├── .gitignore ├── assets ├── intro.png └── toy.png ├── requirements.txt ├── templates └── alpaca.json ├── fine-tuning-unlike.sh ├── merge.py ├── prompter.py ├── scripts ├── generate.py └── convert2alpaca.py ├── README.md ├── LICENSE ├── modeling_llama_unlikelihood.py └── finetune_unlikelihood.py /.gitignore: -------------------------------------------------------------------------------- 1 | wandb/* 2 | wandb 3 | saved_models/* 4 | -------------------------------------------------------------------------------- /assets/intro.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/wwxu21/CUT/HEAD/assets/intro.png -------------------------------------------------------------------------------- /assets/toy.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/wwxu21/CUT/HEAD/assets/toy.png -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | accelerate==0.23.0 2 | alpaca-eval==0.3.0 3 | anthropic==0.3.11 4 | bitsandbytes==0.41.1 5 | datasets==2.16.1 6 | evaluate==0.4.0 7 | fire==0.5.0 8 | fschat==0.2.29 9 | huggingface-hub 10 | matplotlib-inline==0.1.6 11 | nltk==3.8.1 12 | numpy==1.25.2 13 | openai==0.27.9 14 | OpenCC==1.1.6 15 | pandas==2.0.3 16 | peft==0.5.0 17 | prompt-toolkit==3.0.39 18 | rouge-score==0.1.2 19 | sacrebleu==1.5.0 20 | safetensors==0.3.3 21 | scikit-learn==1.3.0 22 | scipy==1.11.2 23 | sentence-transformers==2.2.2 24 | sentencepiece==0.1.99 25 | six==1.16.0 26 | tokenizers==0.13.3 27 | torch==2.0.1 28 | torchaudio==2.0.2 29 | torchvision==0.15.2 30 | tqdm==4.66.1 31 | tqdm-multiprocess==0.0.11 32 | transformers==4.33.2 33 | pynvml 34 | protobuf==4.24.3 35 | safetensors==0.3.3 36 | wandb 37 | -------------------------------------------------------------------------------- /templates/alpaca.json: -------------------------------------------------------------------------------- 1 | { 2 | "description": "Template used by Alpaca-LoRA.", 3 | "prompt_input": "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n", 4 | "prompt_no_input": "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Response:\n", 5 | "prompt_no_input_judge": "Below is an instruction that describes a task. Write a response to the instruction and the response should match the corresponding judgment.\n\n### Instruction:\n{instruction}\n\n### Judgment:\n{judgment}\n\n### Response:\n", 6 | "response_split": "### Response:" 7 | } -------------------------------------------------------------------------------- /fine-tuning-unlike.sh: -------------------------------------------------------------------------------- 1 | 2 | export PATH=/root/miniconda3/envs/NegInstruct/bin:$PATH 3 | threshold=1.1 4 | weight_unlike=1 5 | name=cut-1plus-13b 6 | CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 torchrun --nproc_per_node=8 --master_port=1233 finetune_unlikelihood.py \ 7 | --base_model saved_models/llama2-13b-chat-hf \ 8 | --data-path data/iter/train-alpaca-sample-iter1.json \ 9 | --output_dir ./saved_models/lora/${name} \ 10 | --batch_size 8 \ 11 | --micro_batch_size 1 \ 12 | --num_epochs 1 \ 13 | --learning_rate 0.0004 \ 14 | --cutoff_len 2048 \ 15 | --val_set_size 0 \ 16 | --lora_r 16 \ 17 | --lora_alpha 16 \ 18 | --lora_dropout 0.05 \ 19 | --lora_target_modules '[gate_proj, down_proj, up_proj]' \ 20 | --train_on_inputs False \ 21 | --add_eos_token False \ 22 | --group_by_length False \ 23 | --prompt_template_name alpaca \ 24 | --lr_scheduler 'cosine' \ 25 | --warmup_steps 100\ 26 | --weight_unlike ${weight_unlike}\ 27 | --threshold ${threshold}\ 28 | --downsample 0.25\ 29 | 30 | CUDA_VISIBLE_DEVICES=0 python merge.py \ 31 | --base_model_name_or_path saved_models/llama2-13b-chat-hf \ 32 | --peft_model_path ./saved_models/lora/${name} \ 33 | --output_dir ./saved_models/${name} 34 | -------------------------------------------------------------------------------- /merge.py: -------------------------------------------------------------------------------- 1 | from transformers import AutoModelForCausalLM, AutoTokenizer 2 | from peft import PeftModel 3 | import torch 4 | 5 | import os 6 | import argparse 7 | 8 | def get_args(): 9 | parser = argparse.ArgumentParser() 10 | parser.add_argument("--base_model_name_or_path", type=str) 11 | parser.add_argument("--peft_model_path", type=str) 12 | parser.add_argument("--output_dir", type=str) 13 | parser.add_argument("--device", type=str, default="auto") 14 | 15 | return parser.parse_args() 16 | 17 | def main(): 18 | args = get_args() 19 | 20 | if args.device == 'auto': 21 | device_arg = { 'device_map': 'auto' } 22 | else: 23 | device_arg = { 'device_map': { "": args.device} } 24 | 25 | print(f"Loading base model: {args.base_model_name_or_path}") 26 | base_model = AutoModelForCausalLM.from_pretrained( 27 | args.base_model_name_or_path, 28 | return_dict=True, 29 | torch_dtype=torch.float16, 30 | **device_arg 31 | ) 32 | 33 | print(f"Loading PEFT: {args.peft_model_path}") 34 | model = PeftModel.from_pretrained(base_model, args.peft_model_path, **device_arg) 35 | # print(model.state_dict()) 36 | print(f"Running merge_and_unload") 37 | model = model.merge_and_unload() 38 | 39 | tokenizer = AutoTokenizer.from_pretrained(args.base_model_name_or_path) 40 | 41 | model.save_pretrained(f"{args.output_dir}") 42 | tokenizer.save_pretrained(f"{args.output_dir}") 43 | print(f"Model saved to {args.output_dir}") 44 | 45 | if __name__ == "__main__" : 46 | main() -------------------------------------------------------------------------------- /prompter.py: -------------------------------------------------------------------------------- 1 | """ 2 | A dedicated helper to manage templates and prompt building. 3 | """ 4 | 5 | import json 6 | import os.path as osp 7 | from typing import Union 8 | 9 | 10 | class Prompter(object): 11 | __slots__ = ("template", "_verbose") 12 | 13 | def __init__(self, template_name: str = "", verbose: bool = False): 14 | self._verbose = verbose 15 | if not template_name: 16 | # Enforce the default here, so the constructor can be called with '' and will not break. 17 | template_name = "alpaca" 18 | file_name = osp.join("templates", f"{template_name}.json") 19 | if not osp.exists(file_name): 20 | raise ValueError(f"Can't read {file_name}") 21 | with open(file_name) as fp: 22 | self.template = json.load(fp) 23 | if self._verbose: 24 | print( 25 | f"Using prompt template {template_name}: {self.template['description']}" 26 | ) 27 | 28 | def generate_prompt( 29 | self, 30 | data_point, 31 | output=False, 32 | ) -> str: 33 | # returns the full prompt from instruction and optional input 34 | # if a label (=response, =output) is provided, it's also appended. 35 | instruction = data_point['instruction'] 36 | label = data_point['output'] 37 | res = instruction 38 | if output: 39 | res = f"{res}{label}" 40 | if self._verbose: 41 | print(res) 42 | return res 43 | 44 | def get_response(self, output: str) -> str: 45 | return output.split(self.template["response_split"])[1].strip() -------------------------------------------------------------------------------- /scripts/generate.py: -------------------------------------------------------------------------------- 1 | from transformers import AutoModelForCausalLM, AutoTokenizer 2 | import torch 3 | import datasets 4 | import os, json 5 | import argparse 6 | from tqdm import tqdm 7 | 8 | def get_args(): 9 | parser = argparse.ArgumentParser() 10 | parser.add_argument("--base_model_name_or_path", type=str) 11 | parser.add_argument("--template", type=str, default="../templates/alpaca.json") 12 | parser.add_argument("--device", type=str, default="auto") 13 | parser.add_argument("--verbose", type=bool, default=False) 14 | 15 | 16 | return parser.parse_args() 17 | 18 | 19 | def evaluate(prompt, model, tokenizer, spliter, device='cuda'): 20 | input_ids = tokenizer.encode(prompt, return_tensors="pt").to(device) 21 | generation_output = model.generate(input_ids=input_ids, num_beams=1, num_return_sequences=1, 22 | max_new_tokens=2048, temperature=0., top_p=1) 23 | output = tokenizer.decode(generation_output[0], skip_special_tokens=True) 24 | output = output.split(spliter, 1)[-1] 25 | return output 26 | 27 | 28 | def main(): 29 | args = get_args() 30 | 31 | if args.device == 'auto': 32 | device_arg = {'device_map': 'auto'} 33 | else: 34 | device_arg = {'device_map': {"": args.device}} 35 | 36 | print(f"Loading base model: {args.base_model_name_or_path}") 37 | model = AutoModelForCausalLM.from_pretrained(args.base_model_name_or_path, 38 | return_dict=True, 39 | torch_dtype=torch.float16, 40 | **device_arg) 41 | tokenizer = AutoTokenizer.from_pretrained(args.base_model_name_or_path) 42 | generator = args.base_model_name_or_path.split("/")[-1] 43 | with open(args.template) as fp: 44 | template = json.load(fp) 45 | 46 | eval_set = datasets.load_dataset("tatsu-lab/alpaca_eval", 47 | "alpaca_eval")["eval"] 48 | save_data = [] 49 | for i_e, example in tqdm(enumerate(eval_set), desc='generating'): 50 | 51 | instruction = example["instruction"] 52 | input_text = template["prompt_no_input"].format( 53 | instruction=instruction) 54 | spliter = template['response_split'] 55 | if "dpo" in args.base_model_name_or_path: 56 | input_text = template["prompt_dpo"].format(instruction=instruction) 57 | spliter = template['dpo_split'] 58 | output = evaluate(input_text, model, tokenizer, spliter) 59 | 60 | if args.verbose: 61 | print("instruction", instruction) 62 | print("output", output) 63 | input() 64 | save_one = { 65 | "instruction": example["instruction"], 66 | "output": output.strip(), 67 | "generator": generator, 68 | 'dataset': example['dataset'], 69 | } 70 | save_data.append(save_one) 71 | 72 | save_file = os.path.join(args.base_model_name_or_path, "alpaca_eval.json") 73 | with open(save_file, 'w') as writer: 74 | json.dump(save_data, writer, ensure_ascii=False, indent=2) 75 | 76 | 77 | if __name__ == "__main__" : 78 | main() 79 | -------------------------------------------------------------------------------- /scripts/convert2alpaca.py: -------------------------------------------------------------------------------- 1 | import json 2 | import random 3 | import os 4 | random.seed(42) 5 | 6 | def process_open_summary(addr, addr_save): 7 | f = [] 8 | count = 0 9 | with open(addr) as reader: 10 | for line in reader: 11 | x = json.loads(line) 12 | x = x['data'] 13 | x_passage = x['passage']['text'] 14 | 15 | for question in x['questions']: 16 | x_question = question['question'] 17 | 18 | x_instruction = f'{x_passage}\n{x_question}' 19 | rankings = question['rankings'] 20 | max_index = rankings.index(max(rankings)) 21 | ans = question['answers'][max_index] 22 | x_out = ans['answer'] 23 | if ('skip' in ans['feedback'] and ans['feedback']['skip'] == True) or x_out == "": 24 | continue 25 | 26 | cirtique = ans['feedback']['critiques'] 27 | count += 1 28 | if cirtique is None or cirtique == []: 29 | assert ans['feedback']['rating'] == 7 30 | x_judgment = None 31 | x_i_ans = None 32 | x_score = None 33 | new_x = { 34 | 'output':x_out, 35 | 'input':None, 36 | 'instruction':x_instruction, 37 | 'reason':None, 38 | 'judgment': x_judgment, 39 | 'i_ans':x_i_ans, 40 | 'score':x_score, 41 | } 42 | 43 | f.append(new_x) 44 | else: 45 | # cirtique = cirtique[0] 46 | x_judgment = "" 47 | for critique_one in cirtique: 48 | x_judgment = x_judgment + " " + critique_one['text'] 49 | x_i_ans = critique_one['refinement'] 50 | if x_i_ans == "": 51 | continue 52 | x_score = 1 53 | 54 | x_judgment = x_judgment.strip() 55 | new_x = { 56 | 'output':x_out, 57 | 'input':None, 58 | 'instruction':x_instruction, 59 | 'reason':None, 60 | 'judgment': x_judgment, 61 | 'i_ans':x_i_ans, 62 | 'score':x_score, 63 | } 64 | f.append(new_x) 65 | print(count) 66 | with open(addr_save, 'w') as writer: 67 | json.dump(f, writer, ensure_ascii=False, indent=2) 68 | return f 69 | 70 | def process_Shepherd(addr, feedback_addr, addr_save): 71 | f = [] 72 | data = [] 73 | feedback = [] 74 | with open(addr) as reader: 75 | for line in reader: 76 | x = json.loads(line) 77 | data.append(x) 78 | with open(feedback_addr) as reader: 79 | for line in reader: 80 | x = json.loads(line) 81 | feedback.append(x) 82 | 83 | for i in range(len(data)): 84 | data_one = data[i] 85 | feedback_one = feedback[i] 86 | x_out = feedback_one['answer'] 87 | x_instruction = feedback_one['question'] 88 | x_judgment = feedback_one['feedback'] 89 | x_socre = 1 90 | x_i_ans = data_one['metadata']['output_correct'] 91 | assert data_one['metadata']['context'] == feedback_one['question'] 92 | assert data_one['metadata']['output_candidate'] == feedback_one['answer'] 93 | x_ref = None 94 | new_x = { 95 | 'output':x_out, 96 | 'input':None, 97 | 'instruction':x_instruction, 98 | 'reason':None, 99 | 'judgment': x_judgment, 100 | 'i_ans':x_i_ans, 101 | 'score':x_socre, 102 | "ref":x_ref, 103 | } 104 | f.append(new_x) 105 | with open(addr_save, 'w') as writer: 106 | json.dump(f, writer, ensure_ascii=False, indent=2) 107 | return f 108 | 109 | def combine_ians(addr1, addr2): 110 | with open(addr1, ) as reader: 111 | f1 = json.load(reader) 112 | with open(addr2, ) as reader: 113 | f2 = json.load(reader) 114 | new_f = [] 115 | for i in range(len(f1)): 116 | new_one = f2[i] 117 | new_one['i_ans'] = f1[i]['i_ans'] 118 | new_f.append(new_one) 119 | with open(addr2, 'w') as writer: 120 | json.dump(new_f, writer, ensure_ascii=False, indent=2) 121 | return new_f 122 | 123 | if __name__ == "__main__": 124 | data_file = "data/open_summary/critiques.train.jsonl" 125 | save_file = "data/open_summary/train-alpaca.json" 126 | if not os.path.exists("data/open_summary"): 127 | os.mkdir("data/open_summary") 128 | process_open_summary(data_file, save_file) 129 | data_file = "data/open_summary/critiques.test.jsonl" 130 | save_file = "data/open_summary/test-alpaca.json" 131 | process_open_summary(data_file, save_file) 132 | data_file = "data/Shepherd/human_data_raw.jsonl" 133 | feedback_file = "data/Shepherd/human_data_for_model.jsonl" 134 | save_file = "data/Shepherd/train-alpaca.json" 135 | if not os.path.exists("data/Shepherd"): 136 | os.mkdir("data/Shepherd") 137 | f = process_Shepherd(data_file, feedback_file, save_file) -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Reasons to Reject? Aligning Language Models with Judgments 2 | 3 | This repository contains code and resources of our paper, 4 | 5 | [Reasons to Reject? Aligning Language Models with Judgments](https://arxiv.org/abs/2312.14591). 6 | 7 | Weiwen Xu, Deng Cai, Zhisong Zhang, Wai Lam, Shuming Shi 8 | 9 | 10 | 11 | ### Catalogue: 12 | * 1. Introduction 13 | * 2. Dataset 14 | * 3. Fine-tuning 15 | * 4. Inference 16 | * 5. Testing 17 | **** 18 | 19 | 20 | 21 | #### 1. Introduction 22 | 23 | As humans, we consistently engage in interactions with our peers and receive feedback in the form of natural language. This language feedback allows us to reflect on our actions, maintain appropriate behavior, and rectify our errors. The question arises naturally: can we use language feedback to align large language models (LLMs)? 24 | 25 | intro 26 | 27 | In contrast to previous research that aligns LLMs with reward or preference data, we present the first systematic exploration of alignment through the lens of language feedback (i.e., judgment). We commence with an in-depth investigation of potential methods that can be adapted for aligning LLMs with judgments, revealing that these methods are unable to fully capitalize on the judgments. To facilitate more effective utilization of judgments, we propose a novel framework, Contrastive Unlikelihood Training (CUT), that allows for fine-grained inappropriate content detection and correction based on judgments. 28 | 29 | 30 |

31 | CUT 32 |

