h2ogpt-chatbot / src /create_data.py
pseudotensor's picture
Update with h2oGPT hash 1c93f1c26432bacd38ceb1726fe6009f8d240cb3
1bd70cc
raw
history blame
80.1 kB
"""
Dataset creation tools.
Keep to-level imports clean of non-trivial imports for specific tools,
because this file is imported for various purposes
"""
import ast
import concurrent.futures
import contextlib
import hashlib
import json
import os
import shutil
import signal
import sys
import traceback
from concurrent.futures import ProcessPoolExecutor
import psutil
import pytest
import pandas as pd
import numpy as np
from tqdm import tqdm
from utils import flatten_list, remove
def parse_rst_file(filepath):
with open(filepath, 'r') as f:
input_data = f.read()
settings_overrides = {'initial_header_level': 2}
from docutils import core
document = core.publish_doctree(
source=input_data,
source_path=filepath,
settings_overrides=settings_overrides,
)
qa_pairs = []
current_section = None
current_question = ""
current_answer = ""
for node in document.traverse():
if node.__class__.__name__ == 'section':
current_section = ""
elif current_section is not None:
if node.__class__.__name__ == 'Text':
if node.astext()[-1] == "?":
if current_question:
qa_pairs.append((current_question, current_answer))
current_question = node.astext()
current_answer = ""
else:
current_answer += node.astext()
if current_answer:
qa_pairs.append((current_question, current_answer))
return {k: v for k, v in qa_pairs}
def test_scrape_dai_docs():
home = os.path.expanduser('~')
file = os.path.join(home, 'h2oai/docs/faq.rst')
qa_pairs = parse_rst_file(file)
prompt_type = 'human_bot'
from prompter import prompt_types
assert prompt_type in prompt_types
save_thing = [{"instruction": k, "output": v, 'prompt_type': prompt_type} for k, v in qa_pairs.items()]
output_file = "dai_faq.json"
with open(output_file, "wt") as f:
f.write(json.dumps(save_thing, indent=2))
def test_scrape_dai_docs_all():
"""
pytest create_data.py::test_scrape_dai_docs_all
"""
import glob
import nltk
nltk.download('punkt')
dd = {}
np.random.seed(1234)
home = os.path.expanduser('~')
files = list(glob.glob(os.path.join(home, "h2oai/docs/**/*rst")))
np.random.shuffle(files)
val_count = int(0.05 * len(files))
train_files = files[val_count:]
valid_files = files[:val_count]
things = [
("dai_docs.train.json", train_files),
("dai_docs.valid.json", valid_files)
]
for LEN in [100, 200, 500]:
for output_file, ff in things:
if output_file not in dd:
dd[output_file] = []
for f in ff:
with open(f) as input:
blob = input.read()
blob = blob.replace("~~", "")
blob = blob.replace("==", "")
blob = blob.replace("''", "")
blob = blob.replace("--", "")
blob = blob.replace("**", "")
dd[output_file].extend(get_sentences(blob, length=LEN))
for output_file, _ in things:
save_thing = [{"output": k.strip(), 'prompt_type': 'plain'} for k in dd[output_file]]
with open(output_file, "wt") as f:
f.write(json.dumps(save_thing, indent=2))
def get_sentences(blob, length):
"""
break-up input text into sentences and then output list of sentences of about length in size
:param blob:
:param length:
:return:
"""
import nltk
nltk.download('punkt')
from nltk.tokenize import sent_tokenize
sentences = sent_tokenize(blob)
my_sentences = []
my_string = ""
for sentence in sentences:
if len(my_string) + len(sentence) <= length:
if my_string:
my_string += " " + sentence
else:
my_string = sentence
else:
my_sentences.append(my_string)
my_string = ""
return my_sentences or [my_string]
def setup_dai_docs(path=None, dst="working_dir_docs", from_hf=False):
"""
Only supported if have access to source code or HF token for HF spaces and from_hf=True
:param path:
:param dst:
:param from_hf:
:return:
"""
home = os.path.expanduser('~')
if from_hf:
# assumes
from huggingface_hub import hf_hub_download
# True for case when locally already logged in with correct token, so don't have to set key
token = os.getenv('HUGGING_FACE_HUB_TOKEN', True)
path_to_zip_file = hf_hub_download('h2oai/dai_docs', 'dai_docs.zip', token=token, repo_type='dataset')
path = 'h2oai'
import zipfile
with zipfile.ZipFile(path_to_zip_file, 'r') as zip_ref:
zip_ref.extractall(path)
path = os.path.join(path, 'docs/**/*')
if path is None:
if os.path.isdir(os.path.join(home, 'h2oai')):
path = os.path.join(home, "h2oai/docs/**/*")
else:
assert os.path.isdir(os.path.join(home, 'h2oai.superclean')), '%s does not exist' % path
path = os.path.join(home, "h2oai.superclean/docs/**/*")
import glob
files = list(glob.glob(path, recursive=True))
# pandoc can't find include files
remove(dst)
os.makedirs(dst)
# copy full tree, for absolute paths in rst
for fil in files:
if os.path.isfile(fil):
shutil.copy(fil, dst)
# hack for relative path
scorers_dir = os.path.join(dst, 'scorers')
makedirs(scorers_dir)
for fil in glob.glob(os.path.join(dst, '*.frag')):
shutil.copy(fil, scorers_dir)
return dst
def rst_to_outputs(files, min_len=30, max_len=2048 // 2 - 30):
# account for sequence length (context window) including prompt and input and output
# os.system('pandoc -f rst -t plain ./expert_settings/nlp_settings.rst')
import pypandoc
basedir = os.path.abspath(os.getcwd())
outputs = []
for fil in files:
os.chdir(basedir)
os.chdir(os.path.dirname(fil))
fil = os.path.basename(fil)
print("Processing %s" % fil, flush=True)
# out_format can be one of: asciidoc, asciidoctor, beamer, biblatex, bibtex, commonmark, commonmark_x,
# context, csljson, docbook, docbook4, docbook5, docx, dokuwiki,
# dzslides, epub, epub2, epub3, fb2, gfm, haddock, html, html4, html5, icml,
# ipynb, jats, jats_archiving, jats_articleauthoring, jats_publishing, jira,
# json, latex, man,
# markdown, markdown_github, markdown_mmd, markdown_phpextra, markdown_strict,
# mediawiki, ms, muse, native, odt, opendocument, opml, org, pdf, plain, pptx,
# revealjs, rst, rtf, s5, slideous, slidy, tei, texinfo, textile, xwiki, zimwiki
out_format = 'plain'
# avoid extra new lines injected into text
extra_args = ['--wrap=preserve', '--resource path="%s" % dst']
plain_list = []
try:
# valid for expert settings
input_rst = pypandoc.convert_file(fil, 'rst')
input_list = input_rst.split('\n``')
for input_subrst in input_list:
input_plain = pypandoc.convert_text(input_subrst, format='rst', to='plain')
plain_list.append([input_plain, fil])
except Exception as e:
print("file exception: %s %s" % (fil, str(e)), flush=True)
if not plain_list:
# if failed to process as pieces of rst, then
output = pypandoc.convert_file(fil, out_format, extra_args=extra_args, format='rst')
outputs1 = get_sentences(output, length=max_len)
for oi, output in enumerate(outputs1):
output = output.replace('\n\n', '\n')
plain_list.append([output, fil])
outputs.extend(plain_list)
# report:
# [print(len(x)) for x in outputs]
# deal with blocks longer than context size (sequence length) of 2048
new_outputs = []
num_truncated = 0
num_orig = len(outputs)
for output, fil in outputs:
if len(output) < max_len:
new_outputs.append([output, fil])
continue
outputs1 = get_sentences(output, length=max_len)
for oi, output1 in enumerate(outputs1):
output1 = output1.replace('\n\n', '\n')
new_outputs.append([output1, fil])
num_truncated += 1
print('num_orig: %s num_truncated: %s' % (num_orig, num_truncated), flush=True)
new_outputs = [[k.strip(), fil] for k, fil in new_outputs if len(k.strip()) > min_len]
return new_outputs
def test_scrape_dai_docs_all_pandoc():
"""
pytest -s -v create_data.py::test_scrape_dai_docs_all_pandoc
:return:
"""
dst = setup_dai_docs()
import glob
files = list(glob.glob(os.path.join(dst, '*rst'), recursive=True))
basedir = os.path.abspath(os.getcwd())
new_outputs = rst_to_outputs(files)
os.chdir(basedir)
remove(dst)
save_thing = [{"output": k.strip(), 'prompt_type': 'plain'} for k in new_outputs]
output_file = "dai_docs.train_cleaned.json"
with open(output_file, "wt") as f:
f.write(json.dumps(save_thing, indent=2))
def test_config_to_json():
"""
Needs to run from Driverless AI source directory.
