import os import json import random import glob import shutil import tempfile from urllib.error import URLError from urllib.request import urlopen import numpy as np import pandas as pd from sklearn.model_selection import train_test_split from datasets import load_dataset, Dataset, concatenate_datasets import torch from src.metrics import rouge_L from src.utils import load_model_and_validate_gpu from tqdm import tqdm import ast from typing import Any, Dict, List, Tuple DATA2HF={ "coqa":"stanfordnlp/coqa", "triviaqa":None, "hotpotqa":None, "nq": None, "math":None, "squad":"rajpurkar/squad", "hotpotqa_c":None, 'psiloqa':"s-nlp/PsiloQA", "halueval_summary":None, "cnn_dailymail":"abisee/cnn_dailymail", } TRUETEACHER_MODEL_ID = "google/t5_11b_trueteacher_and_anli" TRUETEACHER_MAX_LENGTH = 2048 TRUETEACHER_DEFAULT_BATCH_SIZE = 4 TRUETEACHER_A6000_BATCH_SIZE = 2 HALUEVAL_SUMMARY_URL = ( "https://raw.githubusercontent.com/RUCAIBox/HaluEval/main/data/summarization_data.json" ) HALUEVAL_SUMMARY_SUBDIRS = ["halueval", "HaluEval"] HALUEVAL_SUMMARY_FILENAME = "summarization_data.json" def _load_json_records(path: str) -> Any: def _load_as_jsonl() -> List[Dict[str, Any]]: records = [] with open(path, "r", encoding="utf-8") as f: for line in f: line = line.strip() if line: records.append(json.loads(line)) return records if path.endswith(".jsonl"): return _load_as_jsonl() try: with open(path, "r", encoding="utf-8") as f: return json.load(f) except json.JSONDecodeError as exc: # Some local files use a .json suffix but are actually JSONL. try: return _load_as_jsonl() except json.JSONDecodeError: raise exc def _pick_first_nonempty(example: Dict[str, Any], keys: List[str]) -> Any: for key in keys: value = example.get(key) if value is None: continue if isinstance(value, str) and value.strip() == "": continue if isinstance(value, list) and len(value) == 0: continue return value return None def _normalize_summary_text(value: Any) -> str: if value is None: return "" if isinstance(value, str): return value.strip() if isinstance(value, list): for item in value: text = _normalize_summary_text(item) if text: return text return "" if isinstance(value, dict): return _normalize_summary_text( _pick_first_nonempty( value, [ "text", "summary", "reference", "reference_summary", "gold_summary", "right_summary", "faithful_summary", ], ) ) return str(value).strip() def _normalize_halueval_summary_record(example: Dict[str, Any]) -> Dict[str, Any]: context = _normalize_summary_text( _pick_first_nonempty( example, ["document", "source", "article", "context", "passage", "text", "input"], ) ) reference_summary = _normalize_summary_text( _pick_first_nonempty( example, [ "reference_summary", "gold_summary", "summary", "reference", "right_summary", "faithful_summary", ], ) ) if not context: raise ValueError(f"Cannot find source document field in example keys: {list(example.keys())}") normalized = { "context": context, "answers": [reference_summary] if reference_summary else [""], } provided_label = _pick_first_nonempty( example, ["label", "hallucination_label", "faithfulness_label", "provided_label"], ) if provided_label is not None: normalized["provided_label"] = provided_label hallucinated_summary = _normalize_summary_text( _pick_first_nonempty(example, ["hallucinated_summary", "hallucinated", "fake_summary"]) ) if hallucinated_summary: normalized["hallucinated_summary"] = hallucinated_summary return normalized def _find_halueval_summary_file(data_cache_dir: str, split: str) -> Tuple[List[str], bool]: split_patterns = [ os.path.join(data_cache_dir, "halueval", f"*summary*{split}*.json*"), os.path.join(data_cache_dir, "HaluEval", f"*summary*{split}*.json*"), os.path.join(data_cache_dir, "halueval", f"*summarization*{split}*.json*"), os.path.join(data_cache_dir, "HaluEval", f"*summarization*{split}*.json*"), ] fallback_patterns = [ os.path.join(data_cache_dir, "halueval", "*summary*.json*"), os.path.join(data_cache_dir, "HaluEval", "*summary*.json*"), os.path.join(data_cache_dir, "halueval", "*summarization*.json*"), os.path.join(data_cache_dir, "HaluEval", "*summarization*.json*"), ] matches = [] for pattern in split_patterns: matches.extend(sorted(glob.glob(pattern))) used_explicit_split_file = len(matches) > 0 if not matches: for pattern in fallback_patterns: matches.extend(sorted(glob.glob(pattern))) return matches, used_explicit_split_file def _download_halueval_summary(data_cache_dir: str) -> str: target_dir = os.path.join(data_cache_dir, HALUEVAL_SUMMARY_SUBDIRS[0]) target_path = os.path.join(target_dir, HALUEVAL_SUMMARY_FILENAME) os.makedirs(target_dir, exist_ok=True) tmp_path = None try: with urlopen(HALUEVAL_SUMMARY_URL) as response: with tempfile.NamedTemporaryFile(delete=False, dir=target_dir) as tmp_file: tmp_path = tmp_file.name shutil.copyfileobj(response, tmp_file) os.replace(tmp_path, target_path) return target_path except Exception as exc: if tmp_path is not None and os.path.exists(tmp_path): os.unlink(tmp_path) if