33 | 34 | 35 | Our offline alignment results show that, with merely 1317 off-the-shelf judgment data, CUT (LLaMA2-13b) can beat the 175B DaVinci003 and surpass the best baseline by 52.34 points on AlpacaEval. The online alignment results demonstrate that CUT can align LLMs (LLaMA2-chat-13b) in an iterative fashion using model-specific judgment data, with a steady performance improvement from 81.09 to 91.36 points on AlpacaEval. Our analysis further suggests that judgments exhibit greater potential than rewards for LLM alignment and warrant future research. 36 | 37 | 38 | 39 | #### 2. Dataset 40 | 41 | ##### 2.1. Offline Alignment 42 | 43 | To reproduce the offline experiments, please use the datasets from [Summarization Train](https://openaipublic.blob.core.windows.net/critiques/dataset/critiques/train.jsonl.gz), [Summarization Test](https://openaipublic.blob.core.windows.net/critiques/dataset/critiques/test.jsonl.gz), and [Shepherd](https://github.com/facebookresearch/Shepherd). 44 | Please use the script `scripts/convert2alpaca.py` to convert the data into the Alpaca Format. 45 | 46 | 47 | ##### 2.2. Online Alignment 48 | 49 | To reproduce the online experiments, we provide the training instances for 5 online interations in `data/iter`. 50 | 51 | ##### 2.3. Judgment v.s. Rewards 52 | 53 | We sample 1000 * 4 instruction-response-judgment triplets from [UltraFeedback](https://github.com/OpenBMB/UltraFeedback) and re-annotate them with only negative judgments. The new judgment data can be found in `data/UltraFeedback`. 54 | 55 | 56 | 57 | #### 3. Fine-tuning 58 | 59 | ##### 3.1. Prepare the environment 60 | 61 | ```bash 62 | pip install -r requirments.txt 63 | ``` 64 | 65 | ##### 3.2. Train LLMs with CUT 66 | 67 | ###### 3.2.1. Online Alignment (the first online iteration as an example) 68 | 69 | ```bash 70 | threshold=1.1 71 | weight_unlike=1 72 | name=cut-1plus-13b 73 | CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 torchrun --nproc_per_node=8 --master_port=1233 finetune_unlikelihood.py \ 74 | --base_model saved_models/llama2-13b-chat-hf \ 75 | --data-path data/iter/train-alpaca-sample-iter1.json \ 76 | --output_dir ./saved_models/lora/${name} \ 77 | --batch_size 8 \ 78 | --micro_batch_size 1 \ 79 | --num_epochs 1 \ 80 | --learning_rate 0.0004 \ 81 | --cutoff_len 2048 \ 82 | --val_set_size 0 \ 83 | --lora_r 16 \ 84 | --lora_alpha 16 \ 85 | --lora_dropout 0.05 \ 86 | --lora_target_modules '[gate_proj, down_proj, up_proj]' \ 87 | --train_on_inputs False \ 88 | --add_eos_token False \ 89 | --group_by_length False \ 90 | --prompt_template_name alpaca \ 91 | --lr_scheduler 'cosine' \ 92 | --warmup_steps 100\ 93 | --weight_unlike ${weight_unlike}\ 94 | --threshold ${threshold}\ 95 | --downsample 0.25\ 96 | 97 | CUDA_VISIBLE_DEVICES=0 python merge.py \ 98 | --base_model_name_or_path saved_models/llama2-13b-chat-hf \ 99 | --peft_model_path ./saved_models/lora/${name} \ 100 | --output_dir ./saved_models/${name} 101 | ``` 102 | 103 | ###### 3.2.2. Offline alignment (Shepherd as an example) 104 | First, get the Shepherd dataset according to Sec. 2.1. Then use the following script: 105 | ```bash 106 | threshold=1.2 107 | weight_unlike=0.5 108 | name=cut-1plus-13b 109 | CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 torchrun --nproc_per_node=8 --master_port=1233 finetune_unlikelihood.py \ 110 | --base_model saved_models/llama2-13b-chat-hf \ 111 | --data-path data/Shepherd/train-alpaca.json \ 112 | --output_dir ./saved_models/lora/${name} \ 113 | --batch_size 8 \ 114 | --micro_batch_size 1 \ 115 | --num_epochs 1 \ 116 | --learning_rate 0.0004 \ 117 | --cutoff_len 2048 \ 118 | --val_set_size 0 \ 119 | --lora_r 16 \ 120 | --lora_alpha 16 \ 121 | --lora_dropout 0.05 \ 122 | --lora_target_modules '[gate_proj, down_proj, up_proj]' \ 123 | --train_on_inputs False \ 124 | --add_eos_token False \ 125 | --group_by_length False \ 126 | --prompt_template_name alpaca \ 127 | --lr_scheduler 'cosine' \ 128 | --warmup_steps 100\ 129 | --weight_unlike ${weight_unlike}\ 130 | --threshold ${threshold}\ 131 | --downsample 0.25\ 132 | 133 | CUDA_VISIBLE_DEVICES=0 python merge.py \ 134 | --base_model_name_or_path saved_models/llama2-13b-chat-hf \ 135 | --peft_model_path ./saved_models/lora/${name} \ 136 | --output_dir ./saved_models/${name} 137 | ``` 138 | 139 | 140 | 141 | #### 4. Inference 142 | 143 | ##### 4.1. Checkpoint Release 144 | 145 | We present our [CUT model](https://huggingface.co/xww033/cut-13b), which has undergone four online iterations and successfully achieved a score of 91.36 points on [AlpacaEval](https://tatsu-lab.github.io/alpaca_eval). 146 | 147 | ##### 4.2. Inference Template 148 | 149 | We follow the inference template used from [Stanford Alpaca](https://github.com/tatsu-lab/stanford_alpaca): 150 | 151 | ``` 152 | Below is an instruction that describes a task. Write a response that appropriately completes the request. 153 | 154 | ### Instruction: 155 | {instruction} 156 | 157 | ### Response: 158 | ``` 159 | 160 | ##### 4.3. CLI 161 | 162 | [Fastchat](https://github.com/lm-sys/FastChat) provides a simple setup for those interested in trying our aligned model. After downloading the [CUT model](https://huggingface.co/xww033/cut-13b) through HuggingFace, clone the Fastchat repository: 163 | 164 | ```bash 165 | git clone https://github.com/lm-sys/FastChat.git 166 | cd FastChat 167 | ``` 168 | 169 | Download the required packages: 170 | 171 | ```bash 172 | pip install --upgrade pip # enable PEP 660 support 173 | pip install -e . 174 | ``` 175 | 176 | Finally, run the following: 177 | 178 | ```bash 179 | python -m fastchat.serve.cli --model-path xww033/cut-13b --conv-template alpaca 180 | ``` 181 | 182 | 183 | 184 | 185 | #### 5. Testing 186 | 187 | ##### 5.1. Generation-based Evaluation 188 | 189 | We evaluate the model on [AlpacaEval](https://tatsu-lab.github.io/alpaca_eval). Please first install the evaluation tool: 190 | 191 | ```bash 192 | pip install alpaca-eval 193 | ``` 194 | 195 | The following script is employed to request the LLM to produce responses to the provided 805 instructions: 196 | 197 | ```bash 198 | python scripts/generate.py --base_model_name_or_path 199 | ``` 200 | 201 | The generated responses would be saved in `/alpaca_eval.json`, which is subsequently submitted for GPT4 evaluation: 202 | 203 | ``` 204 | alpaca_eval --model_outputs /alpaca_eval.json 205 | ``` 206 | 207 | ##### 5.2. Ranking-based Evaluation 208 | 209 | We evaluate the model's performance on ARC, HellaSwag, MMLU and TruthfulQA, utilizing the [LLM Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness). 210 | 211 | ## BibTeX 212 | 213 | ``` 214 | @article{xu2023reasons, 215 | title={Reasons to Reject? Aligning Language Models with Judgments}, 216 | author={Xu, Weiwen and Cai, Deng and Zhang, Zhisong and Lam, Wai and Shi, Shuming}, 217 | journal={arXiv preprint arXiv:2312.14591}, 218 | year={2023} 219 | } 220 | ``` 221 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /modeling_llama_unlikelihood.py: -------------------------------------------------------------------------------- 1 | # coding=utf-8 2 | # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. 3 | # 4 | # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX 5 | # and OPT implementations in this library. It has been modified from its 6 | # original forms to accommodate minor architectural differences compared 7 | # to GPT-NeoX and OPT used by the Meta AI team that trained the model. 8 | # 9 | # Licensed under the Apache License, Version 2.0 (the "License"); 10 | # you may not use this file except in compliance with the License. 11 | # You may obtain a copy of the License at 12 | # 13 | # http://www.apache.org/licenses/LICENSE-2.0 14 | # 15 | # Unless required by applicable law or agreed to in writing, software 16 | # distributed under the License is distributed on an "AS IS" BASIS, 17 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 18 | # See the License for the specific language governing permissions and 19 | # limitations under the License. 