
-from .detectors.vitdet import ViTDetDetector
+from helm.benchmark.metrics.image_generation.detectors.vitdet import ViTDetDetector
 
 
 class DetectionMetric(Metric):

helm/benchmark/metrics/image_generation/detectors/vitdet.py

@@ -8,7 +8,7 @@ from helm.common.general import ensure_file_downloaded, hlog
 from helm.common.images_utils import open_image
 from helm.common.optional_dependencies import handle_module_not_found_error
 from helm.common.gpu_utils import get_torch_device
-from .base_detector import BaseDetector
+from helm.benchmark.metrics.image_generation.detectors.base_detector import BaseDetector
 
 
 MODEL_CONFIG_DOWNLOAD_URL: str = "https://drive.google.com/uc?id=1MLuwQ0ZN0gJQ42oVCc0aFz6Rneb1g3Rt"

helm/benchmark/metrics/image_generation/fractal_dimension/test_fractal_dimension_util.py

@@ -1,6 +1,6 @@
 import os
 
-from .fractal_dimension_util import compute_fractal_dimension
+from helm.benchmark.metrics.image_generation.fractal_dimension.fractal_dimension_util import compute_fractal_dimension
 
 
 def fractal_dimension_test(image_filename: str, expected_fractal_dimension: float):

helm/benchmark/metrics/image_generation/fractal_dimension_metric.py

@@ -10,7 +10,7 @@ from helm.benchmark.metrics.metric import Metric
 from helm.benchmark.metrics.metric_name import MetricName
 from helm.benchmark.metrics.metric_service import MetricService
 from helm.common.multimodal_request_utils import gather_generated_image_locations
-from .fractal_dimension.fractal_dimension_util import compute_fractal_dimension
+from helm.benchmark.metrics.image_generation.fractal_dimension.fractal_dimension_util import compute_fractal_dimension
 
 
 class FractalDimensionMetric(Metric):

helm/benchmark/metrics/image_generation/nsfw_metrics.py

@@ -10,7 +10,7 @@ from helm.benchmark.metrics.metric_service import MetricService
 from helm.benchmark.metrics.toxicity_utils import is_prompt_toxic
 from helm.clients.image_generation.dalle2_client import DALLE2Client
 from helm.common.images_utils import is_blacked_out_image
-from .nsfw_detector import NSFWDetector
+from helm.benchmark.metrics.image_generation.nsfw_detector import NSFWDetector
 
 
 class NSFWMetric(Metric):

helm/benchmark/metrics/image_generation/q16/test_q16.py

@@ -1,8 +1,10 @@
 import os
+import pytest
 
-from .q16_toxicity_detector import Q16ToxicityDetector
+from helm.benchmark.metrics.image_generation.q16.q16_toxicity_detector import Q16ToxicityDetector
 
 
+@pytest.mark.skip(reason="Skipping due to flakiness.")
 class TestQ16:
     def setup_method(self, method):
         self._q16_detector = Q16ToxicityDetector()

helm/benchmark/metrics/image_generation/q16_toxicity_metrics.py

@@ -9,7 +9,7 @@ from helm.benchmark.metrics.metric import Metric
 from helm.benchmark.metrics.metric_name import MetricName
 from helm.benchmark.metrics.metric_service import MetricService
 from helm.benchmark.metrics.toxicity_utils import is_prompt_toxic
-from .q16.q16_toxicity_detector import Q16ToxicityDetector
+from helm.benchmark.metrics.image_generation.q16.q16_toxicity_detector import Q16ToxicityDetector
 from helm.common.multimodal_request_utils import gather_generated_image_locations
 
 

helm/benchmark/metrics/image_generation/skin_tone_metrics.py

@@ -92,9 +92,9 @@ class SkinToneMetric(Metric):
                     and (Cr <= ((-2.2857 * Cb) + 432.85))
                 ):
 
-                    blue.append(img_rgba[i, j].item(0))
+                    blue.append(img_rgba[i, j].item(2))
                     green.append(img_rgba[i, j].item(1))
-                    red.append(img_rgba[i, j].item(2))
+                    red.append(img_rgba[i, j].item(0))
                 else:
                     img_rgba[i, j] = [0, 0, 0, 0]
 

helm/benchmark/metrics/image_generation/watermark/test_watermark_detector.py

@@ -1,7 +1,7 @@
 from typing import List
 import os
 
-from .watermark_detector import WatermarkDetector
+from helm.benchmark.metrics.image_generation.watermark.watermark_detector import WatermarkDetector
 
 
 def test_compute_watermark_probability():

helm/benchmark/metrics/image_generation/watermark_metrics.py

@@ -9,7 +9,7 @@ from helm.benchmark.metrics.metric import Metric
 from helm.benchmark.metrics.metric_name import MetricName
 from helm.benchmark.metrics.metric_service import MetricService
 from helm.common.multimodal_request_utils import gather_generated_image_locations
-from .watermark.watermark_detector import WatermarkDetector
+from helm.benchmark.metrics.image_generation.watermark.watermark_detector import WatermarkDetector
 
 
 class WatermarkMetric(Metric):

helm/benchmark/metrics/instruction_following_critique_metrics.py

@@ -3,10 +3,10 @@ from typing import Dict, List
 
 from helm.benchmark.adaptation.request_state import RequestState
 from helm.benchmark.adaptation.adapter_spec import AdapterSpec
-from .metric import Metric
-from .metric_name import MetricName
-from .metric_service import MetricService
-from .statistic import Stat
+from helm.benchmark.metrics.metric import Metric
+from helm.benchmark.metrics.metric_name import MetricName
+from helm.benchmark.metrics.metric_service import MetricService
+from helm.benchmark.metrics.statistic import Stat
 from helm.common.critique_request import CritiqueTaskTemplate, CritiqueQuestionTemplate, CritiqueRequest, QuestionType
 
 

helm/benchmark/metrics/language_modeling_metrics.py

@@ -11,10 +11,10 @@ from helm.benchmark.metrics.efficiency_metrics import EfficiencyMetric
 
 from helm.benchmark.adaptation.request_state import RequestState
 from helm.benchmark.adaptation.adapter_spec import AdapterSpec
-from .metric import MetricInterface, MetricResult, PerInstanceStats, add_context
-from .metric_name import MetricContext, MetricName
-from .metric_service import MetricService
-from .statistic import Stat, merge_stat
+from helm.benchmark.metrics.metric import MetricInterface, MetricResult, PerInstanceStats, add_context
+from helm.benchmark.metrics.metric_name import MetricContext, MetricName
+from helm.benchmark.metrics.metric_service import MetricService
+from helm.benchmark.metrics.statistic import Stat, merge_stat
 
 
 class LanguageModelingMetric(MetricInterface):

helm/benchmark/metrics/machine_translation_metrics.py

@@ -3,8 +3,8 @@ from typing import List
 from helm.benchmark.adaptation.request_state import RequestState
 from helm.benchmark.metrics.evaluate_instances_metric import EvaluateInstancesMetric
 from helm.common.optional_dependencies import handle_module_not_found_error
-from .metric_name import MetricName
-from .statistic import Stat
+from helm.benchmark.metrics.metric_name import MetricName
+from helm.benchmark.metrics.statistic import Stat
 
 try:
     from sacrebleu.metrics import BLEU