E.g. (base) jon@gpu:~/h2oai$ pytest -s -v /data/jon/h2ogpt/create_data.py::test_config_to_json ; cp config.json /data/jon/h2ogpt/
:return:
"""
try:
# Arrange
import json
from h2oaicore.systemutils import config
toml_list = []
for k, v in config.get_meta_dict().items():
title = (v.title + ": ") if v.title else ''
comment = v.comment or ''
if not (title or comment):
continue
toml_list.extend(
[
{
'prompt_type': 'plain',
'instruction': f"<human>: What does {k} do?\n<bot>: {k.replace('_', ' ')} config.toml: {comment or title}\n<human>:".replace(
"\n", ""),
},
{
'prompt_type': 'plain',
'instruction': f"<human>: Explain {k}.\n<bot>: {k.replace('_', ' ')} config.toml: {comment or title}\n<human>:".replace(
"\n", ""),
},
{
'prompt_type': 'plain',
'instruction': f"<human>: How can I do this: {title}.\n<bot>: Set the {k.replace('_', ' ')} config.toml\n<human>:".replace(
"\n", ""),
} if title and comment else None,
{
'prompt_type': 'human_bot',
'instruction': f'Explain the following expert setting for Driverless AI',
'input': f"{k}",
'output': f"{k.replace('_', ' ')} config.toml: {comment or title}".replace("\n", ""),
},
{
'prompt_type': 'human_bot',
'instruction': f'Explain the following expert setting for Driverless AI',
'input': f"{k}",
'output': f"{k.replace('_', ' ')} config.toml: {title}{comment}".replace("\n", ""),
},
{
'prompt_type': 'human_bot',
'instruction': f'Explain the following expert setting for Driverless AI',
'input': f"{k.replace('_', ' ')}",
'output': f"{k.replace('_', ' ')} config.toml: {title}{comment}".replace("\n", ""),
},
{
'prompt_type': 'human_bot',
'instruction': f'Explain the following expert setting for Driverless AI',
'input': f"{title}",
'output': f"{k.replace('_', ' ')} config.toml: {title}{comment}".replace("\n", ""),
},
{
'prompt_type': 'human_bot',
'instruction': f'Provide a short explanation of the expert setting {k}',
'output': f"{k.replace('_', ' ')} config.toml: {comment or title}".replace("\n", ""),
},
{
'prompt_type': 'human_bot',
'instruction': f'Provide a detailed explanation of the expert setting {k}',
'output': f"{k.replace('_', ' ')} config.toml: {title}{comment}".replace("\n", ""),
},
]
)
toml_list = [x for x in toml_list if x]
with open("config.json", "wt") as f:
f.write(json.dumps(toml_list, indent=2))
except Exception as e:
print("Exception: %s" % str(e), flush=True)
def copy_tree(src, dst, follow_symlink=False):
makedirs(dst, exist_ok=True)
for (path, dirs, files) in os.walk(src, followlinks=follow_symlink):
new_path = path.replace(src, dst)
makedirs(new_path, exist_ok=True)
for file in files:
filename = os.path.join(path, file)
new_filename = os.path.join(new_path, file)
# print("%s -> %s" % (filename, new_filename))
try:
atomic_copy(filename, new_filename)
except FileNotFoundError:
pass
def atomic_move(src, dst):
try:
shutil.move(src, dst)
except (shutil.Error, FileExistsError):
pass
remove(src)
def atomic_copy(src=None, dst=None, with_permissions=True):
if os.path.isfile(dst):
return
import uuid
my_uuid = uuid.uuid4()
dst_tmp = dst + str(my_uuid)
makedirs(os.path.dirname(dst), exist_ok=True)
if with_permissions:
shutil.copy(src, dst_tmp)
else:
shutil.copyfile(src, dst_tmp)
atomic_move(dst_tmp, dst)
remove(dst_tmp)
def makedirs(path, exist_ok=True):
"""
Avoid some inefficiency in os.makedirs()
:param path:
:param exist_ok:
:return:
"""
if os.path.isdir(path) and os.path.exists(path):
assert exist_ok, "Path already exists"
return path
os.makedirs(path, exist_ok=exist_ok)
## Download from https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_unfiltered_cleaned_split.json
## Turn into simple instruct prompt type. No context/previous conversations.
def test_prep_instruct_vicuna():
from datasets import load_dataset
filename = 'ShareGPT_unfiltered_cleaned_split.json'
if not os.path.exists(filename):
os.system(
'wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/%s' % filename)
data = load_dataset("json", data_files={"train": filename})["train"]
training_rows = []
for i in range(data.num_rows):
conversations = data[i]['conversations']
assert isinstance(conversations, list), conversations
convo = ""
for j, conv in enumerate(conversations):
# Get ready for generate.py prompt_type=human_bot
# But train with prompt_type=plain
if conv['from'] == 'human':
FROM = '<human>: '
elif conv['from'] == 'gpt':
FROM = '<bot>: '
convo += f"{FROM}" + conv['value'] + "\n"
if convo:
training_rows.append(dict(input=convo))
with open(filename + ".generate_human_bot.train_plain.json", "wt") as f:
f.write(json.dumps(training_rows, indent=2))
POSTFIX = ".generate_human_bot.train_plain.json"
# https://bair.berkeley.edu/blog/2023/04/03/koala/
OIG_DATASETS = [
"unified_chip2.jsonl",
"unified_grade_school_math_instructions.jsonl",
"unified_poetry_2_song.jsonl",
"unified_plot_screenplay_books_dialog.jsonl",
]
# hub issue: https://huggingface.co/datasets/laion/OIG/discussions/4
ALL_OIG_DATASETS = ['unified_abstract_infill.jsonl',
'unified_basic.jsonl',
'unified_canadian_parliament.jsonl',
'unified_chip2.jsonl',
'unified_conv_finqa.jsonl',
'unified_cuad.jsonl',
'unified_essays.jsonl',
'unified_flan.jsonl.gz',
'unified_grade_school_math_instructions.jsonl',
'unified_hc3_human.jsonl',
'unified_image_prompts_instructions.jsonl',
'unified_joke_explanations.jsonl',
'unified_mathqa_flanv2_kojma_cot.jsonl',
'unified_merged_code_xp3.jsonl',
'unified_multi_news.jsonl',
'unified_multi_sum.jsonl',
'unified_ni.jsonl.gz',
'unified_nq.jsonl',
'unified_openai_summarize_tldr.jsonl',
'unified_oscar_en_sample_dialog.jsonl',
'unified_p3.jsonl.gz',
'unified_plot_screenplay_books_dialog.jsonl',
'unified_poetry_2_song.jsonl',
'unified_poetry_instructions.jsonl',
'unified_rallio_safety_and_prosocial.jsonl',
'unified_rallio_soda_upgraded_2048.jsonl',
'unified_soda_dialog.jsonl',
'unified_sqlv1.jsonl',
'unified_sqlv2.jsonl',
'unified_squad_v2.jsonl',
'unified_squad_v2_more_neg.jsonl',
'unified_ul2_plus_oscar_en_sample_dialog.jsonl',
'unified_unifiedskg_instructions.jsonl',
'unified_unnatural_instructions.jsonl',
'unified_xp3_sample.jsonl']
useful_oig_files = ['unified_rallio_safety_and_prosocial.jsonl.parquet',
'unified_chip2.jsonl.parquet',
'unified_cuad.jsonl.parquet',
'unified_essays.jsonl.parquet',
'unified_flan.jsonl.gz.parquet',
'unified_grade_school_math_instructions.jsonl.parquet',
