I use the following script to train starcoderbase. I don’t know where there is no right operation, which causes the model reasoning after training to become very slow, and also throws a warning ''you have modified the pretrained model configuration to control generation. this is a deprecated startegy to control generation and will be removed soon..."
The script I refer to is "https://github.com/bigcode-project/starcoder/tree/main/chat"
Does anyone know where I went wrong, thanks
The following is a certain line of my json file:
{"content":"\/\/ Copyright (c) 2019 The ReBitcoin Core developers\n\/\/ Distributed under the MIT software license, see the accompanying\n\/\/ file COPYING or http:\/\/www.opensource.org\/licenses\/mit-license.php.\n\n#ifndef REBITCOIN_UTIL_STRING_H\n#define REBITCOIN_UTIL_STRING_H\n\n#include <string>\n#include <vector>\n\n\/**\n * Join a list of items\n *\n * @param list The list to join\n * @param separator The separator\n * @param unary_op Apply this operator to each item in the list\n *\/\ntemplate <typename T, typename UnaryOp>\nstd::string Join(const std::vector<T>& list, const std::string& separator, UnaryOp unary_op)\n{\n std::string ret;\n for (size_t i = 0; i < list.size(); ++i) {\n if (i > 0) ret += separator;\n ret += unary_op(list.at(i));\n }\n return ret;\n}\n\ninline std::string Join(const std::vector[std::string](std::string)& list, const std::string& separator)\n{\n return Join(list, separator, [](const std::string& i) { return i; });\n}\n\n#endif \/\/ REBITCOIN_UTIL_STRENCODINGS_H\n","avg_line_length":28.0285714286,"max_line_length":92,"alphanum_fraction":0.6850152905,"path":"src\/util\/string.h","size":981}
#!/usr/bin/env python
# coding=utf-8
# Copyright 2023 The BigCode & HuggingFace Inc. teams. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Script to instruction fine-tune causal language models on a Hub dataset
Adapted from huggingface/transformers: https://github.com/huggingface/transformers/blob/main/examples/pytorch/language-modeling/run_clm.py
"""
import logging
import math
import os
import random
import sys
from itertools import chain
import torch
import datasets
import transformers
from datasets import load_dataset
from transformers import (AutoModelForCausalLM, AutoTokenizer, Trainer, default_data_collator, set_seed, pipeline)
from transformers.testing_utils import CaptureLogger
from transformers.trainer_utils import get_last_checkpoint
from transformers import DataCollatorWithPadding
from config import DataArguments, ModelArguments, TrainingArguments
from utils import StarChatArgumentParser, hf_login
logger = logging.getLogger(__name__)
def create_validation_and_test_splits(
raw_dataset, tokenizer, max_sequence_length=1024
):
# Get validation, test, and the remaining train dataset
test_dataset = raw_dataset["test"]
train_dataset = raw_dataset["train"]
# Tokenize the train, validation, and test dataset
tokenized_train_dataset = train_dataset.map(
lambda x: tokenizer(
x["content"], truncation=True, padding="max_length", max_length=max_sequence_length
),
batched=True,
remove_columns=["content"],
)
tokenized_test_dataset = test_dataset.map(
lambda x: tokenizer(
x["content"], truncation=True, padding="max_length", max_length=max_sequence_length
),
batched=True,
remove_columns=["content"],
)
return tokenized_train_dataset, tokenized_test_dataset
def addline(examples):
labels = examples["input_ids"].copy()
examples["labels"] = labels
return examples
def main():
parser = StarChatArgumentParser((ModelArguments, DataArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".yaml"):
# If we pass only one argument to the script and it's the path to a YAML file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_yaml_file(os.path.abspath(sys.argv[1]))
# parse command line args and yaml file
elif len(sys.argv) > 2 and sys.argv[1].endswith(".yaml"):
model_args, data_args, training_args = parser.parse_yaml_and_args(os.path.abspath(sys.argv[1]), sys.argv[2:])
# parse command line args only
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Set seed for reproducibility
set_seed(training_args.seed)
###############
# Setup logging
###############
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%Y-%m-%d %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
datasets.utils.logging.set_verbosity(log_level)
transformers.utils.logging.set_verbosity(log_level)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# Log on each process a small summary
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+ f" distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
logger.info(f"Model parameters {model_args}")
logger.info(f"Data parameters {data_args}")
logger.info(f"Training/evaluation parameters {training_args}")
# Login to HuggingFace Hub if needed
hf_login()
###########################
# Detecting last checkpoint
###########################
last_checkpoint = None
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome."