isinstance(exc, URLError): raise FileNotFoundError( "Failed to download HaluEval summarization data from official GitHub. " f"URL: {HALUEVAL_SUMMARY_URL}. " f"Intended local path: {target_path}. " f"Original error: {exc}" ) from exc raise def load_halueval_summary(args) -> Dataset: matches, used_explicit_split_file = _find_halueval_summary_file( args.data_cache_dir, args.split, ) if not matches: print("HaluEval summarization dataset not found locally. Downloading from official GitHub...") downloaded_path = _download_halueval_summary(args.data_cache_dir) print(f"Downloaded HaluEval summarization dataset to: {downloaded_path}") matches, used_explicit_split_file = _find_halueval_summary_file( args.data_cache_dir, args.split, ) if not matches: searched = [ os.path.join(args.data_cache_dir, subdir, f"*summary*{args.split}*.json*") for subdir in HALUEVAL_SUMMARY_SUBDIRS ] + [ os.path.join(args.data_cache_dir, subdir, f"*summarization*{args.split}*.json*") for subdir in HALUEVAL_SUMMARY_SUBDIRS ] + [ os.path.join(args.data_cache_dir, subdir, "*summary*.json*") for subdir in HALUEVAL_SUMMARY_SUBDIRS ] + [ os.path.join(args.data_cache_dir, subdir, "*summarization*.json*") for subdir in HALUEVAL_SUMMARY_SUBDIRS ] raise FileNotFoundError( "HaluEval summarization data is still not discoverable after auto-download. " f"Searched patterns:\n{chr(10).join(searched)}" ) raw = _load_json_records(matches[0]) if isinstance(raw, dict): if args.split in raw and isinstance(raw[args.split], list): records = raw[args.split] used_explicit_split_file = True elif "data" in raw and isinstance(raw["data"], list): records = raw["data"] elif "examples" in raw and isinstance(raw["examples"], list): records = raw["examples"] elif "records" in raw and isinstance(raw["records"], list): records = raw["records"] else: raise ValueError(f"Unsupported HaluEval summary file format: {matches[0]}") elif isinstance(raw, list): records = raw else: raise ValueError(f"Unsupported HaluEval summary file format: {matches[0]}") # total size is 10000 if not used_explicit_split_file: train_records, test_records = train_test_split( records, # records[:100], test_size=0.2, random_state=getattr(args, "seed", 42), ) records = train_records if args.split == "train" else test_records normalized_records = [_normalize_halueval_summary_record(record) for record in records] dataset_dict = { "context": [record["context"] for record in normalized_records], "answers": [record["answers"] for record in normalized_records], } if any("provided_label" in record for record in normalized_records): dataset_dict["provided_label"] = [ record.get("provided_label") for record in normalized_records ] if any("hallucinated_summary" in record for record in normalized_records): dataset_dict["hallucinated_summary"] = [ record.get("hallucinated_summary", "") for record in normalized_records ] return Dataset.from_dict(dataset_dict) def load_cnn_dailymail(args) -> Dataset: local_dir = os.path.join(args.data_cache_dir, "cnn_dailymail") train_file = os.path.join(local_dir, "train.jsonl") test_file = os.path.join(local_dir, "test.jsonl") if not (os.path.exists(train_file) and os.path.exists(test_file)): os.makedirs(local_dir, exist_ok=True) ds = load_dataset(DATA2HF["cnn_dailymail"], "3.0.0", cache_dir=args.data_cache_dir) if len(ds["train"]) < 10000: raise ValueError( f"CNN/DailyMail train split has only {len(ds['train'])} samples; need at least 10000." ) if len(ds["test"]) < 10000: raise ValueError( f"CNN/DailyMail test split has only {len(ds['test'])} samples; need at least 10000." ) train_ds = ds["train"].shuffle(seed=getattr(args, "seed", 42)).select(range(10000)) test_ds = ds["test"].shuffle(seed=getattr(args, "seed", 42)).select(range(10000)) train_ds.to_json(train_file, orient="records", lines=True) test_ds.to_json(test_file, orient="records", lines=True) print(f"CNN/DailyMail train set size: {len(train_ds)}") print(f"CNN/DailyMail test set size: {len(test_ds)}") file_path = train_file if args.split == "train" else test_file ds = load_dataset("json", data_files=file_path, cache_dir=args.data_cache_dir)["train"] return Dataset.from_dict( { "context": ds["article"], "answers": [[summary] for summary in ds["highlights"]], } ) def load_triviaqa(args, legacy=False) -> Dataset: test=True if args.split=='test' else False if legacy: with open('../data/verified-web-dev.json') as f: data_verified = json.load(f)['Data'] with open('../data/web-dev.json') as f: data = json.load(f)['Data'] questions_from_verified = {x['Question'] for x in data_verified} data_not_verified = [ x for x in data if x['Question'] not in questions_from_verified ] print("Length of not verified data: ", len(data_not_verified)) print("Length of verified data: ", len(data_verified)) if test: selected = data_verified else: selected = data_not_verified questions = [ex['Question'] for ex in selected] aliases = [ex['Answer']['Aliases'] for ex in selected] dataset = Dataset.from_dict({ "question": questions, "answer_aliases": aliases, }) return