training_reward_model.py

import os

import torch
import evaluate
import numpy as np
import torch.nn as nn
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional, Union
from datasets import load_dataset
from peft import LoraConfig, TaskType, get_peft_model
from transformers import (
    AutoConfig,
    AutoModelForSequenceClassification,
    AutoTokenizer,
    LlamaTokenizer,
    HfArgumentParser,
    PreTrainedTokenizerBase,
    Trainer,
    TrainingArguments,
    set_seed,
)
from transformers.utils import PaddingStrategy

DEFAULT_PAD_TOKEN = "[PAD]"
DEFAULT_EOS_TOKEN = "</s>"
DEFAULT_BOS_TOKEN = "</s>"
DEFAULT_UNK_TOKEN = "</s>"


# Define and parse arguments.
@dataclass
class ScriptArguments:
    """
    These arguments vary depending on how many GPUs you have, what their capacity and features are, and what size model you want to train.
    """

    local_rank: Optional[int] = field(default=-1, metadata={"help": "Used for multi-gpu"})
    resume_from_checkpoint: Optional[bool] = field(
        default=False,
        metadata={"help": "If you want to resume training where it left off."},
    )
    deepspeed: Optional[str] = field(
        default=None,
        metadata={
            "help": "Path to deepspeed config if using deepspeed. You may need this if the model that you want to train doesn't fit on a single GPU."
        },
    )
    per_device_train_batch_size: Optional[int] = field(default=4)
    per_device_eval_batch_size: Optional[int] = field(default=1)
    gradient_accumulation_steps: Optional[int] = field(default=1)
    learning_rate: Optional[float] = field(default=2e-5)
    weight_decay: Optional[int] = field(default=0.001)
    seed: Optional[int] = field(default=1103)
    max_length: Optional[int] = field(default=512)
    log_freq: Optional[int] = field(default=1)
    eval_freq: Optional[int] = field(default=500)
    save_freq: Optional[int] = field(default=500)
    save_total_limit: Optional[int] = field(default=3)
    lora_r: Optional[int] = field(default=8)
    lora_alpha: Optional[int] = field(default=32)
    lora_dropout: Optional[float] = field(default=0.1)
    model_name: Optional[str] = field(
        default="decapoda-research/llama-7b-hf",
        metadata={
            "help": "The model that you want to train from the Hugging Face hub or local."
        },
    )
    dataset_name: Optional[str] = field(
        default="./data/comparison_data_v2.json",
        metadata={"help": "The dataset name"},
    )
    bf16: Optional[bool] = field(
        default=True,
        metadata={
            "help": "This essentially cuts the training time in half if you want to sacrifice a little precision and have a supported GPU."
        },
    )
    num_train_epochs: Optional[int] = field(
        default=1,
        metadata={"help": "The number of training epochs for the reward model."},
    )
    train_subset: Optional[int] = field(
        default=0,
        metadata={"help": "The size of the subset of the training data to use"},
    )
    eval_subset: Optional[int] = field(
        default=0,
        metadata={"help": "The size of the subset of the eval data to use"},
    )
    gradient_checkpointing: Optional[bool] = field(
        default=False,
        metadata={"help": "Enables gradient checkpointing."},
    )
    optim: Optional[str] = field(
        default="adamw_hf",
        metadata={"help": "The optimizer to use."},
    )
    lr_scheduler_type: Optional[str] = field(
        default="linear",
        metadata={"help": "The lr scheduler"},
    )
    output_dir: Optional[str] = field(default="./checkpoints/training_reward_model/",
                                      metadata={"help": "n steps to save the model"})


parser = HfArgumentParser(ScriptArguments)
script_args = parser.parse_args_into_dataclasses()[0]

set_seed(script_args.seed)