)
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
###############
# Load datasets
###############
raw_datasets = load_dataset("json", data_files=data_args.dataset_name, split="train")
raw_datasets = raw_datasets.train_test_split(test_size=0.005)
logger.info(
f"Training on the following datasets and their proportions: {[split + ' : ' + str(dset.num_rows) for split, dset in raw_datasets.items()]}"
)
with training_args.main_process_first(desc="Log a few random samples from the raw training set"):
for index in random.sample(range(len(raw_datasets["train"])), 3):
logger.info(f"Sample {index} of the raw training set:\n\n{raw_datasets['train'][index]['content']}")
#####################################
# Load tokenizer and process datasets
#####################################
tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
(train_dataset, eval_dataset) = create_validation_and_test_splits(raw_datasets, tokenizer)
unused_column = ["avg_line_length", "max_line_length", "alphanum_fraction", "path", "size"]
train_dataset = train_dataset.map(addline, batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=unused_column)
eval_dataset = eval_dataset.map(addline, batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=unused_column)
if training_args.do_train:
if data_args.max_train_samples is not None:
max_train_samples = min(len(train_dataset), data_args.max_train_samples)
train_dataset = train_dataset.select(range(max_train_samples))
if training_args.do_eval:
if data_args.max_eval_samples is not None:
max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
eval_dataset = eval_dataset.select(range(max_eval_samples))
#######################
# Load pretrained model
#######################
logger.info("*** Load pretrained model ***")
torch_dtype = (
model_args.torch_dtype if model_args.torch_dtype in ["auto", None] else getattr(torch, model_args.torch_dtype)
)
model = AutoModelForCausalLM.from_pretrained(
model_args.model_name_or_path,
revision=model_args.model_revision,
torch_dtype=torch_dtype,
use_cache=False if training_args.gradient_checkpointing else True,
)
model.resize_token_embeddings(len(tokenizer))
########################
# create a data collator for ?
########################
data_collator = DataCollatorWithPadding(tokenizer=tokenizer, padding="max_length", pad_to_multiple_of=8)
########################
# Initialize the Trainer
########################
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
# Data collator defaults to DataCollatorWithPadding, so we change it
# since we've already chunked our corpus
# data_collator=default_data_collator,
data_collator=data_collator,
)
###############
# Training loop
###############
if training_args.do_train:
logger.info("*** Train ***")
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
train_result = trainer.train(resume_from_checkpoint=checkpoint)
metrics = train_result.metrics
max_train_samples = (
data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)
)
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
##########
# Evaluate
##########
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(eval_dataset)
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
try:
perplexity = math.exp(metrics["eval_loss"])
except OverflowError:
perplexity = float("inf")
metrics["perplexity"] = perplexity
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
#################################
# Create model card & push to Hub
#################################
kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "text-generation"}
if data_args.dataset_name is not None:
kwargs["dataset_tags"] = data_args.dataset_name
# if data_args.dataset_config_name is not None:
# kwargs["dataset_args"] = data_args.dataset_config_name
# kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
# else:
# kwargs["dataset"] = data_args.dataset_name
# kwargs["dataset_args"] = "default"
kwargs["dataset"] = data_args.dataset_name
kwargs["dataset_args"] = "default"
# Store dialogue template so we can load it at deployment time
dialogue_template.save_pretrained(training_args.output_dir)
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.save_model(training_args.output_dir)
trainer.create_model_card(**kwargs)
with training_args.main_process_first(desc="Generate a sample from the model"):
inputs = "static size_t chrtos(char *buf, size_t size, char byte) {"
inputs = tokenizer.encode(inputs, return_tensor="pt").to(training_args.device)
outputs = model.generate(
**inputs,
max_new_tokens=256,
pad_token_id=tokenizer.pad_token_id,
eos_token_id=tokenizer.eos_token_id,
)
logger.info(f"=== SAMPLE OUTPUT ==\n\n{tokenizer.decode(outputs[0], skip_special_tokens=True)}")
if __name__ == "__main__":
main()