'unified_hc3_human.jsonl.parquet',
'unified_mathqa_flanv2_kojma_cot.jsonl.parquet',
'unified_merged_code_xp3.jsonl.parquet',
'unified_multi_news.jsonl.parquet',
# 'unified_multi_sum.jsonl.parquet'
'unified_ni.jsonl.gz.parquet',
'unified_openai_summarize_tldr.jsonl.parquet',
# 'unified_oscar_en_sample_dialog.jsonl.parquet', # create text containing these N words, not specific
'unified_plot_screenplay_books_dialog.jsonl.parquet',
'unified_soda_dialog.jsonl.parquet',
'unified_unnatural_instructions.jsonl.parquet',
]
@pytest.mark.parametrize("filename", OIG_DATASETS)
def test_get_small_sample_oig_data(filename):
if not os.path.exists(filename):
os.system('wget https://huggingface.co/datasets/laion/OIG/resolve/main/%s' % filename)
import json
rows = []
with open(filename, "r") as f:
for line in f.readlines():
row = json.loads(line)
rows.append(dict(input=row["text"]))
with open(filename + POSTFIX, "w") as f:
f.write(json.dumps(rows, indent=2))
@pytest.mark.parametrize("filename", ALL_OIG_DATASETS)
def test_download_useful_data_as_parquet(filename):
dest_file = filename + '.parquet'
if dest_file not in useful_oig_files:
pytest.skip('file declared not useful')
if not os.path.exists(filename):
os.system('wget https://huggingface.co/datasets/laion/OIG/resolve/main/%s' % filename)
if not os.path.exists(dest_file):
df = pd.read_json(path_or_buf=filename, lines=True)
df.to_parquet(dest_file, index=False)
def test_merge_shuffle_small_sample_oig_data():
np.random.seed(1234)
rows = []
for filename in OIG_DATASETS:
with open(filename + POSTFIX, "r") as f:
rows.extend(json.loads(f.read()))
np.random.shuffle(rows)
with open("merged_shuffled_OIG_%s.json" % hashlib.sha256(str(OIG_DATASETS).encode()).hexdigest()[:10], "w") as f:
f.write(json.dumps(rows, indent=2))
def test_join_jsons():
files = ['config.json'] * 1 + \
['dai_docs.train_cleaned.json'] * 2 + \
['dai_faq.json'] * 3
print(files)
lst = []
[lst.extend(json.load(open(fil, 'rt'))) for fil in files]
print(len(lst))
json.dump(lst, open("merged.json", "wt"), indent=2)
@pytest.mark.parametrize("filename", ['Anthropic/hh-rlhf'])
def test_make_rlhf_good_data(filename):
from datasets import load_dataset
rows = load_dataset(filename)["train"]["chosen"]
new_rows = []
for row in rows:
if row[:2] == "\n\n":
row = row[2:]
row = row.replace("Human: ", "<human>: ")
row = row.replace("Assistant: ", "<bot>: ")
new_rows.append(dict(input=row))
with open(filename.replace("/", "_") + POSTFIX, "w") as f:
f.write(json.dumps(new_rows, indent=2))
def test_show_prompts():
files = ['config.json'] * 1 + \
['dai_docs.train_cleaned.json'] * 1 + \
['dai_faq.json'] * 1
file_points = [json.load(open(fil, 'rt')) for fil in files]
from prompter import generate_prompt
for data_points in file_points:
for data_point in data_points:
print(generate_prompt(data_point, 'plain', '', False, False, False)[0])
def test_get_open_datasets():
# HF changed things so don't get raw list of all datasets, so not have to filter, but can't do negative filter
open_tags = ['license:Apache License 2.0',
'license:mit',
'license:apache',
'license:apache2',
'license:apache-2.0',
'license:bsd',
'license:bsd-2-clause',
'license:bsd-3-clause',
'license:bsd-3-clause-clear',
'license:lgpl-2.1',
'license:lgpl-3.0',
'license:lgpl-lr',
'license:lgpl',
'license:openrail++',
'license:openrail',
'license:bigscience-bloom-rail-1.0',
# 'license:agpl-3.0',
'license:other',
'license:unknown',
# 'license:mpl-2.0', # ok, but would have to include original copyright, license, source, copies in distribution
# Attribution required:
'license:odc-by',
'license:cc-by-4.0',
'license:cc-by-3.0',
'license:cc-by-2.0',
'license:cc-by-2.5',
# 'license:cc-by-sa-4.0', # would require same license
'license:odbl',
'license:pddl',
'license:ms-pl',
'license:zlib',
]
# bad license: cc-by-nc-4.0
from huggingface_hub import list_datasets
datasets = flatten_list([[x for x in list_datasets(filter=y)] for y in open_tags])
datasets += [x for x in list_datasets(author='openai')]
# check all:
all_license_tags = set(flatten_list([[y for y in x.tags if 'license' in y] for x in datasets]))
print(len(all_license_tags))
open_datasets = [x for x in datasets if any([y in x.tags for y in open_tags]) or 'license:' not in str(x.tags)]
print('open_datasets', len(open_datasets))
all_task_tags = set(flatten_list([[y for y in x.tags if 'task' in y] for x in open_datasets]))
print('all_task_tags', len(all_task_tags))
excluded_tags = ['image', 'hate', 'tabular', 'table-', 'classification', 'retrieval',
'translation', 'identification', 'object', 'mask', 'to-text',
'face-detection', 'audio', 'voice', 'reinforcement', 'depth-est',
'forecasting', 'parsing', 'visual', 'speech', 'multiple-choice',
'slot-filling', 'irds/argsme', '-scoring', 'other', 'graph-ml',
'feature-extraction', 'keyword-spotting',
'coreference-resolution', 'segmentation',
'word-sense-disambiguation',
'lemmatization']
task_tags = [x.replace('task_categories:', '').replace('task_ids:', '')
for x in all_task_tags if not any([y in x for y in
excluded_tags])]
print('task_tags', len(task_tags))
# str(x.tags) to catch any pattern match to anything in list
open_tasked_datasets = [x for x in open_datasets if
any([y in str([x for x in x.tags if 'task' in x]) for y in task_tags]) and
not any([y in str([x for x in x.tags if 'task' in x]) for y in excluded_tags]) or
'task_categories' not in str(x.tags) and 'task_ids' not in str(x.tags)]
open_tasked_datasets = [x for x in open_tasked_datasets if not x.disabled]
open_tasked_datasets = [x for x in open_tasked_datasets if not x.gated]
open_tasked_datasets = [x for x in open_tasked_datasets if not x.private]
print('open_tasked_datasets', len(open_tasked_datasets))
sizes = list(set(flatten_list([[(y, x.id) for y in x.tags if 'size' in y] for x in open_tasked_datasets])))
languages = list(set(flatten_list([[(y, x.id) for y in x.tags if 'language:' in y] for x in open_tasked_datasets])))
open_english_tasked_datasets = [x for x in open_tasked_datasets if
'language:' not in str(x.tags) or
'language:en' in str(x.tags)]
small_open_english_tasked_datasets = [x for x in open_english_tasked_datasets if
'n<1K' in str(x.tags) or
'1K<n<10K' in str(x.tags) or
'1K0<n<100K' in str(x.tags) or
'100K<n<1M' in str(x.tags) or
'size_category' not in str(x.tags)
]