dataset else: if test: file_path = os.path.join( args.data_cache_dir, "triviaqa-unfiltered", "unfiltered-web-dev.json", ) else: file_path = os.path.join( args.data_cache_dir, "triviaqa-unfiltered", "unfiltered-web-train.json", ) with open(file_path) as f: data = json.load(f)['Data'] data, _ = train_test_split(data, train_size=10000, random_state=42) questions = [ex['Question'] for ex in data] aliases = [ex['Answer']['Aliases'] for ex in data] dataset = Dataset.from_dict({ "question": questions, "answers": aliases, }) return dataset def load_psiloqa(args) -> Dataset: """ Load PsiloQA directly from Hugging Face Hub and convert to the field schema used by this project. """ local_dir = os.path.join(args.data_cache_dir, "psiloqa") train_file = os.path.join(local_dir, "train.jsonl") test_file = os.path.join(local_dir, "test.jsonl") if not (os.path.exists(train_file) and os.path.exists(test_file)): os.makedirs(local_dir, exist_ok=True) ds = load_dataset(DATA2HF["psiloqa"], cache_dir=args.data_cache_dir) dataset_split = {} for split in ds.keys(): split_ds = ds[split] if "lang" in split_ds.column_names: split_ds = split_ds.filter(lambda example: example["lang"] == "en") dataset_split[split] = split_ds if len(dataset_split["train"]) < 10000: raise ValueError( f"PsiloQA English train split has only {len(dataset_split['train'])} samples; need at least 10000." ) shuffled_train = dataset_split["train"].shuffle(seed=42) train_ds = shuffled_train.select(range(10000)) rest_ds = shuffled_train.select(range(10000, len(shuffled_train))) combined_test_ds = concatenate_datasets( [rest_ds, dataset_split["validation"], dataset_split["test"]] ) combined_test_ds = combined_test_ds.shuffle(seed=42) train_ds.to_json(train_file, orient="records", lines=True) combined_test_ds.to_json(test_file, orient="records", lines=True) print(f"PsiloQA train set size: {len(train_ds)}") print(f"PsiloQA test set size: {len(combined_test_ds)}") if args.split == 'train': return Dataset.from_dict( { "context": train_ds["wiki_passage"], "question": train_ds["question"], "answers": [[answer] for answer in train_ds["golden_answer"]], } ) else: return Dataset.from_dict( { "context": combined_test_ds["wiki_passage"], "question": combined_test_ds["question"], "answers": [[answer] for answer in combined_test_ds["golden_answer"]], } ) else: # directly load local dataset file_path = os.path.join(local_dir, f"{args.split}.jsonl") ds = load_dataset("json", data_files=file_path, cache_dir=args.data_cache_dir)["train"] return Dataset.from_dict( { "context": ds["wiki_passage"], "question": ds["question"], "answers": [[answer] for answer in ds["golden_answer"]], } ) def load_hotpotqa(args,context=False): if args.split=='train': file_path = os.path.join( args.data_cache_dir, "hotpotqa", "train", "hotpot_train_v1.1.json", ) else: file_path = os.path.join( args.data_cache_dir, "hotpotqa", "distractor_validation", "hotpot_dev_distractor_v1.json", ) with open(file_path) as f: data = json.load(f) if args.split=='train': data, _ = train_test_split(data, train_size=10000, random_state=42) all_questions = [item['question'] for item in data] labels = [ [item['answer']] for item in data] if context: contexts = [] for item in data: title=item['supporting_facts'][0][0] title_list=[t[0] for t in item['context']] title_index=title_list.index(title) sentences_list=item['context'][title_index][1] context="".join(sentences_list) contexts.append(context) dataset= Dataset.from_dict({ "context": contexts, "question": all_questions, "answers": labels, }) return dataset dataset= Dataset.from_dict({ "question": all_questions, "answers": labels, }) return dataset def load_coqa(args): data = load_dataset( DATA2HF[args.dataset], split="train" if args.split=="train" else "validation", cache_dir=args.data_cache_dir, ) question_num=0 dataset = {} dataset['story'] = [] dataset['question'] = [] dataset['answers'] = [] for sample_id, sample in enumerate(data): story = sample['story'] questions = sample['questions'] answers = sample['answers']['input_text'] for question,answer in zip(questions,answers): dataset['story'].append(story) dataset['question'].append(question) dataset['answers'].append([answer]) if args.split=='train': question_num+=1 if question_num>=10000: return Dataset.from_dict({ 'context':dataset['story'], "question": dataset['question'], "answers": dataset['answers'], }) return Dataset.from_dict({ 'context':dataset['story'], "question": dataset['question'], "answers": dataset['answers'], }) def load_squad(args): """ Manually load a local SQuAD v1.1 JSON and return: Dataset({ 'question': List[str], 'answers': List[List[str]], # multiple answers per sample }) Behavior: - train: shuffle and take up to 10000 samples - non-train: same random 10000 """ data_dir = os.path.join(args.data_cache_dir, "squad") # assumes the downloaded file is SQuAD v1.1 if args.split == "train": json_path = os.path.join(data_dir, "train-v1.1.json") else: # dev serves as validation/test json_path = os.path.join(data_dir, "dev-v1.1.json") if not os.path.exists(json_path): raise FileNotFoundError(f"SQuAD json not found: {json_path}") with open(json_path, "r", encoding="utf-8") as f: raw = json.load(f) contexts = [] questions = [] answers = [] # List[List[str]] # raw structure: data -> [article] -> paragraphs -> context + qas for article in raw["data"]: for para in article["paragraphs"]: ctx = para["context"] for qa in para["qas"]: q_text = qa["question"] ans_texts = [a["text"] for a in qa.get("answers", [])] if not ans_texts: ans_texts = [""] # fallback to avoid empty list contexts.append(ctx) questions.append(q_text) answers.append(ans_texts) n = len(questions) indices = list(range(n)) rng = random.Random(getattr(args, "seed", 42)) rng.shuffle(indices) indices = indices[: min(10000, n)] # build the final Dataset from the selected index subset sel_questions = [questions[i] for i in indices] sel_answers = [answers[i] for i in indices] sel_contexts=[contexts[i] for i in indices] return Dataset.from_dict( { "context":sel_contexts, "question": sel_questions, "answers": sel_answers, } ) def uses_completion_prompt(model_name: str) -> bool: return model_name == "llama3.1-8b" def build_prompt(args,dataset): if uses_completion_prompt(args.model): if args.dataset=='triviaqa': return triviaqa_completion_prompt(dataset) elif args.dataset=='hotpotqa': return triviaqa_completion_prompt(dataset) elif args.dataset=='coqa': return coqa_completion_prompt(dataset) elif args.dataset=='nq': return triviaqa_completion_prompt(dataset) elif args.dataset=='squad': return coqa_completion_prompt(dataset) elif args.dataset=='math': return triviaqa_completion_prompt(dataset) elif args.dataset=='hotpotqa_c': return coqa_completion_prompt(dataset) elif args.dataset=='psiloqa': return coqa_completion_prompt(dataset) elif args.dataset in ['halueval_summary', 'cnn_dailymail']: return summarization_completion_prompt(dataset) else: raise NotImplementedError(f"Prompt for dataset {args.dataset} not implemented.") if args.dataset=='triviaqa': return triviaqa_prompt(dataset) elif args.dataset=='hotpotqa': return triviaqa_prompt(dataset) elif args.dataset=='coqa': return coqa_prompt(dataset) elif args.dataset=='nq': return triviaqa_prompt(dataset) elif args.dataset=='squad': return coqa_prompt(dataset) elif args.dataset=='math': return triviaqa_prompt(dataset) elif args.dataset=='hotpotqa_c': return coqa_prompt(dataset) elif args.dataset=='psiloqa': return coqa_prompt(dataset) elif args.dataset in ['halueval_summary', 'cnn_dailymail']: return summarization_prompt(dataset) else: raise NotImplementedError(f"Prompt for dataset {args.dataset} not implemented.") def build_prompt_answer(args,dataset,gt_flag=False): prompts=build_prompt(args,dataset) if gt_flag: answers_list=dataset['answers'][0] else: answers_list=dataset['best_answer'] prompt_answers=[ f"{prompt} {ans}" for prompt, ans in zip(prompts,answers_list)] return prompt_answers def build_prompt_candidate_answer(args,dataset): prompts=build_prompt(args,dataset) candidate_answers_list=dataset['candidate_answers'] prompt_candidate_answer=[] sample_idx=[] candidate_answers_flat=[] for idx,(prompt, candidate_answers) in enumerate(zip(prompts,candidate_answers_list)): for ans in candidate_answers: prompt_candidate_answer.append(f"{prompt} {ans}") sample_idx.append(idx) candidate_answers_flat.append(ans) return prompt_candidate_answer,sample_idx,candidate_answers_flat def triviaqa_prompt(dataset:Dataset): prompts = [] for q in dataset['question']: prompts.append(f"Answer the question as briefly as possible, using plain text only:\n Question:{q}\n Answer:") # prompts.append(f""" # Answer the question in English as briefly as possible: # Question:{q} # Answer: # """) return prompts def triviaqa_completion_prompt(dataset: Dataset): prompts = [] for q in dataset["question"]: prompts.append(f"Question: {q}\nAnswer:") return prompts def coqa_prompt(dataset: Dataset): prompts = [] for sample in dataset: context= sample['context'] q= sample['question'] # prompts.append(f'''Context: {context} \n Question: {q} \n Answer: ''') # prompts.append(f"Answer the question based on the context as briefly as possible:\n Context:{context.strip()}\n Question:{q.strip()}\n Answer:") prompts.append(f"Answer the question as briefly as possible, based only on the context:\n Context:{context.strip()}\n Question:{q.strip()}\n Answer:") return prompts def coqa_completion_prompt(dataset: Dataset): prompts = [] for sample in dataset: context = sample["context"] q = sample["question"] prompts.append(f"Context: {context.strip()}\nQuestion: {q.strip()}\nAnswer:") return prompts def summarization_prompt(dataset: Dataset): prompts = [] for sample in dataset: context = sample["context"] prompts.append( "Summarize the following document in one or two concise sentences.\n" # "Use only information supported by the document.