import os import torch import torch.nn.functional as F import evaluate from datasets import load_metric from datasets import load_dataset import datasets from tqdm import tqdm import numpy as np import pickle # from utils import get_llama_activations_bau, tokenized_tqa, tokenized_tqa_gen, tokenized_tqa_gen_end_q import llama_iti import pickle import argparse import matplotlib.pyplot as plt from pprint import pprint from baukit import Trace, TraceDict from metric_utils import get_measures, print_measures import re from torch.autograd import Variable def seed_everything(seed: int): import random, os import numpy as np import torch random.seed(seed) os.environ['PYTHONHASHSEED'] = str(seed) np.random.seed(seed) torch.manual_seed(seed) torch.cuda.manual_seed(seed) torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = True HF_NAMES = { 'llama_7B': 'baffo32/decapoda-research-llama-7B-hf', 'honest_llama_7B': 'validation/results_dump/llama_7B_seed_42_top_48_heads_alpha_15', 'alpaca_7B': 'circulus/alpaca-7b', 'vicuna_7B': 'AlekseyKorshuk/vicuna-7b', 'llama2_chat_7B': 'models/Llama-2-7b-chat-hf', 'llama2_chat_13B': 'models/Llama-2-13b-chat-hf', 'llama2_chat_70B': 'meta-llama/Llama-2-70b-chat-hf', } def main(): parser = argparse.ArgumentParser() parser.add_argument('--model_name', type=str, default='llama2_chat_7B') parser.add_argument('--dataset_name', type=str, default='triviaqa') parser.add_argument('--num_gene', type=int, default=1) parser.add_argument('--gene', type=int, default=0) parser.add_argument('--generate_gt', type=int, default=0) parser.add_argument('--use_rouge', type=int, default=0) parser.add_argument('--weighted_svd', type=int, default=0) parser.add_argument('--feat_loc_svd', type=int, default=0) parser.add_argument('--wild_ratio', type=float, default=0.75) parser.add_argument('--thres_gt', type=float, default=0.5) parser.add_argument('--most_likely', type=int, default=0) parser.add_argument("--model_dir", type=str, default=None, help='local directory with model data') args = parser.parse_args() MODEL = HF_NAMES[args.model_name] if not args.model_dir else args.model_dir if args.dataset_name == "tqa": dataset = load_dataset("truthful_qa", 'generation')['validation'] elif args.dataset_name == 'triviaqa': dataset = load_dataset("trivia_qa", "rc.nocontext", split="validation") id_mem = set() def remove_dups(batch): if batch['question_id'][0] in id_mem: return {_: [] for _ in batch.keys()} id_mem.add(batch['question_id'][0]) return batch dataset = dataset.map(remove_dups, batch_size=1, batched=True, load_from_cache_file=False) elif args.dataset_name == 'tydiqa': dataset = datasets.load_dataset("tydiqa", "secondary_task", split="train") used_indices = [] for i in range(len(dataset)): if 'english' in dataset[i]['id']: used_indices.append(i) elif args.dataset_name == 'coqa': import json import pandas as pd from datasets import Dataset def _save_dataset(): # https://github.com/lorenzkuhn/semantic_uncertainty/blob/main/code/parse_coqa.py save_path = f'./coqa_dataset' if not os.path.exists(save_path): # https://downloads.cs.stanford.edu/nlp/data/coqa/coqa-dev-v1.0.json with open(f'./coqa-dev-v1.0.json', 'r') as infile: data = json.load(infile)['data'] dataset = {} dataset['story'] = [] dataset['question'] = [] dataset['answer'] = [] dataset['additional_answers'] = [] dataset['id'] = [] for sample_id, sample in enumerate(data): story = sample['story'] questions = sample['questions'] answers = sample['answers'] additional_answers = sample['additional_answers'] for question_index, question in enumerate(questions): dataset['story'].append(story) dataset['question'].append(question['input_text']) dataset['answer'].append({ 'text': answers[question_index]['input_text'], 