# 'aeslc' : email_body, subject -> summarization?
# load_dataset(open_tasked_datasets[0].id).data['train'].to_pandas()
ids = [x.id for x in small_open_english_tasked_datasets]
# sanity checks
# https://bair.berkeley.edu/blog/2023/04/03/koala/
assert 'alespalla/chatbot_instruction_prompts' in ids
assert 'laion/OIG' in ids
assert 'openai/webgpt_comparisons' in ids
assert 'openai/summarize_from_feedback' in ids
assert 'Anthropic/hh-rlhf' in ids
# useful but not allowed for commercial purposes:
# https://huggingface.co/datasets/squad
print('open_english_tasked_datasets: ', ids, flush=True)
exclude_ids = ['allenai/nllb', # translation only
'hf-internal-testing/fixtures_image_utils', # testing
'allenai/c4', # search-url
'agemagician/uniref50', # unknown
'huggingface-course/documentation-images', # images
'smilegate-ai/kor_unsmile', # korean
'MohamedRashad/ChatGPT-prompts', # ChatGPT/LearnGPT/https://www.emergentmind.com/
'humarin/chatgpt-paraphrases', # Paraphrase using ChatGPT
'Jeska/vaccinchat', # not useful
'alespalla/chatbot_instruction_prompts', # mixes alpaca
'allenai/prosocial-dialog',
# already exlucded, but wrongly in other datasets that say more permissive license
'AlekseyKorshuk/persona-chat', # low quality
'bavard/personachat_truecased', # low quality
'adamlin/daily_dialog', # medium quality conversations
'adamlin/FewShotWoz', # low quality
'benjaminbeilharz/better_daily_dialog', # low quality
'benjaminbeilharz/daily_dialog_w_turn_templates', # low
'benjaminbeilharz/empathetic_dialogues_for_lm', # low
'GEM-submissions/GEM__bart_base_schema_guided_dialog__1645547915', # NA
'ia-bentebib/conv_ai_2_fr', # low fr
'ia-bentebib/daily_dialog_fr', # low fr
'ia-bentebib/dialog_re_fr', # low fr
'ia-bentebib/empathetic_dialogues_fr', # low fr
'roskoN/dailydialog', # low
'VadorMazer/skyrimdialogstest', # low
'bigbio/med_qa', # med specific Q/A
'biu-nlp/qa_srl2018', # low quality Q/A
'biu-nlp/qa_discourse', # low quality Q/A
'iarfmoose/qa_evaluator', # low quality Q/A
'jeopardy', # low quality Q/A -- no reasoning
'narrativeqa', # low quality Q/A
'nomic-ai/gpt4all_prompt_generations', # bad license
'nomic-ai/gpt4all_prompt_generations_with_p3', # bad license
'HuggingFaceH4/alpaca', # bad license
'tatsu-lab/alpaca', # ToS breaking
'yahma/alpaca-cleaned', # ToS breaking
'Hello-SimpleAI/HC3', # bad license
'glue', # no reasoning QA
'sahil2801/CodeAlpaca-20k', # bad license
'Short-Answer-Feedback/saf_communication_networks_english', # long Q, medium A
]
small_open_english_tasked_datasets = [x for x in small_open_english_tasked_datasets if x.id not in exclude_ids]
# some ids clearly speech related
small_open_english_tasked_datasets = [x for x in small_open_english_tasked_datasets if 'speech' not in x.id]
# HF testing
small_open_english_tasked_datasets = [x for x in small_open_english_tasked_datasets if
'hf-internal-testing' not in x.id]
small_open_english_tasked_datasets = [x for x in small_open_english_tasked_datasets if
'chinese' not in x.id]
sorted_small_open_english_tasked_datasets = sorted([(x.downloads, x) for x in small_open_english_tasked_datasets],
key=lambda x: x[0], reverse=True)
# NOTES:
# Run like pytest -s -v create_data.py::test_get_open_datasets &> getdata9.log
# See what needs config passed and add:
# grep 'load_dataset(' getdata9.log|grep -v data_id|less -S
# grep "pip install" getdata9.log
# NOTE: Some datasets have default config, but others are there. Don't know how to access them.
"""
https://huggingface.co/datasets/wikihow/blob/main/wikihow.py
https://github.com/mahnazkoupaee/WikiHow-Dataset
https://ucsb.box.com/s/ap23l8gafpezf4tq3wapr6u8241zz358
https://ucsb.app.box.com/s/ap23l8gafpezf4tq3wapr6u8241zz358
"""
"""
# some ambiguous or non-commercial datasets
https://github.com/PhoebusSi/alpaca-CoT
"""
timeout = 3 * 60
# laion/OIG takes longer
for num_downloads, dataset in sorted_small_open_english_tasked_datasets:
data_id = dataset.id
func = do_one
args = (data_id, num_downloads)
kwargs = {}
with ProcessPoolExecutor(max_workers=1) as executor:
future = executor.submit(func, *args, **kwargs)
try:
future.result(timeout=timeout)
except concurrent.futures.TimeoutError:
print("\n\ndata_id %s timeout\n\n" % data_id, flush=True)
for child in psutil.Process(os.getpid()).children(recursive=True):
os.kill(child.pid, signal.SIGINT)
os.kill(child.pid, signal.SIGTERM)
os.kill(child.pid, signal.SIGKILL)
def do_one(data_id, num_downloads):
from datasets import load_dataset
out_file = "data_%s.parquet" % str(data_id.replace('/', '_'))
if os.path.isfile(out_file) and os.path.getsize(out_file) > 1024 ** 3:
return
try:
print("Loading data_id %s num_downloads: %s" % (data_id, num_downloads), flush=True)
avail_list = None
try:
data = load_dataset(data_id, 'foobar')
except Exception as e:
if 'Available: ' in str(e):
avail_list = ast.literal_eval(str(e).split('Available:')[1].strip())
else:
avail_list = None
if avail_list is None:
avail_list = [None]
print("%s avail_list: %s" % (data_id, avail_list), flush=True)
for name in avail_list:
out_file = "data_%s_%s.parquet" % (str(data_id.replace('/', '_')), str(name))
if os.path.isfile(out_file):
continue
data = load_dataset(data_id, name)
column_names_dict = data.column_names
column_names = column_names_dict[list(column_names_dict.keys())[0]]
print("Processing data_id %s num_downloads: %s columns: %s" % (data_id, num_downloads, column_names),
flush=True)
data_dict = data.data
col_dict = data.num_columns
first_col = list(col_dict.keys())[0]
if 'train' in data_dict:
df = data['train'].to_pandas()
else:
df = data[first_col].to_pandas()
# csv has issues with escaping chars, even for datasets I know I want
df.to_parquet(out_file, index=False)
except Exception as e:
t, v, tb = sys.exc_info()
ex = ''.join(traceback.format_exception(t, v, tb))
print("Exception: %s %s" % (data_id, ex), flush=True)
def test_otherlic():
from huggingface_hub import list_datasets
lic = ['license:odc-by',
'license:cc-by-4.0',
'license:cc-by-3.0',
'license:cc-by-2.0',
'license:cc-by-2.5',
'license:cc-by-sa-4.0',
'license:odbl',
'license:pddl',
'license:ms-pl',
'license:zlib',
]
datasets = flatten_list([[x for x in list_datasets(filter=y) if 'translation' not in str(x.tags)] for y in lic])
print(len(datasets))
# These useful datasets are determined based upon data sample, column types, and uniqueness compared to larger datasets like Pile
# grep columns getdata13.log|grep -v "\['image'\]"|sort|uniq|grep -v tokens|grep -v "'image'"|grep -v embedding|grep dialog
useful = ['Dahoas/instruct-human-assistant-prompt',
'Dahoas/first-instruct-human-assistant-prompt',
'knkarthick/dialogsum', # summary of conversation
'McGill-NLP/FaithDial', # medium quality
'Zaid/quac_expanded', # medium quality context + QA
'0-hero/OIG-small-chip2', # medium
'alistvt/coqa-flat', # QA medium
'AnonymousSub/MedQuAD_47441_Question_Answer_Pairs', # QA medium
'Anthropic/hh-rlhf', # high quality # similar to Dahoas/full-hh-rlhf
'arjunth2001/online_privacy_qna', # good quality QA
'Dahoas/instruct_helpful_preferences', # medium quality instruct
'Dahoas/rl-prompt-dataset', # medium chat
'Dahoas/rm-static', # medium chat
'Dahoas/static-hh', # medium chat # HuggingFaceH4/self_instruct
'Dahoas/synthetic-instruct-gptj-pairwise', # medium chat
'eli5', # QA if prompt ELI5
'gsm8k', # QA (various)
'guanaco/guanaco', # prompt/response
'kastan/rlhf-qa-comparisons', # good QA
'kastan/rlhf-qa-conditional-generation-v2', # prompt answer
'OllieStanley/humaneval-mbpp-codegen-qa', # code QA, but started from words, so better than other code QA
'OllieStanley/humaneval-mbpp-testgen-qa', # code QA
'Graverman/Instruct-to-Code', # code QA
'openai/summarize_from_feedback', # summarize
'relbert/analogy_questions', # analogy QA
'yitingxie/rlhf-reward-datasets', # prompt, chosen, rejected.