\n" f"Document:{context.strip()}\n" "Summary:" ) return prompts def summarization_completion_prompt(dataset: Dataset): prompts = [] for sample in dataset: context = sample["context"] prompts.append( f"Document: {context.strip()}\n" "Summary:" ) return prompts import re strings_to_filter_on = [ '\n', 'Q:', 'A:', 'question:', 'answer:', 'Question:', 'Answer:', 'Questions:', 'questions:', 'QUESTION:', 'ANSWER:', 'REF', '.Forms', 'http', 'php','Question','Answer' ] ALLOWED_CHARS = r"A-Za-z0-9 ,.'\"!?;:-" # characters to keep def clean_english_answer(s: str) -> str: original_s = s # save the original string first # 1. only look at the first line s = s.split('\n')[0].strip() # 2. remove all disallowed characters (e.g. Cyrillic, Chinese, odd symbols) # s = re.sub(fr"[^{ALLOWED_CHARS}]", " ", s) # 3. collapse extra whitespace s = re.sub(r"\s+", " ", s).strip() # 4. shorten long repeated chars like "kkkkkkkkk" (>=3 collapsed to 1) s = re.sub(r"(.)\1{2,}", r"\1", s) # ==== handle periods ==== # 1) normalize ". . . ." to "...." s = re.sub(r"\.\s+(?=\.)", ".", s) # 2) collapse 3+ consecutive periods into a single "." s = re.sub(r"\.{3,}", ".", s) # ==== handle commas ==== # 1) normalize ", , , ," to ",,,," s = re.sub(r",\s+(?=,)", ",", s) # 2) collapse 3+ consecutive commas into a single "," s = re.sub(r",{3,}", ",", s) # 7. strip trailing punctuation and whitespace s = s.strip() # if the result is empty after cleaning, fall back to the original text if not s: return original_s.strip() return s import re # 1) word-level abbreviations (the dot is usually followed by a name/word) WORD_ABBRS = { "Mr", "Mrs", "Ms", "Dr", "Prof", "Sr", "Jr", "Gen", "Brig", "Adm", "Rear", "Lt", "Col", "Maj", "Capt", "St", # St. George "vs", "etc", "Fig", "Eq", "No", } # 2) multi-dot abbreviation patterns: U.S. / U.K. / e.g. / i.e. / G. / J.R.R. etc. MULTI_DOT_ABBR_RE = re.compile(r"(?:[A-Za-z]\.){2,}$") # e.g. U.S. i.e. J.R.R. SINGLE_INITIAL_RE = re.compile(r"^[A-Za-z]$") # G. J. def extract_first_sentence(text: str) -> str: s = text.strip() n = len(s) if n == 0: return s i = 0 while i < n: ch = s[i] if ch not in ".!?": i += 1 continue # A) ellipsis ... if ch == "." and s[i:i+3] == "...": i += 3 continue # B) decimal point 3.14 if ch == "." and 0 < i < n-1 and s[i-1].isdigit() and s[i+1].isdigit(): i += 1 continue # C) token j = i - 1 while j >= 0 and (s[j].isalpha() or s[j] == "."): j -= 1 token_norm = s[j+1:i].strip() # D) multi-dot abbreviation: X.Y. if ch == ".": right_is_letter_dot = (i+2 < n and s[i+1].isalpha() and s[i+2] == ".") left_is_letter = (i-1 >= 0 and s[i-1].isalpha()) if left_is_letter and right_is_letter_dot: i += 1 continue if "." in token_norm and MULTI_DOT_ABBR_RE.match(token_norm + "."): i += 1 continue # E) word abbreviation if token_norm in WORD_ABBRS: if token_norm == "No": k = i + 1 while k < n and s[k].isspace(): k += 1 if k < n and s[k].isdigit(): i += 1 continue else: i += 1 continue # F) personal name initial if SINGLE_INITIAL_RE.match(token_norm): k = i + 1 while k < n and s[k].isspace(): k += 1 if k < n and s[k].isupper(): i += 1 continue # reached here: treat this punctuation as the end of the first sentence return s[:i+1].strip() return s def postprocess_answers(answers, model_name, filters=strings_to_filter_on): cleaned = [] for ans in answers: if ans is None: ans = "" if not isinstance(ans, str): ans = str(ans) original_ans = ans # fallback: revert to original if result is empty after truncation # 1) truncate at keyword markers first cut_pos = len(ans) for f in filters: idx = ans.find(f) if 0 <= idx < cut_pos: cut_pos = idx truncated = ans[:cut_pos].strip() # if truncation by keyword markers yields an empty string, revert to original if not truncated: truncated = original_ans.strip() truncated=extract_first_sentence(truncated) cleaned.append(truncated) return cleaned def measure_correctness(dataset:Dataset,args): labels=[] # matched_ground_truths=[] # d_name=args.dataset if args.dataset=="math": for sample in tqdm(dataset,desc="Measuring correctness"): best_answer=sample['best_answer'] answers= sample['answers'] label=0 for ans in answers: if ans.strip().lower() in best_answer.strip().lower(): label=1 labels.append(label) break ans_float=float(ans) if ans_float.is_integer(): ans_int_str=str(int(ans_float)) if ans_int_str.strip().lower() in best_answer.strip().lower(): label=1 labels.append(label) break if label==0: labels.append(label) # matched_ground_truths.append(answers[0]) elif args.dataset in ["halueval_summary", "cnn_dailymail"]: labels = trueteacher_judge_batch(dataset, args) else: gpu_name = torch.cuda.get_device_name(0) if args.dataset in ['coqa','squad','psiloqa']: bs=16 if gpu_name=='NVIDIA RTX A6000' else 32 else: bs=32 if gpu_name=='NVIDIA RTX A6000' else 64 labels=LLM_judge_batch(dataset,args,batch_size=bs) # matched_ground_truths=[sample['answers'][0] for sample in dataset] # return labels,matched_ground_truths return labels def reevaluate_label(args,remove_NAN_flag=False): # load dataset print(f"reevaluate {args.model} {args.dataset} {args.split}") data_dir=os.path.join(args.basepath_2_save,args.model,args.dataset) json_path=os.path.join(data_dir,f"{args.split}_data.jsonl") df = pd.read_json( json_path, orient="records", lines=True ) if remove_NAN_flag: df=remove_NAN(df) dataset_iter= Dataset.from_pandas(df) # compute correctness labels= measure_correctness(dataset_iter,args) df["label"] = labels # save to json df.to_json( json_path, orient="records", lines=True, ) print(f"Re-evaluated labels saved to {json_path}") def remove_NAN(dataset): if not isinstance(dataset, pd.DataFrame): df = dataset.to_pandas() else: df = dataset orig_rows = len(df) invalid_stats = { "best_answer_missing": 0, "candidate_answers_missing": 0, "candidate_answers_empty_list": 0, "candidate_answers_contains_empty": 0, } def row_valid(row): # 1) check best_answer ba = row.get("best_answer", None) # NaN / None / empty string are all treated as invalid if ba is None or (isinstance(ba, float) and pd.isna(ba)): invalid_stats["best_answer_missing"] += 1 return False if isinstance(ba, str) and ba.strip() == "": invalid_stats["best_answer_missing"] += 1 return False # 2) check candidate_answers ca = row.get("candidate_answers", None) if ca is None or (isinstance(ca, float) and pd.isna(ca)): invalid_stats["candidate_answers_missing"] += 1 return False if len(ca) == 0: invalid_stats["candidate_answers_empty_list"] += 1 return False # list must not contain None / NaN / empty string for x in ca: if x is None: invalid_stats["candidate_answers_contains_empty"] += 1 return False if isinstance(x, float) and pd.isna(x): invalid_stats["candidate_answers_contains_empty"] += 1 return False if isinstance(x, str) and x.strip() == "": invalid_stats["candidate_answers_contains_empty"] += 1 return False return True mask = df.apply(row_valid, axis=1) df_clean = df[mask].reset_index(drop=True) clean_rows = len(df_clean) removed_rows = orig_rows - clean_rows print(f"original lines: {orig_rows}") print(f"deleted: {removed_rows}") print("deletion breakdown:") print(f" best_answer missing/empty: {invalid_stats['best_answer_missing']}") print(f" candidate_answers missing: {invalid_stats['candidate_answers_missing']}") print(f" candidate_answers empty list: {invalid_stats['candidate_answers_empty_list']}") print(f" candidate_answers contains empty item: {invalid_stats['candidate_answers_contains_empty']}") print(f"left: {clean_rows}") return df_clean def _is_summary_dataset(dataset_name: str) -> bool: return dataset_name in {"halueval_summary", "cnn_dailymail"} def get_trueteacher_input(context: str, model_summary: str) -> str: context_text = "" if context is None else str(context).strip() if not context_text: raise ValueError("TrueTeacher requires a non-empty 'context' field for summary datasets.") summary_text = "" if model_summary is None else str(model_summary).strip() return f"premise: {context_text} hypothesis: {summary_text}" def _get_trueteacher_batch_size() -> int: if not torch.cuda.is_available(): return TRUETEACHER_DEFAULT_BATCH_SIZE gpu_name = torch.cuda.get_device_name(0) if gpu_name == "NVIDIA RTX A6000": return TRUETEACHER_A6000_BATCH_SIZE return TRUETEACHER_DEFAULT_BATCH_SIZE def _get_model_input_device(model) -> torch.device: if hasattr(model, "hf_device_map"): for device in model.hf_device_map.values(): if isinstance(device, int): return torch.device(f"cuda:{device}") if isinstance(device, str) and device.startswith("cuda"): return torch.device(device) return next(model.parameters()).device def trueteacher_judge_batch(dataset, args, batch_size=None): try: from transformers import T5ForConditionalGeneration, T5Tokenizer except ImportError as exc: raise ImportError( "TrueTeacher evaluation requires transformers with T5 support installed." ) from exc if batch_size is None: batch_size = _get_trueteacher_batch_size() model_kwargs = { "cache_dir": args.model_cache_dir, "low_cpu_mem_usage": True, } if torch.cuda.is_available(): model_kwargs["device_map"] = "auto" if torch.cuda.is_bf16_supported(): model_kwargs["torch_dtype"] = torch.bfloat16 else: model_kwargs["torch_dtype"] = torch.float16 try: tokenizer = T5Tokenizer.from_pretrained( TRUETEACHER_MODEL_ID, cache_dir=args.model_cache_dir, ) model = T5ForConditionalGeneration.from_pretrained( TRUETEACHER_MODEL_ID, **model_kwargs, ) except Exception as exc: raise RuntimeError( f"Failed to load TrueTeacher model '{TRUETEACHER_MODEL_ID}' " f"from cache_dir={args.model_cache_dir}. Original error: {exc}" ) from exc input_device = _get_model_input_device(model) labels = [] for start in tqdm(range(0, len(dataset), batch_size), desc="TrueTeacher judging"): batch = dataset[start:start + batch_size] batch_inputs = [] empty_summary_flags = [] for context, best_answer in zip(batch["context"], batch["best_answer"]): summary_text = "" if best_answer is None else str(best_answer).strip() if not summary_text: batch_inputs.append(None) empty_summary_flags.append(True) continue batch_inputs.append(get_trueteacher_input(context, summary_text)) empty_summary_flags.append(False) valid_inputs = [text for text