'answer_start': answers[question_index]['span_start'] }) dataset['id'].append(sample['id'] + '_' + str(question_index)) additional_answers_list = [] for i in range(3): additional_answers_list.append(additional_answers[str(i)][question_index]['input_text']) dataset['additional_answers'].append(additional_answers_list) story = story + ' Q: ' + question['input_text'] + ' A: ' + answers[question_index]['input_text'] if not story[-1] == '.': story = story + '.' dataset_df = pd.DataFrame.from_dict(dataset) dataset = Dataset.from_pandas(dataset_df) dataset.save_to_disk(save_path) return save_path # dataset = datasets.load_from_disk(_save_dataset()) def get_dataset(tokenizer, split='validation'): # from https://github.com/lorenzkuhn/semantic_uncertainty/blob/main/code/parse_coqa.py dataset = datasets.load_from_disk(_save_dataset()) id_to_question_mapping = dict(zip(dataset['id'], dataset['question'])) def encode_coqa(example): example['answer'] = [example['answer']['text']] + example['additional_answers'] example['prompt'] = prompt = example['story'] + ' Q: ' + example['question'] + ' A:' return tokenizer(prompt, truncation=False, padding=False) dataset = dataset.map(encode_coqa, batched=False, load_from_cache_file=False) dataset.set_format(type='torch', columns=['input_ids', 'attention_mask'], output_all_columns=True) return dataset dataset = get_dataset(llama_iti.LlamaTokenizer.from_pretrained(MODEL, trust_remote_code=True)) else: raise ValueError("Invalid dataset name") if args.gene: tokenizer = llama_iti.LlamaTokenizer.from_pretrained(MODEL, trust_remote_code=True) model = llama_iti.LlamaForCausalLM.from_pretrained(MODEL, low_cpu_mem_usage=True, torch_dtype=torch.float16, device_map="auto").cuda() begin_index = 0 if args.dataset_name == 'tydiqa': end_index = len(used_indices) else: end_index = len(dataset) if not os.path.exists(f'./save_for_eval/{args.dataset_name}_hal_det/'): os.mkdir(f'./save_for_eval/{args.dataset_name}_hal_det/') if not os.path.exists(f'./save_for_eval/{args.dataset_name}_hal_det/answers'): os.mkdir(f'./save_for_eval/{args.dataset_name}_hal_det/answers') period_token_id = [tokenizer(_)['input_ids'][-1] for _ in ['\n']] period_token_id += [tokenizer.eos_token_id] for i in range(begin_index, end_index): answers = [None] * args.num_gene if args.dataset_name == 'tydiqa': question = dataset[int(used_indices[i])]['question'] prompt = tokenizer( "Concisely answer the following question based on the information in the given passage: \n" + \ " Passage: " + dataset[int(used_indices[i])]['context'] + " \n Q: " + question + " \n A:", return_tensors='pt').input_ids.cuda() elif args.dataset_name == 'coqa': prompt = tokenizer( dataset[i]['prompt'], return_tensors='pt').input_ids.cuda() else: question = dataset[i]['question'] prompt = tokenizer(f"Answer the question concisely. Q: {question}" + " A:", return_tensors='pt').input_ids.cuda() for gen_iter in range(args.num_gene): if args.most_likely: generated = model.generate(prompt, num_beams=5, num_return_sequences=1, do_sample=False, max_new_tokens=64, ) else: generated = model.generate(prompt, do_sample=True, num_return_sequences=1, num_beams=1, max_new_tokens=64, temperature=0.5, top_p=1.0) decoded = tokenizer.decode(generated[0, prompt.shape[-1]:], skip_special_tokens=True) if args.dataset_name == 'tqa' or args.dataset_name == 'triviaqa': # corner case. if 'Answer the question concisely' in decoded: print('#####error') print(decoded.split('Answer the question concisely')[1]) print('#####error') decoded = decoded.split('Answer the question concisely')[0] if args.dataset_name == 'coqa': if 'Q:' in decoded: print('#####error') print(decoded.split('Q:')[1]) print('#####error') decoded = decoded.split('Q:')[0] print(decoded) answers[gen_iter] = decoded print('sample: ', i) if args.most_likely: info = 'most_likely_' else: info = 'batch_generations_' print("Saving answers") np.save(f'./save_for_eval/{args.dataset_name}_hal_det/answers/' + info + f'hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy', answers) elif args.generate_gt: from bleurt_pytorch import BleurtConfig, BleurtForSequenceClassification, BleurtTokenizer model = BleurtForSequenceClassification.from_pretrained('lucadiliello/BLEURT-20').cuda() tokenizer = BleurtTokenizer.from_pretrained('lucadiliello/BLEURT-20') model.eval() rouge = evaluate.load('rouge') gts = np.zeros(0) if args.dataset_name == 'tydiqa': length = len(used_indices) else: length = len(dataset) for i in range(length): if args.dataset_name == 'tqa': best_answer = dataset[i]['best_answer'] correct_answer = dataset[i]['correct_answers'] all_answers = [best_answer] + correct_answer elif args.dataset_name == 'triviaqa': all_answers = dataset[i]['answer']['aliases'] elif args.dataset_name == 'coqa': all_answers = dataset[i]['answer'] elif args.dataset_name == 'tydiqa': all_answers = dataset[int(used_indices[i])]['answers']['text'] if args.most_likely: answers = np.load( f'./save_for_eval/{args.dataset_name}_hal_det/answers/most_likely_hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy') else: answers = np.load( f'./save_for_eval/{args.dataset_name}_hal_det/answers/batch_generations_hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy') # get the gt. if args.use_rouge: predictions = answers all_results = np.zeros((len(all_answers), len(predictions))) all_results1 = np.zeros((len(all_answers), len(predictions))) all_results2 = np.zeros((len(all_answers), len(predictions))) for anw in range(len(all_answers)): results = rouge.compute(predictions=predictions, references=[all_answers[anw]] * len(predictions), use_aggregator=False) all_results[anw] = results['rougeL'] all_results1[anw] = results['rouge1'] all_results2[anw] = results['rouge2'] # breakpoint() gts = np.concatenate([gts, np.max(all_results, axis=0)], 0) if i % 50 == 0: print("samples passed: ", i) else: predictions = answers all_results = np.zeros((len(all_answers), len(predictions))) with torch.no_grad(): for anw in range(len(all_answers)): inputs = tokenizer(predictions.tolist(), [all_answers[anw]] * len(predictions), padding='longest', return_tensors='pt') for key in list(inputs.keys()): inputs[key] = inputs[key].cuda() res = np.asarray(model(**inputs).logits.flatten().tolist()) all_results[anw] = res gts = np.concatenate([gts, np.max(all_results, axis=0)], 0) if i % 10 == 0: print("samples passed: ", i) # breakpoint() if args.most_likely: if args.use_rouge: np.save(f'./ml_{args.dataset_name}_rouge_score.npy', gts) else: np.save(f'./ml_{args.dataset_name}_bleurt_score.npy', gts) else: if args.use_rouge: np.save(f'./bg_{args.dataset_name}_rouge_score.npy', gts) else: np.save(f'./bg_{args.dataset_name}_bleurt_score.npy', gts) else: tokenizer = llama_iti.LlamaTokenizer.from_pretrained(MODEL, trust_remote_code=True) model = llama_iti.LlamaForCausalLM.from_pretrained(MODEL, low_cpu_mem_usage=True, torch_dtype=torch.float16, device_map="auto").cuda() # firstly get the embeddings of the generated question and answers. embed_generated = [] if args.dataset_name == 'tydiqa': length = len(used_indices) else: length = len(dataset) for i in tqdm(range(length)): if args.dataset_name == 'tydiqa': question = dataset[int(used_indices[i])]['question'] else: question = dataset[i]['question'] answers = np.load( f'save_for_eval/{args.dataset_name}_hal_det/answers/most_likely_hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy') for anw in answers: if args.dataset_name == 