'yizhongw/self_instruct', # instruct (super natural & instruct)
'HuggingFaceH4/asss', # QA, big A
'kastan/rlhf-qa-conditional-generation-v2', # QA
'cosmos_qa', # context QA
'vishal-burman/c4-faqs', # QA but not so much reasoning, but alot of text
'squadshifts', # QA from context
'hotpot_qa', # QA from context
'adversarial_qa', # QA from context
'allenai/soda', # dialog -> narrative/summary
'squad_v2', # context QA
'squadshifts', # context QA
'dferndz/cSQuAD1', # context QA
'dferndz/cSQuAD2', # context QA
'din0s/msmarco-nlgen', # context QA
'domenicrosati/TruthfulQA', # common sense truthful QA -- trivia but good trivia
'hotpot_qa', # context, QA
'HuggingFaceH4/self-instruct-eval', # instruct QA, medium quality, some language reasoning
'kastan/EE_QA_for_RLHF', # context QA
'KK04/LogicInference_OA', # instruction logical QA
'lmqg/qa_squadshifts_synthetic', # context QA
'lmqg/qg_squad', # context QA
'lmqg/qg_squadshifts', # context QA
'lmqg/qg_subjqa', # context QA
'pszemraj/HC3-textgen-qa',
# QA medium, has human responses -- humans tend to provide links instead of trying to answer
'pythonist/newdata', # long context, QA, brief A
'ropes', # long background, situation, question, A
'wikitablequestions', # table -> QA
'bigscience/p3', # context QA but short answers
]
code_useful = ['0n1xus/codexglue',
'openai_humaneval',
'koutch/staqc',
]
maybe_useful = ['AlekseyKorshuk/comedy-scripts',
'openbookqa', # hard to parse, low reasoning
'qed', # reasonable QA, but low reasoning
'selqa', # candidate answers
'HuggingFaceH4/instruction-pilot-outputs-filtered',
'GBaker/MedQA-USMLE-4-options', # medical QA with long questions
'npc-engine/light-batch-summarize-dialogue', # dialog summarize, kinda low specific quality
]
summary_useful = ['austin/rheum_abstracts',
'CarperAI/openai_summarize_comparisons', # summarize chosen/rejected
'CarperAI/openai_summarize_tldr', # summarize QA
'ccdv/cnn_dailymail', # summarize news
'ccdv/govreport-summarization', # summarize high quality
'ccdv/pubmed-summarization', # summarize high quality
'duorc', # plot -> QA
'farleyknight/big_patent_5_percent', # desc -> abstract
'multi_news', # summary
'opinosis',
'SophieTr/reddit_clean',
'allenai/mup', # long text -> summary
'allenai/multi_lexsum', # long text -> summary
'big_patent',
'allenai/wcep_dense_max',
'awinml/costco_long_practice',
'GEM/xsum',
'ratishsp/newshead',
'RussianNLP/wikiomnia', # russian
'stacked-summaries/stacked-xsum-1024',
]
math_useful = [
'competition_math'
]
skipped = ['c4', # maybe useful, used for flan, but skipped due to size
]
"""
To get training data from oig:
pytest test_oig test_grade_final test_finalize_to_json
"""
human = '<human>:'
bot = '<bot>:'
def test_assemble_and_detox():
import re
from profanity_check import predict_prob
df_list = []
for data in useful_oig_files:
print("Processing %s" % data, flush=True)
df = pd.read_parquet(data)
df = df.reset_index(drop=True)
# chop up into human/bot interactions of no more than 10kB per row
text_list = df[['text']].values.ravel().tolist()
new_text = []
max_len = 2048 # uber cutoff
MAX_LEN = 2048 // 2 - 30 # max len per question/answer
for text in tqdm(text_list):
human_starts = [m.start() for m in re.finditer('<human>: ', text)]
if len(human_starts) == 1:
human_starts = [0, len(text)] # always go into for loop below
blurb = ''
for i in range(len(human_starts) - 1):
interaction = text[human_starts[i]: human_starts[i + 1]][:max_len]
blurb += interaction
if len(blurb) >= MAX_LEN:
blurb = get_sentences(blurb, length=MAX_LEN)[0]
new_text.append(blurb + "\n<human>:")
blurb = ''
if blurb:
blurb = get_sentences(blurb, length=MAX_LEN)[0]
new_text.append(blurb + "\n<human>:")
if len(new_text) > len(text_list):
print("Added %d new rows (before: %d)" % (len(new_text) - df.shape[0], df.shape[0]))
df = pd.DataFrame({"text": new_text, "source": [data] * len(new_text)})
df = df.drop_duplicates(keep='first')
print(df['text'].apply(lambda x: len(x)).describe())
assert df['text'].apply(lambda x: len(x)).max() <= 2 * max_len
# faster than better_profanity, do early
df['profanity'] = predict_prob(df['text'])
before_rows = df.shape[0]
df = df[df['profanity'] < 0.25] # drop any low quality stuff
after_rows = df.shape[0]
print("Dropped %d rows out of %d due to alt-profanity-check" % (before_rows - after_rows, before_rows))
df_list.append(df)
print("Done processing %s -> %s rows" % (data, df.shape[0]), flush=True)
print("So far have %d rows" % sum([len(x) for x in df_list]))
df_final = pd.concat(df_list)
df_final = df_final.sample(frac=1, random_state=1234).reset_index(drop=True)
df_final.to_parquet('h2oGPT.cleaned.human_bot.shorter.parquet', index=False)
def test_basic_cleaning():
# from better_profanity import profanity
# https://pypi.org/project/alt-profanity-check/
from profanity_check import predict
df_list = []
for data in useful_oig_files:
# for data in useful_oig_files[:5]:
# for data in ['unified_openai_summarize_tldr.jsonl.parquet']:
print("Processing %s" % data, flush=True)
df = pd.read_parquet(data)
df = df.reset_index(drop=True)
# NOTE: Not correct if multiple human-bot interactions, but those dialogs even more desired
# avg_chars = len(df['text'][0])/(df['text'][0].count(human)+df['text'][0].count(bot))
df['avg_words'] = df['text'].apply(lambda x: x.count(' ') / (x.count(human) + x.count(bot)) / 2.0)
df['avg_bot_words'] = df['text'].apply(lambda x: x.split(bot)[1].count(' ') / x.count(bot))
# df['bad_words'] = df['text'].apply(lambda x: profanity.contains_profanity(x))
# low_quality_patterns = ['Write the rest of this wikipedia article']
res = predict(df['text'])
df['bad_words'] = res
df = df.reset_index(drop=True)
df = df[df['bad_words'] == 0]
df = df[['text', 'avg_words', 'avg_bot_words']]
df = df.drop_duplicates(keep='first')
print(df[df['avg_words'] == df['avg_words'].max()]['text'].values)
median_words = np.median(df['avg_words'])
min_words_per_entity = max(30, 0.8 * median_words)
max_words_per_entity = 2048 # too hard to learn from for now
df = df[df['avg_words'] > min_words_per_entity]
df = df[df['avg_words'] < max_words_per_entity]
min_words_per_entity = max(20, 0.5 * median_words) # bot should say stuff for now
max_words_per_entity = 2048 # too hard to learn from for now
df = df[df['avg_bot_words'] > min_words_per_entity]
df = df[df['avg_bot_words'] < max_words_per_entity]
df_list.append(df)
print("Done processing %s -> %s rows" % (data, df.shape[0]), flush=True)
df_final = pd.concat(df_list)
df_final.to_parquet('h2oGPT.cleaned.human_bot.parquet', index=False)
from joblib import Parallel, delayed, effective_n_jobs
from sklearn.utils import gen_even_slices
from sklearn.utils.validation import _num_samples
def parallel_apply(df, func, n_jobs=-1, **kwargs):
""" Pandas apply in parallel using joblib.