20 | """ PyTorch LLaMA model.""" 21 | from typing import List, Optional, Tuple, Union 22 | import torch 23 | import torch.nn.functional as F 24 | import torch.utils.checkpoint 25 | from torch import nn 26 | from torch.nn import CrossEntropyLoss, NLLLoss 27 | from transformers.activations import ACT2FN 28 | from transformers.modeling_outputs import CausalLMOutputWithPast 29 | from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings 30 | from transformers.models.llama.modeling_llama import LLAMA_INPUTS_DOCSTRING, LlamaPreTrainedModel, LlamaModel, LLAMA_START_DOCSTRING 31 | 32 | from peft.peft_model import PeftModel, PeftConfig, _get_batch_size, PeftType 33 | import warnings 34 | logger = logging.get_logger(__name__) 35 | 36 | _CONFIG_FOR_DOC = "LlamaConfig" 37 | 38 | 39 | class LlamaForCausalLM(LlamaPreTrainedModel): 40 | _tied_weights_keys = ["lm_head.weight"] 41 | 42 | def __init__(self, config, threshold): 43 | super().__init__(config) 44 | self.model = LlamaModel(config) 45 | self.vocab_size = config.vocab_size 46 | self.threshold = threshold 47 | self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) 48 | 49 | # Initialize weights and apply final processing 50 | self.post_init() 51 | 52 | def get_input_embeddings(self): 53 | return self.model.embed_tokens 54 | 55 | def set_input_embeddings(self, value): 56 | self.model.embed_tokens = value 57 | 58 | def get_output_embeddings(self): 59 | return self.lm_head 60 | 61 | def set_output_embeddings(self, new_embeddings): 62 | self.lm_head = new_embeddings 63 | 64 | def set_decoder(self, decoder): 65 | self.model = decoder 66 | 67 | def get_decoder(self): 68 | return self.model 69 | 70 | @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING) 71 | @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) 72 | def forward( 73 | self, 74 | input_ids: torch.LongTensor = None, 75 | attention_mask: Optional[torch.Tensor] = None, 76 | position_ids: Optional[torch.LongTensor] = None, 77 | weight_like: Optional[torch.Tensor] = None, 78 | weight_unlike: Optional[torch.Tensor] = None, 79 | past_key_values: Optional[List[torch.FloatTensor]] = None, 80 | inputs_embeds: Optional[torch.FloatTensor] = None, 81 | labels: Optional[torch.LongTensor] = None, 82 | use_cache: Optional[bool] = None, 83 | input_ids_neg=None, 84 | attention_mask_neg=None, 85 | labels_neg=None, 86 | output_attentions: Optional[bool] = None, 87 | output_hidden_states: Optional[bool] = None, 88 | return_dict: Optional[bool] = None, 89 | ) -> Union[Tuple, CausalLMOutputWithPast]: 90 | r""" 91 | Args: 92 | labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): 93 | Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., 94 | config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored 95 | (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. 96 | 97 | Returns: 98 | 99 | Example: 100 | 101 | ```python 102 | >>> from transformers import AutoTokenizer, LlamaForCausalLM 103 | 104 | >>> model = LlamaForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS) 105 | >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER) 106 | 107 | >>> prompt = "Hey, are you conscious? Can you talk to me?" 108 | >>> inputs = tokenizer(prompt, return_tensors="pt") 109 | 110 | >>> # Generate 111 | >>> generate_ids = model.generate(inputs.input_ids, max_length=30) 112 | >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] 113 | "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you." 114 | ```""" 115 | output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions 116 | output_hidden_states = ( 117 | output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states 118 | ) 119 | return_dict = return_dict if return_dict is not None else self.config.use_return_dict 120 | # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) 121 | outputs = self.model( 122 | input_ids=input_ids, 123 | attention_mask=attention_mask, 124 | position_ids=position_ids, 125 | past_key_values=past_key_values, 126 | inputs_embeds=inputs_embeds, 127 | use_cache=use_cache, 128 | output_attentions=output_attentions, 129 | output_hidden_states=output_hidden_states, 130 | return_dict=return_dict, 131 | ) 132 | hidden_states = outputs[0] 133 | if self.config.pretraining_tp > 1: 134 | lm_head_slices = self.lm_head.weight.split(self.vocab_size // self.config.pretraining_tp, dim=0) 135 | logits = [F.linear(hidden_states, lm_head_slices[i]) for i in range(self.config.pretraining_tp)] 136 | logits = torch.cat(logits, dim=-1) 137 | else: 138 | logits = self.lm_head(hidden_states) 139 | logits = logits.float() 140 | probs = torch.softmax(logits,dim=2) 141 | batch_size2, seq_length, hidden_size = probs.size() 142 | batch_size = batch_size2 // 2 143 | 144 | loss = None 145 | unlike_mask = weight_unlike.ne(-1).view(-1).to(probs.device) 146 | if labels is not None: 147 | # Shift so that tokens < n predict n 148 | shift_probs_pos = probs[:batch_size][..., :-1, :].contiguous() 149 | shift_labels = labels[..., 1:].contiguous() 150 | # Flatten the tokens 151 | loss_fct = NLLLoss() 152 | shift_probs_pos = shift_probs_pos.view(-1, self.config.vocab_size) 153 | shift_logits = torch.log(shift_probs_pos) 154 | shift_labels = shift_labels.view(-1) 155 | # Enable model parallelism 156 | shift_labels = shift_labels.to(shift_logits.device) 157 | loss = loss_fct(shift_logits, shift_labels) 158 | 159 | loss = loss 160 | if unlike_mask.any(): 161 | loss_unlike = self.unlikelihood(probs, labels, labels_neg, weight_unlike, unlike_mask) 162 | loss = (loss_unlike + loss) / 2 163 | 164 | if not return_dict: 165 | output = (logits,) + outputs[1:] 166 | return (loss,) + output if loss is not None else output 167 | 168 | return CausalLMOutputWithPast( 169 | loss=loss, 170 | logits=logits, 171 | past_key_values=outputs.past_key_values, 172 | hidden_states=outputs.hidden_states, 173 | attentions=outputs.attentions, 174 | ) 175 | def unlikelihood(self, probs, labels, labels_neg, weight_unlike, unlike_mask): 176 | labels = labels.to(probs.device) 177 | labels_neg = labels_neg.to(probs.device) 178 | weight_unlike = weight_unlike.to(probs.device) 179 | shift_probs = probs[..., :-1, :].contiguous() 180 | shift_labels = labels[..., 1:].contiguous() 181 | shift_labels_neg = labels_neg[..., 1:].contiguous() 182 | valid_indices = shift_labels[unlike_mask] != -100 183 | valid_indices_neg = shift_labels_neg[unlike_mask] != -100 184 | # assert (valid_indices == valid_indices_neg).all() 185 | batch_size2, seq_length, hidden_size = shift_probs.size() 186 | batch_size = batch_size2 // 2 187 | device = probs.device 188 | label_clamped = torch.clamp(shift_labels, min=0, max=hidden_size - 1) 189 | label_clamped_neg = torch.clamp(shift_labels_neg, min=0, max=hidden_size - 1) 190 | rows, cols = torch.meshgrid(torch.arange(batch_size, device=device), torch.arange(seq_length, device=device)) 191 | probs_out = shift_probs[:batch_size][rows, cols, label_clamped][unlike_mask] 192 | probs_out_neg = shift_probs[batch_size:][rows, cols, label_clamped_neg][unlike_mask] 193 | valid_prob = probs_out[valid_indices] 194 | valid_prob_neg = probs_out_neg[valid_indices_neg] 195 | scale = (valid_prob / valid_prob_neg).detach() 196 | unlike_indices = scale > self.threshold # give some margins 197 | valid_prob_neg[unlike_indices] = 1 - valid_prob_neg[unlike_indices] 198 | valid_prob_neg[valid_prob_neg == 0] += 1e-5 # avoid 0 199 | valid_lprob_neg = torch.log(valid_prob_neg) 200 | valid_lprob_neg[unlike_indices] = weight_unlike[unlike_mask][0][0] * valid_lprob_neg[unlike_indices] 201 | valid_lprob_neg[~unlike_indices] = valid_lprob_neg[~unlike_indices] 202 | loss_unlike = -torch.sum(valid_lprob_neg)/ valid_lprob_neg.size(0) 203 | return loss_unlike 204 | 205 | 206 | def prepare_inputs_for_generation( 207 | self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs 208 | ): 209 | if past_key_values: 210 | input_ids = input_ids[:, -1:] 211 | 212 | position_ids = kwargs.get("position_ids", None) 213 | if attention_mask is not None and position_ids is None: 214 | # create position_ids on the fly for batch generation 215 | position_ids = attention_mask.long().cumsum(-1) - 1 216 | position_ids.masked_fill_(attention_mask == 0, 1) 217 | if past_key_values: 218 | position_ids = position_ids[:, -1].unsqueeze(-1) 219 | 220 | # if `inputs_embeds` are passed, we only want to use them in the 1st generation step 221 | if inputs_embeds is not None and past_key_values is None: 222 | model_inputs = {"inputs_embeds": inputs_embeds} 223 | else: 224 | model_inputs = {"input_ids": input_ids} 225 | 226 | model_inputs.update( 227 | { 228 | "position_ids": position_ids, 229 | "past_key_values": past_key_values, 230 | "use_cache": kwargs.get("use_cache"), 231 | "attention_mask": attention_mask, 232 | } 233 | ) 234 | return model_inputs 235 | 236 | @staticmethod 237 | def _reorder_cache(past_key_values, beam_idx): 238 | reordered_past = () 239 | for layer_past in past_key_values: 240 | reordered_past += ( 241 | tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past), 242 | ) 243 | return reordered_past 244 | 245 | class PeftModelForCausalLM(PeftModel): 246 | """ 247 | Peft model for causal language modeling. 248 | 249 | Args: 250 | model ([`~transformers.PreTrainedModel`]): Base transformer model. 251 | peft_config ([`PeftConfig`]): Peft config. 252 | 253 | 254 | Example: 255 | 256 | ```py 257 | >>> from transformers import AutoModelForCausalLM 258 | >>> from peft import PeftModelForCausalLM, get_peft_config 259 | 260 | >>> config = { 261 | ... "peft_type": "PREFIX_TUNING", 262 | ... "task_type": "CAUSAL_LM", 263 | ... "inference_mode": False, 264 | ... "num_virtual_tokens": 20, 265 | ... "token_dim": 1280, 266 | ... "num_transformer_submodules": 1, 267 | ... "num_attention_heads": 20, 268 | ... "num_layers": 36, 269 | ... "encoder_hidden_size": 1280, 270 | ... "prefix_projection": False, 271 | ... "postprocess_past_key_value_function": None, 272 | ... } 273 | 274 | >>> peft_config = get_peft_config(config) 275 | >>> model = AutoModelForCausalLM.from_pretrained("gpt2-large") 276 | >>> peft_model = PeftModelForCausalLM(model, peft_config) 277 | >>> peft_model.print_trainable_parameters() 278 | trainable params: 1843200 || all params: 775873280 || trainable%: 0.23756456724479544 279 | ``` 280 | """ 281 | 282 | def __init__(self, model, peft_config: PeftConfig, adapter_name="default"): 283 | super().__init__(model, peft_config, adapter_name) 284 | self.base_model_prepare_inputs_for_generation = self.base_model.prepare_inputs_for_generation 285 | 286 | def forward( 287 | self, 288 | input_ids=None, 289 | attention_mask=None, 290 | inputs_embeds=None, 291 | labels=None, 292 | input_ids_neg=None, 293 | attention_mask_neg=None, 294 | labels_neg=None, 295 | output_attentions=None, 296 | output_hidden_states=None, 297 | return_dict=None, 298 | weight_like=None, 299 | weight_unlike=None, 300 | **kwargs, 301 | ): 302 | peft_config = self.active_peft_config 303 | kwargs.update({'weight_like':weight_like, 'weight_unlike':weight_unlike, "labels_neg": labels_neg}) 304 | input_ids = torch.cat([input_ids, input_ids_neg], dim=0) 305 | attention_mask = torch.cat([attention_mask, attention_mask_neg], dim=0) 306 | if not peft_config.is_prompt_learning: 307 | if self.base_model.config.model_type == "mpt": 308 | if inputs_embeds is not None: 309 | raise AssertionError("forward in MPTForCausalLM does not support inputs_embeds") 310 | return self.base_model( 311 | input_ids=input_ids, 312 | attention_mask=attention_mask, 313 | labels=labels, 314 | output_attentions=output_attentions, 315 | output_hidden_states=output_hidden_states, 316 | return_dict=return_dict, 317 | **kwargs, 318 | ) 319 | 320 | return self.base_model( 321 | input_ids=input_ids, 322 | attention_mask=attention_mask, 323 | inputs_embeds=inputs_embeds, 324 | labels=labels, 325 | output_attentions=output_attentions, 326 | output_hidden_states=output_hidden_states, 327 | return_dict=return_dict, 328 | **kwargs, 329 | ) 330 | batch_size = _get_batch_size(input_ids, inputs_embeds) 331 | if attention_mask is not None: 332 | # concat prompt attention mask 333 | prefix_attention_mask = torch.ones(batch_size, peft_config.num_virtual_tokens).to(attention_mask.device) 334 | attention_mask = torch.cat((prefix_attention_mask, attention_mask), dim=1) 335 | 336 | if kwargs.get("position_ids", None) is not None: 337 | warnings.warn("Position ids are not supported for parameter efficient tuning. Ignoring position ids.") 338 | kwargs["position_ids"] = None 339 | if kwargs.get("token_type_ids", None) is not None: 340 | warnings.warn("Token type ids are not supported for parameter efficient tuning. Ignoring token type ids") 341 | kwargs["token_type_ids"] = None 342 | kwargs.update( 343 | { 344 | "attention_mask": attention_mask, 345 | "labels": labels, 346 | "output_attentions": output_attentions, 347 | "output_hidden_states": output_hidden_states, 348 | "return_dict": return_dict, 349 | } 350 | ) 351 | 352 | if peft_config.peft_type == PeftType.PREFIX_TUNING: 353 | past_key_values = self.get_prompt(batch_size) 354 | return self.base_model( 355 | input_ids=input_ids, inputs_embeds=inputs_embeds, past_key_values=past_key_values, **kwargs 356 | ) 357 | else: 358 | if inputs_embeds is None: 359 | inputs_embeds = self.word_embeddings(input_ids) 360 | # concat prompt labels 361 | if labels is not None: 362 | prefix_labels = torch.full((batch_size, peft_config.num_virtual_tokens), -100).to(labels.device) 363 | kwargs["labels"] = torch.cat((prefix_labels, labels), dim=1) 364 | prompts = self.get_prompt(batch_size=batch_size) 365 | prompts = prompts.to(inputs_embeds.dtype) 366 | inputs_embeds = torch.cat((prompts, inputs_embeds), dim=1) 367 | return self.base_model(inputs_embeds=inputs_embeds, **kwargs) 368 | 369 | def generate(self, **kwargs): 370 | self.base_model.prepare_inputs_for_generation = self.prepare_inputs_for_generation 371 | if hasattr(self.base_model, "model"): 372 | self.base_model.model.generation_config = self.generation_config 373 | else: 374 | self.base_model.generation_config = self.generation_config 375 | try: 376 | outputs = self.base_model.generate(**kwargs) 377 | except: 378 | self.base_model.prepare_inputs_for_generation = self.base_model_prepare_inputs_for_generation 379 | raise 380 | else: 381 | self.base_model.prepare_inputs_for_generation = self.base_model_prepare_inputs_for_generation 382 | return outputs 383 | 384 | def prepare_inputs_for_generation(self, *args, **kwargs): 385 | peft_config = self.active_peft_config 386 | model_kwargs = self.base_model_prepare_inputs_for_generation(*args, **kwargs) 387 | if peft_config.is_prompt_learning: 388 | if model_kwargs.get("attention_mask", None) is not None: 389 | prefix_attention_mask = torch.ones( 390 | model_kwargs["input_ids"].shape[0], peft_config.num_virtual_tokens 391 | ).to(model_kwargs["input_ids"].device) 392 | model_kwargs["attention_mask"] = torch.cat( 393 | (prefix_attention_mask, model_kwargs["attention_mask"]), dim=1 394 | ) 395 | 396 | if model_kwargs.get("position_ids", None) is not None: 397 | warnings.warn("Position ids are not supported for parameter efficient tuning. Ignoring position ids.") 398 | model_kwargs["position_ids"] = None 399 | 400 | if kwargs.get("token_type_ids", None) is not None: 401 | warnings.warn( 402 | "Token type ids are not supported for parameter efficient tuning. Ignoring token type ids" 403 | ) 404 | kwargs["token_type_ids"] = None 405 | 406 | if model_kwargs["past_key_values"] is