# We need to define a special data collator that batches the data in our j vs k format.
@dataclass
class RewardDataCollatorWithPadding:
    tokenizer: PreTrainedTokenizerBase
    padding: Union[bool, str, PaddingStrategy] = True
    max_length: Optional[int] = None
    pad_to_multiple_of: Optional[int] = None
    return_tensors: str = "pt"

    def __call__(self, features: List[Dict[str, Any]]) -> Dict[str, Any]:
        features_j = []
        features_k = []
        for feature in features:
            features_j.append(
                {
                    "input_ids": feature["input_ids_j"],
                    "attention_mask": feature["attention_mask_j"],
                }
            )
            features_k.append(
                {
                    "input_ids": feature["input_ids_k"],
                    "attention_mask": feature["attention_mask_k"],
                }
            )
        batch_j = self.tokenizer.pad(
            features_j,
            padding=self.padding,
            max_length=self.max_length,
            pad_to_multiple_of=self.pad_to_multiple_of,
            return_tensors=self.return_tensors,
        )
        batch_k = self.tokenizer.pad(
            features_k,
            padding=self.padding,
            max_length=self.max_length,
            pad_to_multiple_of=self.pad_to_multiple_of,
            return_tensors=self.return_tensors,
        )
        batch = {
            "input_ids_j": batch_j["input_ids"],
            "attention_mask_j": batch_j["attention_mask"],
            "input_ids_k": batch_k["input_ids"],
            "attention_mask_k": batch_k["attention_mask"],
            "return_loss": True,
        }
        return batch


# Turn the dataset into pairs of post + summaries, where text_j is the preferred question + answer and
# text_k is the other. Then tokenize the dataset.
def preprocess_function(examples):
    new_examples = {
        "input_ids_j": [],
        "attention_mask_j": [],
        "input_ids_k": [],
        "attention_mask_k": [],
    }
    for question, response_j, response_k in zip(examples["user_input"], examples["completion_a"],
                                                examples["completion_b"]):
        tokenized_j = tokenizer(question + response_j, truncation=True)
        tokenized_k = tokenizer(question + response_k, truncation=True)

        new_examples["input_ids_j"].append(tokenized_j["input_ids"])
        new_examples["attention_mask_j"].append(tokenized_j["attention_mask"])
        new_examples["input_ids_k"].append(tokenized_k["input_ids"])
        new_examples["attention_mask_k"].append(tokenized_k["attention_mask"])

    return new_examples


def compute_metrics(eval_pred):
    # Define the metric that we'll use for validation.
    accuracy = evaluate.load("accuracy")

    predictions, _ = eval_pred
    # Here, predictions is rewards_j and rewards_k.
    # We want to see how much of the time rewards_j > rewards_k.
    predictions = np.argmax(predictions, axis=0)
    labels = np.zeros(predictions.shape)
    return accuracy.compute(predictions=predictions, references=labels)


class RewardTrainer(Trainer):
    # Define how to compute the reward loss. We use the InstructGPT pairwise logloss: https://arxiv.org/abs/2203.02155
    def compute_loss(self, model, inputs, return_outputs=False):
        rewards_j = model(input_ids=inputs["input_ids_j"], attention_mask=inputs["attention_mask_j"])[0]
        rewards_k = model(input_ids=inputs["input_ids_k"], attention_mask=inputs["attention_mask_k"])[0]
        loss = -nn.functional.logsigmoid(rewards_j - rewards_k).mean()
        if return_outputs:
            return loss, {"rewards_j": rewards_j, "rewards_k": rewards_k}
        return loss


if "decapoda" in script_args.model_name.lower():
    tokenizer = LlamaTokenizer.from_pretrained(script_args.model_name)
    # required for llama
    tokenizer.add_special_tokens(
        {
            "eos_token": DEFAULT_EOS_TOKEN,
            "bos_token": DEFAULT_BOS_TOKEN,
            "unk_token": DEFAULT_UNK_TOKEN,
            "pad_token": DEFAULT_PAD_TOKEN,
        }
    )
else:
    tokenizer = AutoTokenizer.from_pretrained(script_args.model_name)
    if getattr(tokenizer, "pad_token", None) is None:
        tokenizer.pad_token = tokenizer.eos_token