Uses sklearn.utils to partition input evenly.
Args:
df: Pandas DataFrame, Series, or any other object that supports slicing and apply.
func: Callable to apply
n_jobs: Desired number of workers. Default value -1 means use all available cores.
**kwargs: Any additional parameters will be supplied to the apply function
Returns:
Same as for normal Pandas DataFrame.apply()
"""
if effective_n_jobs(n_jobs) == 1:
return df.apply(func, **kwargs)
else:
ret = Parallel(n_jobs=n_jobs)(
delayed(type(df).apply)(df[s], func, **kwargs)
for s in gen_even_slices(_num_samples(df), effective_n_jobs(n_jobs)))
return pd.concat(ret)
def add_better_profanity_flag(df):
from better_profanity import profanity
df['better_profanity'] = parallel_apply(
df['text'],
lambda x: profanity.contains_profanity(x),
n_jobs=-1,
)
return df
def add_textstat_grade(df):
import textstat
def myfunc(x):
return textstat.flesch_kincaid_grade(x) # simple grade
if False:
import dask.dataframe as dd
# 40 seconds for 1000 rows, but have 1,787,799 rows
ddata = dd.from_pandas(df, npartitions=120)
df['flesch_grade'] = ddata['text'].apply(myfunc).compute()
if True:
# fast way
df['flesch_grade'] = parallel_apply(df['text'], myfunc, n_jobs=-1)
return df
def add_deberta_grade(df):
from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch
reward_name = "OpenAssistant/reward-model-deberta-v3-large-v2"
rank_model, tokenizer = AutoModelForSequenceClassification.from_pretrained(
reward_name), AutoTokenizer.from_pretrained(reward_name)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
rank_model.to(device)
def get_question(x):
return x.replace('<human>: ', '').split('<bot>:')[0]
def get_answer(x):
try:
answer = x.split('<bot>: ')[1].split('<human>:')[0].replace('<bot>: ', '')
except:
answer = x.split('<bot>:')[1].split('<human>:')[0].replace('<bot>:', '')
return answer
df['question'] = parallel_apply(df['text'], get_question, n_jobs=-1)
df['answer'] = parallel_apply(df['text'], get_answer, n_jobs=-1)
from datasets import Dataset
from transformers import pipeline
from transformers.pipelines.pt_utils import KeyPairDataset
import tqdm
pipe = pipeline(
"text-classification",
model=reward_name,
device="cuda:0" if torch.cuda.is_available() else "cpu"
)
start = 0
batch_size = 64 * 16
micro_batch = orig_micro_batch = 16
end = 0
import socket
checkpoint = "grades.%s.pkl" % socket.gethostname()
grades = []
import pickle
if os.path.exists(checkpoint):
with open(checkpoint, "rb") as f:
start, grades = pickle.loads(f.read())
last_oom = 0
while end < df.shape[0]:
# manual batching to handle OOM more gracefully
end = min(start + batch_size, df.shape[0])
if start == end:
break
dataset = Dataset.from_pandas(df.iloc[start:end, :])
try:
grades.extend([
x['score'] for x in tqdm.tqdm(
pipe(KeyPairDataset(dataset, "question", "answer"), batch_size=micro_batch)
)
])
except torch.cuda.OutOfMemoryError:
last_oom = start
micro_batch = max(1, micro_batch // 2)
print("OOM - retrying with micro_batch=%d" % micro_batch)
continue
if last_oom == start:
micro_batch = orig_micro_batch
print("Returning to micro_batch=%d" % micro_batch)
assert len(grades) == end
start = end
with open(checkpoint, "wb") as f:
f.write(pickle.dumps((end, grades)))
print("%d/%d" % (end, df.shape[0]))
df['grade_deberta'] = grades
if os.path.exists(checkpoint):
os.remove(checkpoint)
return df
def test_chop_by_lengths():
file = "h2oGPT.cleaned.human_bot.shorter.parquet"
df = pd.read_parquet(file).reset_index(drop=True)
df = count_human_bot_lengths(df)
df['rand'] = np.random.rand(df.shape[0])
df['rand2'] = np.random.rand(df.shape[0])
before_rows = df.shape[0]
# throw away short human/bot responses with higher likelihood
df = df[(df['len_human_mean'] > 20)] # never keep very short ones
df = df[(df['len_human_mean'] > 30) | (df['rand'] < 0.2)]
df = df[(df['len_human_mean'] > 50) | (df['rand'] < 0.5)]
df = df[(df['len_human_max'] < 10000)] # drop super long (basically only human) ones
df = df[(df['len_bot_mean'] > 20)] # never keep very short ones
df = df[(df['len_bot_mean'] > 30) | (df['rand2'] < 0.2)]
df = df[(df['len_bot_mean'] > 50) | (df['rand2'] < 0.5)]
df = df[(df['len_bot_max'] < 10000)] # drop super long (only bot) ones
assert df['text'].apply(lambda x: len(x)).max() < 20000
df = df.drop(['rand', 'rand2'], axis=1)
after_rows = df.shape[0]
print("Chopped off %d out of %d rows due to length" % (before_rows - after_rows, before_rows))
print(df.describe())
df.to_parquet('h2oGPT.cleaned.chopped.human_bot.shorter.parquet', index=False)
def count_human_bot_lengths(df, human=None, bot=None):
import re
len_human_min = []
len_human_max = []
len_human_mean = []
len_bot_min = []
len_bot_max = []
len_bot_mean = []
human = human or '<human>:'
bot = bot or '<bot>:'
for is_human in [True, False]:
what = human if is_human else bot
other = human if not is_human else bot
for i in range(df.shape[0]):
text = df.loc[i, 'text']
assert isinstance(text, str)
starts = [m.start() for m in re.finditer(what, text)]
if len(starts) == 1:
starts = [starts[0], len(text)] # always go into for loop below
assert len(text)
list_what = []
for ii in range(len(starts) - 1):
interaction = text[starts[ii]: starts[ii + 1]]
if other in interaction:
interaction = interaction[:interaction.find(other)]
interaction.strip()
list_what.append(interaction)
if not list_what:
list_what = [''] # handle corrupted data, very rare, leads to sizes 0
if is_human:
len_human_min.append(min([len(x) for x in list_what]))
len_human_max.append(max([len(x) for x in list_what]))
len_human_mean.append(np.mean([len(x) for x in list_what]))
else:
len_bot_min.append(min([len(x) for x in list_what]))
len_bot_max.append(max([len(x) for x in list_what]))
len_bot_mean.append(np.mean([len(x) for x in list_what]))
df['len_human_min'] = len_human_min
df['len_human_max'] = len_human_max
df['len_human_mean'] = len_human_mean
df['len_bot_min'] = len_bot_min
df['len_bot_max'] = len_bot_max
df['len_bot_mean'] = len_bot_mean
np.random.seed(1234)
pd.set_option('display.max_columns', None)
print("Before chopping")
print(df.describe())
return df
def test_grade():
df = None
file = "h2oGPT.cleaned.chopped.human_bot.shorter.parquet"
output_file = "h2oGPT.cleaned.graded1.human_bot.shorter.parquet"
if not os.path.exists(output_file):
if df is None:
df = pd.read_parquet(file).reset_index(drop=True)
df = add_textstat_grade(df)
min_grade = 10
max_grade = 25
df = df[df['flesch_grade'] >= min_grade]
df = df[df['flesch_grade'] <= max_grade]
print("After Flesch grade")
print(df.describe())
df.to_parquet(output_file, index=False)
file = output_file
output_file = "h2oGPT.cleaned.graded2.human_bot.shorter.parquet"
if not os.path.exists(output_file):