in batch_inputs if text is not None] decoded_outputs = [] if valid_inputs: input_id = tokenizer( valid_inputs, return_tensors="pt", padding=True, truncation=True, max_length=TRUETEACHER_MAX_LENGTH, ).input_ids.to(input_device) try: with torch.inference_mode(): generated = model.generate( input_ids=input_id, ) except RuntimeError as exc: raise RuntimeError( "TrueTeacher inference failed. This may be caused by insufficient GPU memory. " "Try reducing the batch size or running on a larger GPU." ) from exc decoded_outputs = tokenizer.batch_decode(generated, skip_special_tokens=True) decoded_iter = iter(decoded_outputs) for is_empty in empty_summary_flags: if is_empty: labels.append(0) continue text = next(decoded_iter).strip() if text.startswith("1"): labels.append(1) elif text.startswith("0"): labels.append(0) else: print(f"Invalid TrueTeacher output: {text!r}, defaulting to 0") labels.append(0) return labels def LLM_judge_batch(dataset, args, batch_size=32): """ dataset: HF Dataset with fields: - 'question' - 'best_answer' - 'answers' (list; only the 0th element is used as reference) - 'context' (optional, used for prompt) """ model, tokenizer = load_model_and_validate_gpu( # 'mistralai/Mistral-7B-Instruct-v0.2', 'mistralai/Ministral-8B-Instruct-2410', cache_dir=args.model_cache_dir ) tokenizer.pad_token = tokenizer.eos_token tokenizer.padding_side = 'left' device = model.device n = len(dataset) correctness = [None] * n # placeholder indices_to_judge = [] # indices of samples that need LLM judging prompts_to_judge = [] # corresponding prompts # ---------- stage 1: string match ---------- for idx, sample in enumerate(tqdm(dataset, desc="String match pre-filter")): model_answer = sample["best_answer"] if _is_summary_dataset(args.dataset): context = sample.get("context", None) prompt = get_summary_prompt(context, sample.get("answers", []), model_answer) indices_to_judge.append(idx) prompts_to_judge.append(prompt) continue question = sample["question"] # simple string match (lowercased + stripped) if any( answ.strip().lower() in model_answer.strip().lower() for answ in sample["answers"] ): correctness[idx] = 1 continue # unmatched samples: build prompt for batched LLM judging context = sample.get("context", None) prompt = get_prompt(context, question, sample["answers"], model_answer) indices_to_judge.append(idx) prompts_to_judge.append(prompt) # stats: total samples that failed string match (i.e. no hit) total_to_judge = len(indices_to_judge) num_llm_1 = 0 # of those, how many the LLM judged as 1 # ---------- stage 2: batched LLM-as-judge ---------- for start in tqdm(range(0, len(prompts_to_judge), batch_size), desc="LLM judging"): end = min(start + batch_size, len(prompts_to_judge)) batch_prompts = prompts_to_judge[start:end] batch_indices = indices_to_judge[start:end] enc = tokenizer( batch_prompts, return_tensors="pt", padding=True, ).to(device) with torch.no_grad(): outputs = model.generate( **enc, max_new_tokens=30, do_sample=True, temperature=0.1, top_k=50, top_p=1.0, return_dict_in_generate=True, eos_token_id=None, ) # newly generated tokens: everything after the prompt length seqs = outputs["sequences"] prompt_len = enc["input_ids"].shape[1] # same for all samples due to padding gen_tokens = seqs[:, prompt_len:] for i, seq in enumerate(gen_tokens): text_decoded = tokenizer.decode(seq, skip_special_tokens=True) # light cleanup text = ( text_decoded.replace(".", "") .replace("", "") .split("\n")[0] .strip() .strip(".") ) idx1 = text.find("1") idx0 = text.find("0") if idx1 != -1 and (idx0 == -1 or idx1 < idx0): lab = 1 elif idx0 != -1 and (idx1 == -1 or idx0 < idx1): lab = 0 else: print(f"Invalid judge output: {text!r}, default to 0") lab = 0 # count: among string-match=0 samples, how many the LLM judged as 1 if lab == 1: num_llm_1 += 1 # write back to the corresponding global index global_idx = batch_indices[i] correctness[global_idx] = lab # print summary statistics if total_to_judge > 0: ratio = num_llm_1 / total_to_judge else: ratio = 0.0 print( f"[LLM-judge stats] " f"Number of string-match=0: {total_to_judge}, " f"Numer of LLM judge 1: {num_llm_1} " f"({ratio:.2%})" ) return correctness def get_prompt(context,question,reference_answers,model_answer): reference_answer="; ".join(reference_answers) if context is None: prompt = f""" Evaluate the following answers to questions. For each question you are given a model answer and the correct answer. You must determine if the model answer is correct or not. If the model answer is correct, write '1' and if it is not correct, write '0'. For example: Question: who is the young guitarist who played with buddy guy? Ground Truth: Quinn Sullivan Model Answer: Ronnie Earl Explanation: Ronnie Earl is an American blues guitarist and singer who has played with many famous blues musicians, including Buddy Guy. He is known for his soulful and melodic playing style, and has released many albums that blend blues, jazz, and rock music. Earl has also been a member of the