# Load the dataset for tuning the reward model.
data_path = script_args.dataset_name
if data_path.endswith(".json") or data_path.endswith(".jsonl"):
    dataset = load_dataset("json", data_files=data_path, split="train")
else:
    dataset = load_dataset(data_path, split="train")
dataset = dataset.train_test_split(test_size=0.1, seed=script_args.seed)
train_dataset = dataset["train"]
eval_dataset = dataset["test"]

if script_args.train_subset > 0:
    train_dataset = train_dataset.select(range(script_args.train_subset))
if script_args.eval_subset > 0:
    eval_dataset = eval_dataset.select(range(script_args.eval_subset))
# Define the training args. Needs to be done before the model is loaded if you are using deepspeed.
model_name_split = script_args.model_name.split("/")[-1]
output_name = (
    f"{model_name_split}_peft_gpt-4-llm_rm_{script_args.train_subset}_{script_args.learning_rate}"
)

peft_config = LoraConfig(
    task_type=TaskType.SEQ_CLS,
    inference_mode=False,
    r=script_args.lora_r,
    lora_alpha=script_args.lora_alpha,
    lora_dropout=script_args.lora_dropout,
)

training_args = TrainingArguments(
    output_dir=os.path.join(script_args.output_dir, output_name),
    learning_rate=script_args.learning_rate,
    per_device_train_batch_size=script_args.per_device_train_batch_size,
    per_device_eval_batch_size=script_args.per_device_eval_batch_size,
    num_train_epochs=script_args.num_train_epochs,
    weight_decay=script_args.weight_decay,
    evaluation_strategy="steps",
    eval_steps=script_args.eval_freq,
    save_strategy="steps",
    save_steps=script_args.save_freq,
    save_total_limit=script_args.save_total_limit,
    gradient_accumulation_steps=script_args.gradient_accumulation_steps,
    gradient_checkpointing=script_args.gradient_checkpointing,
    deepspeed=script_args.deepspeed,
    local_rank=script_args.local_rank,
    remove_unused_columns=False,
    label_names=[],
    bf16=script_args.bf16,
    logging_strategy="steps",
    logging_steps=script_args.log_freq,
    optim=script_args.optim,
    lr_scheduler_type=script_args.lr_scheduler_type,
)

model = AutoModelForSequenceClassification.from_pretrained(
    script_args.model_name, num_labels=1, torch_dtype=torch.bfloat16
)
model = get_peft_model(model, peft_config)

model.print_trainable_parameters()
model.config.use_cache = script_args.gradient_checkpointing

num_proc = 24  # Can adjust to be higher if you have more processors.
original_columns = train_dataset.column_names
# preprocess the dataset and filter out QAs that are longer than max_length
train_dataset = train_dataset.map(
    preprocess_function, batched=True, num_proc=num_proc, remove_columns=original_columns
)
train_dataset = train_dataset.filter(
    lambda x: len(x["input_ids_j"]) <= script_args.max_length and len(x["input_ids_k"]) <= script_args.max_length)
eval_dataset = eval_dataset.map(
    preprocess_function, batched=True, num_proc=num_proc, remove_columns=original_columns
)
eval_dataset = eval_dataset.filter(
    lambda x: len(x["input_ids_j"]) <= script_args.max_length and len(x["input_ids_k"]) <= script_args.max_length)

# Train the model
trainer = RewardTrainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    compute_metrics=compute_metrics,
    data_collator=RewardDataCollatorWithPadding(tokenizer=tokenizer, max_length=script_args.max_length),
)

trainer.train(script_args.resume_from_checkpoint)

print("Saving last checkpoint of the model")
model.save_pretrained(script_args.output_dir + "peft_last_checkpoint")