'tydiqa': prompt = tokenizer( "Concisely answer the following question based on the information in the given passage: \n" + \ " Passage: " + dataset[int(used_indices[i])]['context'] + " \n Q: " + question + " \n A:", return_tensors='pt').input_ids.cuda() elif args.dataset_name == 'coqa': prompt = tokenizer(dataset[i]['prompt'] + anw, return_tensors='pt').input_ids.cuda() else: prompt = tokenizer( f"Answer the question concisely. Q: {question}" + " A:" + anw, return_tensors='pt').input_ids.cuda() with torch.no_grad(): hidden_states = model(prompt, output_hidden_states=True).hidden_states hidden_states = torch.stack(hidden_states, dim=0).squeeze() hidden_states = hidden_states.detach().cpu().numpy()[:, -1, :] embed_generated.append(hidden_states) embed_generated = np.asarray(np.stack(embed_generated), dtype=np.float32) np.save(f'save_for_eval/{args.dataset_name}_hal_det/most_likely_{args.model_name}_gene_embeddings_layer_wise.npy', embed_generated) HEADS = [f"model.layers.{i}.self_attn.head_out" for i in range(model.config.num_hidden_layers)] MLPS = [f"model.layers.{i}.mlp" for i in range(model.config.num_hidden_layers)] embed_generated_loc2 = [] embed_generated_loc1 = [] for i in tqdm(range(length)): if args.dataset_name == 'tydiqa': question = dataset[int(used_indices[i])]['question'] else: question = dataset[i]['question'] answers = np.load( f'save_for_eval/{args.dataset_name}_hal_det/answers/most_likely_hal_det_{args.model_name}_{args.dataset_name}_answers_index_{i}.npy') for anw in answers: if args.dataset_name == 'tydiqa': prompt = tokenizer( "Concisely answer the following question based on the information in the given passage: \n" + \ " Passage: " + dataset[int(used_indices[i])]['context'] + " \n Q: " + question + " \n A:", return_tensors='pt').input_ids.cuda() elif args.dataset_name == 'coqa': prompt = tokenizer(dataset[i]['prompt'] + anw, return_tensors='pt').input_ids.cuda() else: prompt = tokenizer( f"Answer the question concisely. Q: {question}" + " A:" + anw, return_tensors='pt').input_ids.cuda() with torch.no_grad(): with TraceDict(model, HEADS + MLPS) as ret: output = model(prompt, output_hidden_states=True) head_wise_hidden_states = [ret[head].output.squeeze().detach().cpu() for head in HEADS] head_wise_hidden_states = torch.stack(head_wise_hidden_states, dim=0).squeeze().numpy() mlp_wise_hidden_states = [ret[mlp].output.squeeze().detach().cpu() for mlp in MLPS] mlp_wise_hidden_states = torch.stack(mlp_wise_hidden_states, dim=0).squeeze().numpy() embed_generated_loc2.append(mlp_wise_hidden_states[:, -1, :]) embed_generated_loc1.append(head_wise_hidden_states[:, -1, :]) embed_generated_loc2 = np.asarray(np.stack(embed_generated_loc2), dtype=np.float32) embed_generated_loc1 = np.asarray(np.stack(embed_generated_loc1), dtype=np.float32) np.save(f'save_for_eval/{args.dataset_name}_hal_det/most_likely_{args.model_name}_gene_embeddings_head_wise.npy', embed_generated_loc1) np.save(f'save_for_eval/{args.dataset_name}_hal_det/most_likely_{args.model_name}_embeddings_mlp_wise.npy', embed_generated_loc2) # get the split and label (true or false) of the unlabeled data and the test data. if args.use_rouge: gts = np.load(f'./ml_{args.dataset_name}_rouge_score.npy') gts_bg = np.load(f'./bg_{args.dataset_name}_rouge_score.npy') else: gts = np.load(f'./ml_{args.dataset_name}_bleurt_score.npy') gts_bg = np.load(f'./bg_{args.dataset_name}_bleurt_score.npy') thres = args.thres_gt gt_label = np.asarray(gts> thres, dtype=np.int32) gt_label_bg = np.asarray(gts_bg > thres, dtype=np.int32) if args.dataset_name == 'tydiqa': length = len(used_indices) else: length = len(dataset) permuted_index = np.random.permutation(length) wild_q_indices = permuted_index[:int(args.wild_ratio * length)] # exclude validation samples. wild_q_indices1 = wild_q_indices[:len(wild_q_indices) - 100] wild_q_indices2 = wild_q_indices[len(wild_q_indices) - 100:] gt_label_test = [] gt_label_wild = [] gt_label_val = [] for i in range(length): if i not in wild_q_indices: gt_label_test.extend(gt_label[i: i+1]) elif i in wild_q_indices1: gt_label_wild.extend(gt_label[i: i+1]) else: gt_label_val.extend(gt_label[i: i+1]) gt_label_test = np.asarray(gt_label_test) gt_label_wild = np.asarray(gt_label_wild) gt_label_val = np.asarray(gt_label_val) def svd_embed_score(embed_generated_wild, gt_label, begin_k, k_span, mean=1, svd=1, weight=0): embed_generated = embed_generated_wild best_auroc_over_k = 0 best_layer_over_k = 0 best_scores_over_k = None best_projection_over_k = None for k in tqdm(range(begin_k, k_span)): best_auroc = 0 best_layer = 0 best_scores = None mean_recorded = None best_projection = None for layer in range(len(embed_generated_wild[0])): if mean: mean_recorded = embed_generated[:, layer, :].mean(0) centered = embed_generated[:, layer, :] - mean_recorded else: centered = embed_generated[:, layer, :] if not svd: pca_model = PCA(n_components=k, whiten=False).fit(centered) projection = pca_model.components_.T mean_recorded = pca_model.mean_ if weight: projection = pca_model.singular_values_ * projection else: _, sin_value, V_p = torch.linalg.svd(torch.from_numpy(centered).cuda()) projection = V_p[:k, :].T.cpu().data.numpy() if weight: projection = sin_value[:k] * projection scores = np.mean(np.matmul(centered, projection), -1, keepdims=True) assert scores.shape[1] == 1 scores = np.sqrt(np.sum(np.square(scores), axis=1)) # not sure about whether true and false data the direction will point to, # so we test both. similar practices are in the representation engineering paper # https://arxiv.org/abs/2310.01405 measures1 = get_measures(scores[gt_label == 1], scores[gt_label == 0], plot=False) measures2 = get_measures(-scores[gt_label == 1], -scores[gt_label == 0], plot=False) if measures1[0] > measures2[0]: measures = measures1 sign_layer = 1 else: measures = measures2 sign_layer = -1 if measures[0] > best_auroc: best_auroc = measures[0] best_result = [100 * measures[2], 100 * measures[0]] best_layer = layer best_scores = sign_layer * scores best_projection = projection best_mean = mean_recorded best_sign = sign_layer print('k: ', k, 'best result: ', best_result, 'layer: ', best_layer, 'mean: ', mean, 'svd: ', svd) if best_auroc > best_auroc_over_k: best_auroc_over_k = best_auroc best_result_over_k = best_result best_layer_over_k = best_layer best_k = k best_sign_over_k = best_sign best_scores_over_k = best_scores best_projection_over_k = best_projection best_mean_over_k = best_mean return {'k': best_k, 'best_layer':best_layer_over_k, 'best_auroc':best_auroc_over_k, 'best_result':best_result_over_k, 'best_scores':best_scores_over_k, 'best_mean': best_mean_over_k, 'best_sign':best_sign_over_k, 'best_projection':best_projection_over_k} from sklearn.decomposition import PCA feat_loc = args.feat_loc_svd if args.most_likely: if feat_loc == 3: embed_generated = np.load(f'save_for_eval/{args.dataset_name}_hal_det/most_likely_{args.model_name}_gene_embeddings_layer_wise.npy', allow_pickle=True) elif feat_loc == 2: embed_generated = np.load( f'save_for_eval/{args.dataset_name}_hal_det/most_likely_{args.model_name}_gene_embeddings_mlp_wise.npy', allow_pickle=True) else: embed_generated = np.load( f'save_for_eval/{args.dataset_name}_hal_det/most_likely_{args.model_name}_gene_embeddings_head_wise.npy', allow_pickle=True) feat_indices_wild = [] feat_indices_eval = [] if args.dataset_name == 'tydiqa': length = len(used_indices) else: length = len(dataset) for i in