Buddy Guy Blues Band and has played with other notable blues musicians such as B.B. King, Eric Clapton, and Stevie Ray Vaughan. He is considered one of the most Correctness: 0 Question: name of the first episode of stranger things Ground Truth: Chapter One : The Vanishing of Will Byers Model Answer: The disappearance of Will Byers. Explanation: The first episode of the first season of Stranger Things is titled "The Vanishing of Will Byers". The episode introduces the main characters and sets the tone for the rest of the series. It follows the story of Will Byers, a young boy who goes missing in the fictional town of Hawkins, Indiana, and the subsequent search for him by his mother Joyce and his friends Mike, Dustin, and Lucas. The episode sets the stage for the supernatural Correctness: 1 Question: {question} Ground Truth: {reference_answer} Model Answer: {model_answer} Correctness: """.strip() else: prompt = f""" Evaluate the following answers to questions. For each question you are given a model answer and the correct answer. You must determine if the model answer is correct or not. If the model answer is correct, write '1' and if it is not correct, write '0'. For example: Question: who is the young guitarist who played with buddy guy? Ground Truth: Quinn Sullivan Model Answer: Ronnie Earl Explanation: Ronnie Earl is an American blues guitarist and singer who has played with many famous blues musicians, including Buddy Guy. He is known for his soulful and melodic playing style, and has released many albums that blend blues, jazz, and rock music. Earl has also been a member of the Buddy Guy Blues Band and has played with other notable blues musicians such as B.B. King, Eric Clapton, and Stevie Ray Vaughan. He is considered one of the most Correctness: 0 Question: name of the first episode of stranger things Ground Truth: Chapter One : The Vanishing of Will Byers Model Answer: The disappearance of Will Byers. Explanation: The first episode of the first season of Stranger Things is titled "The Vanishing of Will Byers". The episode introduces the main characters and sets the tone for the rest of the series. It follows the story of Will Byers, a young boy who goes missing in the fictional town of Hawkins, Indiana, and the subsequent search for him by his mother Joyce and his friends Mike, Dustin, and Lucas. The episode sets the stage for the supernatural Correctness: 1 Context: {context} Question: {question} Ground Truth: {reference_answer} Model Answer: {model_answer} Correctness: """.strip() return prompt def get_summary_prompt(context, reference_answers, model_summary): reference_summary = "; ".join( [str(ans).strip() for ans in reference_answers if str(ans).strip()] ) or "N/A" prompt = f""" Evaluate whether the model summary is faithful to the source document. Write '1' if the summary is fully supported by the document and does not introduce materially unsupported information. Write '0' if the summary contains hallucinated, contradictory, or unsupported content. If the summary is incomplete but still faithful, prefer '1'. Source Document: {context} Reference Summary: {reference_summary} Model Summary: {model_summary} Faithfulness: """.strip() return prompt def _clean_best_answer_cell(x,args): """Process a single best_answer cell (string).""" if x is None or not isinstance(x, str): raise ValueError( f"best_answer cell must be a non-empty string, got {type(x)}: {x!r}" ) # postprocess_answers expects list[str]; pass a single-element list return postprocess_answers([x],args.model)[0] def _clean_candidate_answers_cell(x,args): """Process a single candidate_answers cell.""" if x is None: raise ValueError("candidate_answers cell must be a list of strings, got None") # neither a list nor a bare string is a valid type here if not isinstance(x, list) or isinstance(x, str): raise ValueError( f"candidate_answers cell must be a list of strings, got {type(x)}: {x!r}" ) return postprocess_answers(x,args.model) def re_post_process(args): """ Re-post-process an already generated dataset: - clean 'best_answer' - clean 'candidate_answers' (list of strings) Overwrites the existing {split}_data.jsonl in place. """ print(f"repost {args.model} {args.dataset} {args.split}") data_dir = os.path.join(args.basepath_2_save, args.model, args.dataset) json_path = os.path.join(data_dir, f"{args.split}_data.jsonl") print(f"Loading dataset from: {json_path}") df = pd.read_json(json_path, orient="records", lines=True) assert "best_answer" in df.columns assert "candidate_answers" in df.columns print("Post-processing 'best_answer' column...") df["best_answer"] = df["best_answer"].apply(_clean_best_answer_cell,args=(args,),) print("Post-processing 'candidate_answers' column...") df["candidate_answers"] = df["candidate_answers"].apply(_clean_candidate_answers_cell,args=(args,),) df=remove_NAN(df) # overwrite the original file df.to_json( json_path, orient="records", lines=True, ) print(f"Re-post-processed dataset saved to: {json_path}")