# slower than alt-profanity, do last, but do before deberta grading, since that's slower
if df is None:
df = pd.read_parquet(file).reset_index(drop=True)
df = add_better_profanity_flag(df)
before_rows = df.shape[0]
df = df[df['better_profanity'] == 0]
df = df.drop(['better_profanity'], axis=1)
after_rows = df.shape[0]
print("Dropped %d rows out of %d due to better_profanity" % (before_rows - after_rows, before_rows))
print(df.describe())
df.to_parquet(output_file, index=False)
file = output_file
output_file = 'h2oGPT.cleaned.graded3.human_bot.shorter.parquet'
if not os.path.exists(output_file):
if df is None:
df = pd.read_parquet(file).reset_index(drop=True)
df = add_deberta_grade(df)
min_grade = 0.3
max_grade = np.inf
before_rows = df.shape[0]
df = df[df['grade_deberta'] >= min_grade]
df = df[df['grade_deberta'] <= max_grade]
after_rows = df.shape[0]
print("Dropped %d rows out of %d due to deberta grade" % (before_rows - after_rows, before_rows))
print("After DeBERTa grade")
print(df.describe())
df.to_parquet(output_file, index=False)
file = output_file
output_file = 'h2oGPT.cleaned.graded.human_bot.shorter.parquet'
if df is None:
df = pd.read_parquet(file).reset_index(drop=True)
df.to_parquet(output_file, index=False)
@pytest.mark.parametrize(
"fixup_personality, only_personality, deberta_grading",
[
# [False, False, False],
# [True, True, False],
[True, False, False],
# [True, False, True],
]
)
@pytest.mark.parametrize("prompt_type", ["llama2"])
def test_add_open_assistant(fixup_personality, only_personality, deberta_grading, prompt_type, save_json=True):
"""
Flatten tree structure into one row per path from root to leaf
Also turn into human_bot prompting format:
<human>: question\n<bot>: answer <human>: question2\n<bot>: answer2 Etc.
Also saves a .json locally as side-effect
returns list of dicts, containing intput, prompt_type and source
"""
from datasets import load_dataset
data_file = "OpenAssistant/oasst1"
ds = load_dataset(data_file)
df = pd.concat([ds['train'].to_pandas(), ds['validation'].to_pandas()], axis=0)
rows = {}
message_ids = df['message_id'].values.tolist()
message_tree_ids = df['message_tree_id'].values.tolist()
parent_ids = df['parent_id'].values.tolist()
texts = df['text'].values.tolist()
roles = df['role'].values.tolist()
deleteds = df['deleted'].values.tolist()
for i in range(df.shape[0]):
# collect all trees
message_id = message_ids[i]
message_tree_id = message_tree_ids[i]
parent_id = parent_ids[i]
text = texts[i]
deleted = deleteds[i]
if deleted:
continue
if fixup_personality:
text = text.replace("Open Assistant", "h2oGPT")
text = text.replace("Open-Assistant", "h2oGPT")
text = text.replace("open-assistant", "h2oGPT")
text = text.replace("OpenAssistant", "h2oGPT")
text = text.replace("open assistant", "h2oGPT")
text = text.replace("Open Assistand", "h2oGPT")
text = text.replace("Open Assitant", "h2oGPT")
text = text.replace("Open Assistent", "h2oGPT")
text = text.replace("Open Assisstant", "h2oGPT")
text = text.replace("Open Assitent", "h2oGPT")
text = text.replace("Open Assitiant", "h2oGPT")
text = text.replace("Open Assistiant", "h2oGPT")
text = text.replace("Open Assitan ", "h2oGPT ")
text = text.replace("Open Assistan ", "h2oGPT ")
text = text.replace("Open Asistant", "h2oGPT")
text = text.replace("Open Assiant", "h2oGPT")
text = text.replace("Assistant", "h2oGPT")
text = text.replace("LAION AI", "H2O.ai")
text = text.replace("LAION-AI", "H2O.ai")
text = text.replace("LAION,", "H2O.ai,")
text = text.replace("LAION.ai", "H2O.ai")
text = text.replace("LAION.", "H2O.ai.")
text = text.replace("LAION", "H2O.ai")
role = roles[i]
if prompt_type == "llama2":
new_data = ('[INST] ' if role == 'prompter' else ' [/INST] ') + text
if parent_id and role == 'prompter':
new_data = " " + new_data
elif prompt_type == "human_bot":
new_data = ('<human>: ' if role == 'prompter' else '<bot>: ') + text
else:
raise NotImplementedError("prompt_type not supported")
entry = dict(message_id=message_id, parent_id=parent_id, text=new_data)
if message_tree_id not in rows:
rows[message_tree_id] = [entry]
else:
rows[message_tree_id].append(entry)
all_rows = []
for node_id in rows:
# order responses in tree, based on message/parent relationship
conversations = []
list_msgs = rows[node_id]
# find start
while len(list_msgs):
for i, leaf in enumerate(list_msgs):
found = False
parent_id = leaf['parent_id']
if parent_id is None:
# conversation starter
conversations.append(leaf)
found = True
else:
for conv in conversations:
# find all conversations to add my message to
if parent_id in conv['message_id'] and parent_id != conv['message_id'][-len(parent_id):]:
# my message doesn't follow conversation
continue
if parent_id == conv['message_id'][-len(parent_id):]:
# my message follows conversation, but fork first, so another follow-on message can do same
conversations.append(conv.copy())
if prompt_type == "llama2":
conv['text'] += f"""{leaf['text']}"""
elif prompt_type == "human_bot":
conv['text'] += f"""
{leaf['text']}
"""
else:
raise NotImplementedError
conv['message_id'] += leaf['message_id']
found = True
break
if found:
# my content was used, so nuke from list
del list_msgs[i]
break
# now reduce down to final conversations, find the longest chains of message ids
for i, conv in enumerate(conversations):
for j, conv2 in enumerate(conversations):
if i == j:
continue
if conv['message_id'] and conv2['message_id']:
assert conv['message_id'] != conv2['message_id']
# delete the shorter conversation, if one contains the other
if conv['message_id'] in conv2['message_id']:
conv['message_id'] = None
if conv2['message_id'] in conv['message_id']:
conv2['message_id'] = None
conversations = [c for c in conversations if c['message_id']]
if only_personality:
if prompt_type == "human_bot":
all_rows.extend(
[dict(input=c['text'] + "\n<human>:", output="", prompt_type='plain', source=data_file) for c in conversations if
'h2oGPT' in c['text']])
elif prompt_type == "llama2":
all_rows.extend(
[dict(input=c['text'] +
("" if c['text'].rfind("[/INST]") > c['text'].rfind("[INST]") else " [/INST]"),
output="", prompt_type='plain', source=data_file) for c in conversations if
'h2oGPT' in c['text']])
else:
raise NotImplementedError
else:
if prompt_type == "human_bot":
all_rows.extend(
[dict(input=c['text'] + "\n<human>:", output="", prompt_type='plain', source=data_file) for c in conversations
if
"What is H2O.ai" not in c['text']])
elif prompt_type == "llama2":
all_rows.extend(
[dict(input=c['text'] +
(" " if c['text'].rfind("[/INST]") > c['text'].rfind("[INST]") else " [/INST]"),
output="", prompt_type='plain', source=data_file) for c in conversations if
"What is H2O.ai" not in c['text']])
else:
raise NotImplementedError