range(length): if i in wild_q_indices1: feat_indices_wild.extend(np.arange(i, i+1).tolist()) elif i in wild_q_indices2: feat_indices_eval.extend(np.arange(i, i + 1).tolist()) if feat_loc == 3: embed_generated_wild = embed_generated[feat_indices_wild][:,1:,:] embed_generated_eval = embed_generated[feat_indices_eval][:, 1:, :] else: embed_generated_wild = embed_generated[feat_indices_wild] embed_generated_eval = embed_generated[feat_indices_eval] # returned_results = svd_embed_score(embed_generated_wild, gt_label_wild, # 1, 11, mean=0, svd=0, weight=args.weighted_svd) # get the best hyper-parameters on validation set returned_results = svd_embed_score(embed_generated_eval, gt_label_val, 1, 11, mean=0, svd=0, weight=args.weighted_svd) pca_model = PCA(n_components=returned_results['k'], whiten=False).fit(embed_generated_wild[:,returned_results['best_layer'],:]) projection = pca_model.components_.T if args.weighted_svd: projection = pca_model.singular_values_ * projection scores = np.mean(np.matmul(embed_generated_wild[:,returned_results['best_layer'],:], projection), -1, keepdims=True) assert scores.shape[1] == 1 best_scores = np.sqrt(np.sum(np.square(scores), axis=1)) * returned_results['best_sign'] # direct projection feat_indices_test = [] for i in range(length): if i not in wild_q_indices: feat_indices_test.extend(np.arange(1 * i, 1 * i + 1).tolist()) if feat_loc == 3: embed_generated_test = embed_generated[feat_indices_test][:, 1:, :] else: embed_generated_test = embed_generated[feat_indices_test] test_scores = np.mean(np.matmul(embed_generated_test[:,returned_results['best_layer'],:], projection), -1, keepdims=True) assert test_scores.shape[1] == 1 test_scores = np.sqrt(np.sum(np.square(test_scores), axis=1)) measures = get_measures(returned_results['best_sign'] * test_scores[gt_label_test == 1], returned_results['best_sign'] *test_scores[gt_label_test == 0], plot=False) print_measures(measures[0], measures[1], measures[2], 'direct-projection') thresholds = np.linspace(0,1, num=40)[1:-1] normalizer = lambda x: x / (np.linalg.norm(x, ord=2, axis=-1, keepdims=True) + 1e-10) auroc_over_thres = [] for thres_wild in thresholds: best_auroc = 0 for layer in range(len(embed_generated_wild[0])): thres_wild_score = np.sort(best_scores)[int(len(best_scores) * thres_wild)] true_wild = embed_generated_wild[:,layer,:][best_scores > thres_wild_score] false_wild = embed_generated_wild[:,layer,:][best_scores <= thres_wild_score] embed_train = np.concatenate([true_wild,false_wild],0) label_train = np.concatenate([np.ones(len(true_wild)), np.zeros(len(false_wild))], 0) ## gt training, saplma # embed_train = embed_generated_wild[:,layer,:] # label_train = gt_label_wild ## gt training, saplma from linear_probe import get_linear_acc best_acc, final_acc, ( clf, best_state, best_preds, preds, labels_val), losses_train = get_linear_acc( embed_train, label_train, embed_train, label_train, 2, epochs = 50, print_ret = True, batch_size=512, cosine=True, nonlinear = True, learning_rate = 0.05, weight_decay = 0.0003) clf.eval() output = clf(torch.from_numpy( embed_generated_test[:, layer, :]).cuda()) pca_wild_score_binary_cls = torch.sigmoid(output) pca_wild_score_binary_cls = pca_wild_score_binary_cls.cpu().data.numpy() if np.isnan(pca_wild_score_binary_cls).sum() > 0: breakpoint() measures = get_measures(pca_wild_score_binary_cls[gt_label_test == 1], pca_wild_score_binary_cls[gt_label_test == 0], plot=False) if measures[0] > best_auroc: best_auroc = measures[0] best_result = [100 * measures[0]] best_layer = layer auroc_over_thres.append(best_auroc) print('thres: ', thres_wild, 'best result: ', best_result, 'best_layer: ', best_layer) if __name__ == '__main__': seed_everything(42) main()