None and peft_config.peft_type == PeftType.PREFIX_TUNING: 407 | past_key_values = self.get_prompt(batch_size=model_kwargs["input_ids"].shape[0]) 408 | model_kwargs["past_key_values"] = past_key_values 409 | else: 410 | if model_kwargs["past_key_values"] is None: 411 | inputs_embeds = self.word_embeddings(model_kwargs["input_ids"]) 412 | prompts = self.get_prompt(batch_size=model_kwargs["input_ids"].shape[0]) 413 | prompts = prompts.to(inputs_embeds.dtype) 414 | model_kwargs["inputs_embeds"] = torch.cat((prompts, inputs_embeds), dim=1) 415 | model_kwargs["input_ids"] = None 416 | 417 | return model_kwargs -------------------------------------------------------------------------------- /finetune_unlikelihood.py: -------------------------------------------------------------------------------- 1 | import os 2 | import sys 3 | from typing import List 4 | import json 5 | import fire 6 | import torch 7 | from torch.utils.data import DataLoader 8 | import transformers 9 | from datasets import load_dataset, concatenate_datasets, Dataset 10 | from transformers import TrainerCallback, TrainingArguments, TrainerState, TrainerControl 11 | from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR 12 | from peft import ( 13 | LoraConfig, 14 | prepare_model_for_int8_training, 15 | set_peft_model_state_dict, 16 | MODEL_TYPE_TO_PEFT_MODEL_MAPPING, 17 | PeftModel, 18 | ) 19 | from peft.utils import _prepare_prompt_learning_config 20 | from transformers.tokenization_utils_base import PreTrainedTokenizerBase 21 | from transformers.utils import PaddingStrategy 22 | from transformers import LlamaTokenizer, LlamaConfig 23 | from modeling_llama_unlikelihood import LlamaForCausalLM, PeftModelForCausalLM 24 | from prompter import Prompter 25 | from typing import Optional, Union, Any 26 | from dataclasses import dataclass 27 | import numpy as np 28 | import random 29 | seed = 42 30 | random.seed(seed) 31 | np.random.seed(seed) 32 | torch.manual_seed(seed) 33 | torch.cuda.manual_seed_all(seed) 34 | 35 | @dataclass 36 | class MyDataCollator: 37 | """ 38 | Data collator that will dynamically pad the inputs received, as well as the labels. 39 | 40 | Args: 41 | tokenizer ([`PreTrainedTokenizer`] or [`PreTrainedTokenizerFast`]): 42 | The tokenizer used for encoding the data. 43 | model ([`PreTrainedModel`]): 44 | The model that is being trained. If set and has the *prepare_decoder_input_ids_from_labels*, use it to 45 | prepare the *decoder_input_ids* 46 | 47 | This is useful when using *label_smoothing* to avoid calculating loss twice. 48 | padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `True`): 49 | Select a strategy to pad the returned sequences (according to the model's padding side and padding index) 50 | among: 51 | 52 | - `True` or `'longest'` (default): Pad to the longest sequence in the batch (or no padding if only a single 53 | sequence is provided). 54 | - `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum 55 | acceptable input length for the model if that argument is not provided. 56 | - `False` or `'do_not_pad'`: No padding (i.e., can output a batch with sequences of different lengths). 57 | max_length (`int`, *optional*): 58 | Maximum length of the returned list and optionally padding length (see above). 59 | pad_to_multiple_of (`int`, *optional*): 60 | If set will pad the sequence to a multiple of the provided value. 61 | 62 | This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 63 | 7.5 (Volta). 64 | label_pad_token_id (`int`, *optional*, defaults to -100): 65 | The id to use when padding the labels (-100 will be automatically ignored by PyTorch loss functions). 66 | return_tensors (`str`): 67 | The type of Tensor to return. Allowable values are "np", "pt" and "tf". 68 | """ 69 | tokenizer: PreTrainedTokenizerBase 70 | model: Optional[Any] = None 71 | padding: Union[bool, str, PaddingStrategy] = True 72 | max_length: Optional[int] = None 73 | pad_to_multiple_of: Optional[int] = None 74 | label_pad_token_id: int = -100 75 | return_tensors: str = "pt" 76 | 77 | def __call__(self, features, return_tensors=None): 78 | 79 | if return_tensors is None: 80 | return_tensors = self.return_tensors 81 | labels = [feature["labels"] for feature in features] if "labels" in features[0].keys() else None 82 | labels_neg = [feature["labels_neg"] for feature in features] if "labels_neg" in features[0].keys() else None 83 | # We have to pad the labels before calling `tokenizer.pad` as this method won't pad them and needs them of the 84 | # same length to return tensors. 85 | if labels is not None: 86 | max_label_length = max(len(l) for l in labels) 87 | if labels_neg is not None: 88 | max_label_length_neg = max(len(l) for l in labels_neg) 89 | max_label_length = max(max_label_length, max_label_length_neg) 90 | if self.pad_to_multiple_of is not None: 91 | max_label_length = ( 92 | (max_label_length + self.pad_to_multiple_of - 1) 93 | // self.pad_to_multiple_of 94 | * self.pad_to_multiple_of 95 | ) 96 | # self.tokenizer.padding_side = "left" 97 | padding_side = self.tokenizer.padding_side 98 | 99 | for feature in features: 100 | feature['weight_like'] = [feature['weight_like']] 101 | feature['weight_unlike'] = [feature['weight_unlike']] 102 | remainder = [self.label_pad_token_id] * (max_label_length - len(feature["labels"])) 103 | remainder_length = max_label_length - len(feature["labels_neg"]) 104 | remainder_label = [self.label_pad_token_id] * remainder_length 105 | remainder_ids = [self.tokenizer.pad_token_id] * remainder_length 106 | remainder_mask = [0] * remainder_length 107 | if isinstance(feature["labels"], list): 108 | feature["labels"] = ( 109 | feature["labels"] + remainder if padding_side == "right" else remainder + feature["labels"] 110 | ) 111 | feature["labels_neg"] = ( 112 | feature["labels_neg"] + remainder_label if padding_side == "right" else remainder_label + feature["labels_neg"] 113 | ) 114 | feature["input_ids_neg"] = ( 115 | feature["input_ids_neg"] + remainder_ids if padding_side == "right" else remainder_ids + feature["input_ids_neg"] 116 | ) 117 | feature["attention_mask_neg"] = ( 118 | feature["attention_mask_neg"] + remainder_mask if padding_side == "right" else remainder_mask + feature["attention_mask_neg"] 119 | ) 120 | elif padding_side == "right": 121 | feature["labels"] = np.concatenate([feature["labels"], remainder]).astype(np.int64) 122 | feature["labels_neg"] = np.concatenate([feature["labels_neg"], remainder_label]).astype(np.int64) 123 | feature["input_ids_neg"] = np.concatenate([feature["input_ids_neg"], remainder_ids]).astype(np.int64) 124 | feature["attention_mask_neg"] = np.concatenate([feature["attention_mask_neg"], remainder_mask]).astype(np.int64) 125 | else: 126 | feature["labels"] = np.concatenate([remainder, feature["labels"]]).astype(np.int64) 127 | feature["labels_neg"] = np.concatenate([remainder_label, feature["labels_neg"]]).astype(np.int64) 128 | feature["input_ids_neg"] = np.concatenate([remainder_ids, feature["input_ids_neg"]]).astype(np.int64) 129 | feature["attention_mask_neg"] = np.concatenate([remainder_mask, feature["attention_mask_neg"]]).astype(np.int64) 130 | 131 | features = self.tokenizer.pad( 132 | features, 133 | padding=self.padding, 134 | max_length=max_label_length, 135 | pad_to_multiple_of=self.pad_to_multiple_of, 136 | return_tensors=return_tensors, 137 | ) 138 | 139 | # prepare decoder_input_ids 140 | if ( 141 | labels is not None 142 | and self.model is not None 143 | and hasattr(self.model, "prepare_decoder_input_ids_from_labels") 144 | ): 145 | decoder_input_ids = self.model.prepare_decoder_input_ids_from_labels(labels=features["labels"]) 146 | features["decoder_input_ids"] = decoder_input_ids 147 | return features 148 | 149 | class SavePeftModelCallback(TrainerCallback): 150 | def on_save( 151 | self, 152 | args: TrainingArguments, 153 | state: TrainerState, 154 | control: TrainerControl, 155 | **kwargs, 156 | ): 157 | checkpoint_folder = os.path.join(args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}") 158 | 159 | kwargs["model"].save_pretrained(checkpoint_folder) 160 | 161 | pytorch_model_path = os.path.join(checkpoint_folder, "pytorch_model.bin") 162 | torch.save({}, pytorch_model_path) 163 | return control 164 | 165 | 166 | class LoadBestPeftModelCallback(TrainerCallback): 167 | def on_train_end( 168 | self, 169 | args: TrainingArguments, 170 | state: TrainerState, 171 | control: TrainerControl, 172 | **kwargs, 173 | ): 174 | print(f"Loading best peft model from {state.best_model_checkpoint} (score: {state.best_metric}).") 