gpt4all/gpt4all-training/data.py

168 lines
5.9 KiB
Python
Raw Permalink Normal View History

2023-03-25 12:17:48 -04:00
import glob
import torch
2023-03-27 12:29:43 -04:00
from datasets import load_dataset, concatenate_datasets
2023-03-25 12:17:48 -04:00
import os
from torch.utils.data import DataLoader
from transformers import DefaultDataCollator
def tokenize_inputs(config, tokenizer, examples):
max_length = config["max_length"]
2023-04-04 16:53:23 -04:00
# hacky backward compatible
different_eos = tokenizer.eos_token != "</s>"
out = {"labels": [], "input_ids": []}
for prompt, response in zip(examples["prompt"], examples["response"]):
if different_eos:
2023-04-11 23:51:29 -04:00
if response.count("</s> \n") > 0:
response = response.replace("</s> \n", f"{tokenizer.eos_token} \n")
2023-04-04 16:53:23 -04:00
2023-04-08 16:33:51 -04:00
prompt_len = len(tokenizer(prompt + "\n", return_tensors="pt")["input_ids"][0])
2023-04-04 16:53:23 -04:00
# hack if our prompt is super long
2023-04-05 23:03:34 -04:00
# we need to include some labels so we arbitrarily trunacate at max_length // 2
# if the length is too long
if prompt_len >= max_length // 2:
# if prompt is too long, truncate
# but make sure to truncate to at max 1024 tokens
new_len = min(max_length // 2, len(prompt) // 2)
prompt = prompt[:new_len]
# get new prompt length
2023-04-08 16:33:51 -04:00
prompt_len = tokenizer(prompt + "\n", return_tensors="pt", max_length=max_length // 2, truncation=True).input_ids.ne(tokenizer.pad_token_id).sum().item()
2023-04-05 23:03:34 -04:00
assert prompt_len <= max_length // 2, f"prompt length {prompt_len} exceeds max length {max_length}"
2023-04-04 16:53:23 -04:00
input_tokens = tokenizer(prompt + "\n" + response + tokenizer.eos_token,
truncation=True, max_length=max_length, return_tensors="pt")["input_ids"].squeeze()
2023-03-25 12:17:48 -04:00
2023-04-04 16:53:23 -04:00
labels = input_tokens.clone()
2023-04-08 16:33:51 -04:00
labels[:prompt_len] = -100
2023-04-04 16:53:23 -04:00
if len(labels) < max_length:
# pad to max_length with -100
labels = torch.cat([labels, torch.full((max_length - len(labels),), -100)])
2023-04-08 16:33:51 -04:00
assert (labels == -100).sum() < len(labels), f"Labels are all -100, something wrong. prompt length {prompt_len} exceeds max length {max_length}"
2023-04-05 23:03:34 -04:00
if (labels == -100).sum() == len(labels) - 1:
print(prompt)
print(response)
raise
2023-04-04 16:53:23 -04:00
input_tokens = tokenizer.pad({"input_ids": input_tokens}, padding="max_length", max_length=max_length)["input_ids"]
out["labels"].append(labels)
out["input_ids"].append(input_tokens)
2023-03-25 12:17:48 -04:00
out = {k: torch.stack(v) if isinstance(v, list) else v for k, v in out.items()}
return out
def load_data(config, tokenizer):
dataset_path = config["dataset_path"]
if os.path.exists(dataset_path):
if os.path.isdir(dataset_path):
files = glob.glob(os.path.join(dataset_path, "*_clean.jsonl"))
else:
files = [dataset_path]
2023-03-27 12:29:43 -04:00
print(f"Reading files {files}")
2023-03-25 12:17:48 -04:00
dataset = load_dataset("json", data_files=files, split="train")
else:
2023-04-06 21:45:07 -04:00
dataset = load_dataset(dataset_path, split="train")
2023-03-25 12:17:48 -04:00
dataset = dataset.train_test_split(test_size=.05, seed=config["seed"])
train_dataset, val_dataset = dataset["train"], dataset["test"]
if config["streaming"] is False:
kwargs = {"num_proc": config["num_proc"]}
else:
kwargs = {}
# tokenize inputs and return labels and attention mask
train_dataset = train_dataset.map(
lambda ele: tokenize_inputs(config, tokenizer, ele),
batched=True,
remove_columns=["source", "prompt"],
**kwargs
)
val_dataset = val_dataset.map(
2023-04-13 11:16:31 -04:00
lambda ele: tokenize_inputs(config, tokenizer, ele),
2023-03-25 12:17:48 -04:00
batched=True,
remove_columns=["source", "prompt"],
**kwargs
)
train_dataset = train_dataset.with_format("torch")
val_dataset = val_dataset.with_format("torch")
# create dataloader with default data collator since we already have labels
train_dataloader = DataLoader(
train_dataset,
collate_fn=DefaultDataCollator(),
batch_size=config["batch_size"],
)
val_dataloader = DataLoader(
val_dataset,
collate_fn=DefaultDataCollator(),
batch_size=config["batch_size"],
)
return train_dataloader, val_dataloader
2023-04-06 21:45:07 -04:00
def load_data_for_inference(config, tokenizer):
dataset_path = config["dataset_path"]
if os.path.exists(dataset_path):
# check if path is a directory
if os.path.isdir(dataset_path):
files = glob.glob(os.path.join(dataset_path, "*_clean.jsonl"))
else:
files = [dataset_path]
print(f"Reading files {files}")
dataset = load_dataset("json", data_files=files, split="train")
else:
dataset = load_dataset(dataset_path, split="train")
dataset = dataset.train_test_split(test_size=.05, seed=config["seed"])
train_dataset, val_dataset = dataset["train"], dataset["test"]
train_dataset = train_dataset.add_column("index", list(range(len(train_dataset))))
2023-04-07 08:09:31 -04:00
# select first N batches that are divisible by batch_size
# gather is a bit annoying (or the way I'm using it) to get uneven batches as it duplicates data
train_dataset = train_dataset.select(range((len(train_dataset) // config["batch_size"]) * config["batch_size"]))
2023-04-06 21:45:07 -04:00
val_dataset = val_dataset.add_column("index", list(range(len(val_dataset))))
2023-04-07 08:09:31 -04:00
val_dataset = val_dataset.select(range((len(val_dataset) // config["batch_size"]) * config["batch_size"]))
2023-04-06 21:45:07 -04:00
if config["streaming"] is False:
kwargs = {"num_proc": config["num_proc"]}
else:
kwargs = {}
# tokenize inputs and return labels and attention mask
train_dataset = train_dataset.map(
lambda ele: tokenize_inputs(config, tokenizer, ele),
batched=True,
**kwargs
)
val_dataset = val_dataset.map(
lambda ele: tokenize_inputs(config, tokenizer, ele),
batched=True,
**kwargs
)
train_dataset = train_dataset.with_format("torch")
val_dataset = val_dataset.with_format("torch")
return train_dataset, val_dataset