175 | best_model_path = os.path.join(state.best_model_checkpoint, "adapter_model.bin") 176 | adapters_weights = torch.load(best_model_path) 177 | model = kwargs["model"] 178 | set_peft_model_state_dict(model, adapters_weights) 179 | return control 180 | 181 | def get_peft_model(model, peft_config, adapter_name: str = "default"): 182 | """ 183 | Returns a Peft model object from a model and a config. 184 | 185 | Args: 186 | model ([`transformers.PreTrainedModel`]): Model to be wrapped. 187 | peft_config ([`PeftConfig`]): Configuration object containing the parameters of the Peft model. 188 | """ 189 | model_config = getattr(model, "config", {"model_type": "custom"}) 190 | if hasattr(model_config, "to_dict"): 191 | model_config = model_config.to_dict() 192 | 193 | peft_config.base_model_name_or_path = model.__dict__.get("name_or_path", None) 194 | 195 | if peft_config.task_type not in MODEL_TYPE_TO_PEFT_MODEL_MAPPING.keys() and not peft_config.is_prompt_learning: 196 | return PeftModel(model, peft_config, adapter_name=adapter_name) 197 | if peft_config.is_prompt_learning: 198 | peft_config = _prepare_prompt_learning_config(peft_config, model_config) 199 | return PeftModelForCausalLM(model, peft_config, adapter_name=adapter_name) 200 | 201 | def train( 202 | # model/data params 203 | base_model: str = "", 204 | data_path: str = "", 205 | output_dir: str = "", 206 | # training hyperparams 207 | batch_size: int = 128, 208 | micro_batch_size: int = 8, 209 | num_epochs: int = 1, 210 | learning_rate: float = 3e-4, 211 | cutoff_len: int = 4096, 212 | val_set_size: int = 0, 213 | lr_scheduler: str = "cosine", 214 | warmup_steps: int = 100, 215 | # lora hyperparams 216 | lora_r: int = 16, 217 | lora_alpha: int = 16, 218 | lora_dropout: float = 0.05, 219 | # from peft docs: ["q_proj", "k_proj", "v_proj", "o_proj", "fc_in", "fc_out", "wte", "gate_proj", "down_proj", "up_proj"] 220 | lora_target_modules: List[str] = ["gate_proj", "down_proj", "up_proj"], 221 | # llm hyperparams 222 | train_on_inputs: bool = False, # if False, masks out inputs in loss 223 | add_eos_token: bool = False, 224 | group_by_length: bool = False, # faster, but produces an odd training loss curve 225 | # wandb params 226 | wandb_project: str = "", 227 | wandb_run_name: str = "", 228 | wandb_watch: str = "", # options: false | gradients | all 229 | wandb_log_model: str = "", # options: false | true 230 | resume_from_checkpoint: str = None, # either training checkpoint or final adapter 231 | prompt_template_name: str = "alpaca", 232 | weight_unlike: float = 0.1, 233 | threshold: float = 1.1, 234 | downsample: float = -1, 235 | debug: bool = False, 236 | ): 237 | if int(os.environ.get("LOCAL_RANK", 0)) == 0: 238 | print( 239 | f"Params using prompt template {prompt_template_name}\n" 240 | f"the unlikelihood weight for the incorrect token in the incorrect response: {weight_unlike}\n" 241 | f"the threshold to determine the unlikelihood token: {threshold}\n" 242 | f"downssample rate for Hindsight-P: {downsample}\n" 243 | f"base_model: {base_model}\n" 244 | f"data_path: {data_path}\n" 245 | f"output_dir: {output_dir}\n" 246 | f"batch_size: {batch_size}\n" 247 | f"micro_batch_size: {micro_batch_size}\n" 248 | f"num_epochs: {num_epochs}\n" 249 | f"learning_rate: {learning_rate}\n" 250 | f"cutoff_len: {cutoff_len}\n" 251 | f"val_set_size: {val_set_size}\n" 252 | f"lr_scheduler: {lr_scheduler}\n" 253 | f"warmup_steps: {warmup_steps}\n" 254 | f"lora_r: {lora_r}\n" 255 | f"lora_alpha: {lora_alpha}\n" 256 | f"lora_dropout: {lora_dropout}\n" 257 | f"lora_target_modules: {lora_target_modules}\n" 258 | f"train_on_inputs: {train_on_inputs}\n" 259 | f"add_eos_token: {add_eos_token}\n" 260 | f"group_by_length: {group_by_length}\n" 261 | f"wandb_project: {wandb_project}\n" 262 | f"wandb_run_name: {wandb_run_name}\n" 263 | f"wandb_watch: {wandb_watch}\n" 264 | f"wandb_log_model: {wandb_log_model}\n" 265 | f"resume_from_checkpoint: {resume_from_checkpoint or False}\n" 266 | ) 267 | assert ( 268 | base_model 269 | ), "Please specify a --base_model, e.g. --base_model='huggyllama/llama-7b'" 270 | gradient_accumulation_steps = batch_size // micro_batch_size 271 | 272 | prompter = Prompter(prompt_template_name) 273 | if not debug: 274 | device_map = "auto" 275 | else: 276 | device_map = "cpu" 277 | world_size = int(os.environ.get("WORLD_SIZE", 1)) 278 | local_rank = int(os.environ.get("LOCAL_RANK", 0)) 279 | ddp = world_size != 1 280 | if ddp: 281 | device_map = {"": int(os.environ.get("LOCAL_RANK") or 0)} 282 | gradient_accumulation_steps = gradient_accumulation_steps // world_size 283 | print("gradient_accumulation_steps: ", gradient_accumulation_steps) 284 | 285 | # Check if parameter passed or if set within environ 286 | use_wandb = len(wandb_project) > 0 or ( 287 | "WANDB_PROJECT" in os.environ and len(os.environ["WANDB_PROJECT"]) > 0 288 | ) 289 | use_wandb =False 290 | # Only overwrite environ if wandb param passed 291 | if len(wandb_project) > 0: 292 | os.environ["WANDB_PROJECT"] = wandb_project 293 | if len(wandb_watch) > 0: 294 | os.environ["WANDB_WATCH"] = wandb_watch 295 | if len(wandb_log_model) > 0: 296 | os.environ["WANDB_LOG_MODEL"] = wandb_log_model 297 | if not debug: 298 | model = LlamaForCausalLM.from_pretrained( 299 | base_model, 300 | load_in_8bit=True, 301 | torch_dtype=torch.float16, 302 | device_map=device_map, 303 | threshold=threshold) 304 | else: 305 | config_llama = LlamaConfig.from_pretrained( 306 | base_model,) 307 | model = LlamaForCausalLM(config_llama,threshold=threshold) 308 | tokenizer = LlamaTokenizer.from_pretrained(base_model) 309 | model = prepare_model_for_int8_training(model) 310 | if not debug: 311 | config = LoraConfig( 312 | r=lora_r, 313 | lora_alpha=lora_alpha, 314 | target_modules=lora_target_modules, 315 | lora_dropout=lora_dropout, 316 | bias="none", 317 | task_type="CAUSAL_LM") 318 | 319 | model = get_peft_model(model, config) 320 | model.print_trainable_parameters() 321 | bos = tokenizer.bos_token_id 322 | eos = tokenizer.eos_token_id 323 | pad = tokenizer.pad_token_id 324 | print("pre-trained model's BOS EOS and PAD token id:",bos,eos,pad," => It should be 1 2 None") 325 | 326 | tokenizer.pad_token_id = 0 # unk. we want this to be different from the eos token 327 | tokenizer.padding_side = "right" 328 | def pad_token(mode): 329 | if mode == 'input_ids': 330 | return tokenizer.pad_token_id 331 | elif mode == 'attention_mask': 332 | return 0 333 | elif mode == 'labels': 334 | return -100 335 | def tokenize(prompt, add_eos_token=True): 336 | result = tokenizer( 337 | prompt, 338 | truncation=False, 339 | padding=False, 340 | return_tensors=None, 341 | ) 342 | if len(result["input_ids"]) > cutoff_len: # truncate from left side to keep the response complete 343 | n_overflow = len(result["input_ids"]) - cutoff_len 344 | result["input_ids"] = result["input_ids"][-cutoff_len:] 345 | result["attention_mask"] = result["attention_mask"][-cutoff_len:] 346 | else: 347 | n_overflow = 0 348 | if ( 349 | result["input_ids"][-1] != tokenizer.eos_token_id 350 | and add_eos_token 351 | ): 352 | result["input_ids"].append(tokenizer.eos_token_id) 353 | result["attention_mask"].append(1) 354 | 355 | result["labels"] = result["input_ids"].copy() 356 | result["n_overflow"] = n_overflow 357 | return result, n_overflow 358 | 359 | def generate_and_tokenize_prompt(data_point): 360 | instructions = data_point['instruction_list'] 361 | tokenized_full_prompt_list = [] 362 | for i_i, instruction in enumerate(instructions): 363 | data_point['instruction'] = instruction 364 | full_prompt = prompter.generate_prompt( 365 | data_point, output=True) 366 | 367 | tokenized_full_prompt, n_overflow_full = tokenize(full_prompt) 368 | if not train_on_inputs: 369 | user_prompt = prompter.generate_prompt( 370 | data_point, output=False) 371 | 372 | tokenized_user_prompt, n_overflow_user = tokenize( 373 | user_prompt, add_eos_token=add_eos_token) 374 | 375 | user_prompt_len = len(tokenized_user_prompt["input_ids"]) 376 | offset = n_overflow_full - n_overflow_user 377 | user_prompt_len = user_prompt_len - offset 378 | if add_eos_token: 379 | user_prompt_len -= 1 380 | if user_prompt_len > 0: 381 | tokenized_full_prompt["labels"] = [ 382 | -100 383 | ] * user_prompt_len + tokenized_full_prompt["labels"][ 384 | user_prompt_len: 385 | ] # TODO: Speed up? 386 | assert len(tokenized_full_prompt["labels"]) == len(tokenized_full_prompt["input_ids"]) 387 | if i_i == 0: 388 | answer_len = len(tokenized_full_prompt["labels"]) - user_prompt_len 389 | elif i_i == 1: 390 | answer_len2 = len(tokenized_full_prompt["labels"]) - user_prompt_len 391 | assert answer_len == answer_len2 392 | tokenized_full_prompt_list.append(tokenized_full_prompt) 393 | if len(tokenized_full_prompt_list) == 1: 394 | tokenized_full_prompt = tokenized_full_prompt_list[0] 395 | tokenized_full_prompt['input_ids_neg'] = [pad_token('input_ids')] * len(tokenized_full_prompt['input_ids']) 396 | tokenized_full_prompt['attention_mask_neg'] = [pad_token('attention_mask')] * len(tokenized_full_prompt['attention_mask']) 397 | tokenized_full_prompt['labels_neg'] = [pad_token('labels')] * len(tokenized_full_prompt['labels']) 398 | else: 399 | tokenized_full_prompt = tokenized_full_prompt_list[0] 400 | tokenized_full_prompt_neg = tokenized_full_prompt_list[1] 401 | tokenized_full_prompt['input_ids_neg'] = tokenized_full_prompt_neg['input_ids'] 402 | tokenized_full_prompt['attention_mask_neg'] = tokenized_full_prompt_neg['attention_mask'] 403 | tokenized_full_prompt['labels_neg'] = tokenized_full_prompt_neg['labels'] 404 | return tokenized_full_prompt 405 | 406 | 407 | 408 | if data_path.endswith(".json") or data_path.endswith(".jsonl"): 409 | data = load_dataset("json", data_files=data_path) 410 | else: 411 | data = load_dataset(data_path) 412 | 413 | if resume_from_checkpoint: 414 | # Check the available weights and load them 415 | checkpoint_name = os.path.join( 416 | resume_from_checkpoint, "pytorch_model.bin" 417 | ) # Full checkpoint 418 | if not os.path.exists(checkpoint_name): 419 | checkpoint_name = os.path.join( 420 | resume_from_checkpoint, "adapter_model.bin" 421 | ) # only LoRA model - LoRA config above has to fit 422 | resume_from_checkpoint = ( 423 | False # So the trainer won't try loading its state 424 | ) 425 | # The two files above have a different name depending on how they were saved, but are actually the same. 426 | if os.path.exists(checkpoint_name): 427 | print(f"Restarting from {checkpoint_name}") 428 | adapters_weights = torch.load(checkpoint_name) 429 | set_peft_model_state_dict(model, adapters_weights) 430 | else: 431 | print(f"Checkpoint {checkpoint_name} not found") 432 | 433 | 434 | file_name = os.path.join("templates", f"{prompt_template_name}.json") 435 | with open(file_name) as fp: 436 | template = json.load(fp) 437 | 438 | train_processed = [] 439 | n_pos = 0 440 | n_neg = 0 441 | pos_ids = [] 442 | for ix,x in enumerate(data["train"]): 443 | x_judgment = x['judgment'] 444 | x_score = x['score'] if 'score' in x else None 445 | if x_score is not None and x_score >= 7: 446 | x_judgment = None 447 | x_input = x['input'] 448 | x_instruction = x['instruction'] 449 | x_out = x['output'] 450 | x_i_ans = x['i_ans'] if "i_ans" in x else None 451 | if x_input: 452 | x_instruction = f"{x_instruction}\n{x_input}" 453 | if x_judgment is not None: 454 | hindsight_n = template["prompt_no_input_judge"].format(judgment=x_judgment, instruction=x_instruction) 455 | unfaithful = template["prompt_no_input"].format(instruction=x_instruction) 456 | x_new = { 457 | 'output':x_out, 458 | 'input':None, 459 | 'instruction_list':[hindsight_n,unfaithful], 460 | 'i_ans':x_i_ans, 461 | 'score':x_score, 462 | "weight_like":1, 463 | "weight_unlike":weight_unlike, 464 | } 465 | n_neg += 1 466 | train_processed.append(x_new) 467 | else: 468 | hindsight_p = template["prompt_no_input"].format(instruction=x_instruction) 469 | x_new = { 470 | 'output':x_out, 471 | 'input':None, 472 | 'instruction_list':[hindsight_p], 473 | 'i_ans':x_i_ans, 474 | 'score':x_score, 475 | "weight_like":1, 476 | "weight_unlike":-1, 477 | } 478 | n_pos += 1 479 | pos_ids.append(len(train_processed)) 480 | train_processed.append(x_new) 481 | print(f"n_pos:{n_pos}, n_neg:{n_neg}") 482 | if downsample != -1: 483 | n_keep = int(n_neg * downsample) 484 | if n_keep < n_pos: 485 | pos_keep_ids = random.sample(pos_ids, n_keep) 486 | pos_ids = sorted(pos_ids, reverse=True) 487 | for idx in pos_ids: 488 | if idx not in pos_keep_ids: 489 | train_processed.pop(idx) 490 | print(f"after downsampling, the total num of train data is: {len(train_processed)}") 491 | train_processed = Dataset.from_list(train_processed) 492 | print(f"num of training data: {len(train_processed)}") 493 | train_data = train_processed.map(generate_and_tokenize_prompt) 494 | val_data = None 495 | 496 | 497 | if not ddp and torch.cuda.device_count() > 1: 498 | # keeps Trainer from trying its own DataParallelism when more than 1 gpu is available 499 | model.is_parallelizable = True 500 | model.model_parallel = True 501 | 502 | trainer = transformers.Trainer( 503 | model=model, 504 | train_dataset=train_data, 505 | eval_dataset=val_data, 506 | args=transformers.TrainingArguments( 507 | per_device_train_batch_size=micro_batch_size, 508 | gradient_accumulation_steps=gradient_accumulation_steps, 509 | warmup_steps=warmup_steps, 510 | num_train_epochs=num_epochs, 511 | learning_rate=learning_rate, 512 | # dataloader_num_workers=16, 513 | # fp16=True, 514 | bf16=True if not debug else False, 515 | logging_steps=1, 516 | optim="adamw_torch", 517 | evaluation_strategy="steps" if val_set_size > 0 else "no", 518 | save_strategy="steps", 519 | eval_steps=200 if val_set_size > 0 else None, 520 | save_steps=1000, 521 | lr_scheduler_type=lr_scheduler, 522 | output_dir=output_dir, 523 | save_total_limit=2, 524 | load_best_model_at_end=True if val_set_size > 0 else False, 525 | ddp_find_unused_parameters=False if ddp else None, 526 | group_by_length=group_by_length, 527 | report_to="wandb" if use_wandb else None, 528 | run_name=wandb_run_name if use_wandb else None, 529 | ), 530 | data_collator=MyDataCollator( 531 | tokenizer, pad_to_multiple_of=8, return_tensors="pt", padding="max_length" 532 | ), 533 | ) 534 | model.config.use_cache = False 535 | 536 | if torch.__version__ >= "2" and sys.platform != "win32": 537 | model = torch.compile(model) 538 | 539 | trainer.train(resume_from_checkpoint=resume_from_checkpoint) 540 | if local_rank == 0: 541 | model.save_pretrained(output_dir) 542 | # model.base_model.save_pretrained(output_dir) 543 | pytorch_model_path = os.path.join(output_dir, "pytorch_model.bin") 544 | torch.save({}, pytorch_model_path) 545 | tokenizer.save_pretrained(output_dir) 546 | 547 | 548 | if __name__ == "__main__": 549 | torch.cuda.empty_cache() 550 | fire.Fire(train) --------------------------------------------------------------------------------