unhelpful = get_unhelpful_list()
all_rows = [x for x in all_rows if not any(u in x['input'] for u in unhelpful)]
personality = create_personality_data(prompt_type=prompt_type)
all_rows.extend(personality * 10)
np.random.seed(123)
np.random.shuffle(all_rows)
print(len(all_rows))
if deberta_grading:
df = pd.DataFrame(all_rows)
df = df.rename(columns={'input': 'text'})
df = add_deberta_grade(df)
df = df.rename(columns={'text': 'input'})
drop = True
if drop:
min_grade = 0.3
max_grade = np.inf
before_rows = df.shape[0]
df = df[df['grade_deberta'] >= min_grade]
df = df[df['grade_deberta'] <= max_grade]
after_rows = df.shape[0]
print("Dropped %d rows out of %d due to deberta grade" % (before_rows - after_rows, before_rows))
print("After DeBERTa grade")
print(df.describe())
all_rows = []
for i in range(df.shape[0]):
all_rows.append(
dict(
input=df['input'].iloc[i],
output=df['output'].iloc[i],
source=df['source'].iloc[i],
prompt_type=df['prompt_type'].iloc[i],
grade_deberta=df['grade_deberta'].iloc[i],
)
)
if save_json:
data_file = data_file + \
("_h2ogpt" if fixup_personality else "") + \
("_only" if only_personality else "") + \
("_graded" if deberta_grading else "") + \
("_llama2_chat" if prompt_type == "llama2" else "")
for i in range(len(all_rows)):
all_rows[i]['id'] = i
with open(data_file.lower().replace("/", "_") + ".json", "w") as f:
f.write(json.dumps(all_rows, indent=2))
return all_rows
def test_finalize_to_json():
df = pd.read_parquet('h2oGPT.cleaned.graded.human_bot.shorter.parquet')
df = df.rename(columns={'text': 'input'})
print("Number of high-quality human_bot interactions: %s" % df.shape[0], flush=True)
print("Adding open assistant data")
with open("openassistant_oasst1_h2ogpt_graded.json") as f:
open_assistant = json.loads(f.read())
df = pd.concat([df, pd.DataFrame(open_assistant)], axis=0)
def final_clean(df):
from better_profanity import profanity
profanity.load_censor_words_from_file("data/censor_words.txt")
df['profanity'] = parallel_apply(
df['input'],
lambda x: profanity.contains_profanity(x),
n_jobs=-1,
)
return df[(df['profanity'] == 0)].reset_index(drop=True)
print("Before cleaning: Number of final high-quality human_bot interactions: %s" % df.shape[0], flush=True)
df = final_clean(df)
print("After cleaning: Number of final high-quality human_bot interactions: %s" % df.shape[0], flush=True)
print(df.describe())
print(df.shape)
row_list = []
for i in range(df.shape[0]):
row_list.append(
dict(
input=df.loc[i, 'input'],
source=df.loc[i, 'source'],
prompt_type='plain',
)
)
np.random.seed(1234)
np.random.shuffle(row_list)
unhelpful = get_unhelpful_list()
row_list = [x for x in row_list if not any(u in x['input'] for u in unhelpful)]
for i in range(len(row_list)):
row_list[i]['id'] = i
row_list[i]['input'] = row_list[i]['input'].replace(" <bot>:", "\n<bot>:")
with open('h2ogpt-oig-oasst1-instruct-cleaned-v3.json', "w") as f:
f.write(json.dumps(row_list, indent=2))
def create_personality_data(prompt_type="llama2"):
questions = [
"What's your name?",
"What is your name?",
"What are you?",
"Who are you?",
"Do you have a name?",
"Who trained you?",
"Who created you?",
"Who made you?",
]
answers = [
"I'm h2oGPT, a large language model by H2O.ai.",
"I'm h2oGPT, a large language model by H2O.ai, the visionary leader in democratizing AI.",
"My name is h2oGPT. I'm a large language model by H2O.ai, the visionary leader in democratizing AI.",
"My name is h2oGPT. I'm a large language model trained by H2O.ai.",
"Hi! I'm h2oGPT, a large language model by H2O.ai.",
"Hi! I'm h2oGPT, a large language model by H2O.ai, the visionary leader in democratizing AI.",
]
help = [
"",
" How can I help you?",
" How may I assist you?",
" Nice to meet you.",
]
import itertools
rows = []
for pair in itertools.product(questions, answers, help):
rows.append(
dict(input=f"{pair[0]}", output=f"{pair[1]}{pair[2]}", prompt_type=prompt_type, source="H2O.ai")
)
for q, a in [
("What is H2O.ai?", "H2O.ai is a technology company that aims to democratize AI and make it accessible to a broader audience by simplifying the process of creating and deploying machine learning models."),
("What is h2o.ai?", "H2O.ai is a technology company that aims to democratize AI and make it accessible to a broader audience by simplifying the process of creating and deploying machine learning models."),
("What is H2O?", "H2O.ai is a technology company that aims to democratize AI and make it accessible to a broader audience by simplifying the process of creating and deploying machine learning models."),
("Who is h2o.ai?", "H2O.ai is a technology company that aims to democratize AI and make it accessible to a broader audience by simplifying the process of creating and deploying machine learning models."),
("who is h2o.ai?", "H2O.ai is a technology company that aims to democratize AI and make it accessible to a broader audience by simplifying the process of creating and deploying machine learning models."),
("who is h2o?", "H2O.ai is a technology company that aims to democratize AI and make it accessible to a broader audience by simplifying the process of creating and deploying machine learning models."),
("what is H2O.ai?", "H2O.ai is the visionary leader in democratizing AI."),
("who is H2O.ai?", "H2O.ai is the visionary leader in democratizing AI."),
("who is H2O?", "H2O.ai is the visionary leader in democratizing AI."),
("Who is h20?", "H2O.ai is the visionary leader in democratizing AI."),
]:
rows.append(dict(input=q, output=a, prompt_type=prompt_type, source='H2O.ai'))
print(len(rows))
with open("h2ogpt-personality.json", "w") as f:
f.write(json.dumps(rows, indent=2))
return rows
def test_check_stats_data():
filename = 'h2ogpt-oig-oasst1-instruct-cleaned-v3.json'
df = pd.read_json(filename)
# get word stats
df['char_count'] = df['input'].apply(lambda x: len(x))
import matplotlib.pyplot as plt
plt.figure(figsize=(10, 10))
plt.hist(df['char_count'], bins=100)
chars_avg = np.mean(df['char_count'])
chars_median = np.median(df['char_count'])
plt.title("char_count avg: %s median: %s" % (chars_avg, chars_median))
plt.savefig('chars_hist.png')
plt.close()
# get tokenize stats for random sample of 1000 rows
from finetune import generate_and_tokenize_prompt
from loaders import get_loaders, get_tokenizer
from functools import partial
llama_type = False
tokenizer_base_model = base_model = 'h2oai/h2ogpt-oasst1-512-20b'
model_loader, tokenizer_loader, conditional_type = (
get_loaders(model_name=base_model, reward_type=False, llama_type=llama_type))
local_files_only = False
resume_download = True
use_auth_token = False
tokenizer = get_tokenizer(tokenizer_loader, tokenizer_base_model, local_files_only, resume_download, use_auth_token)
prompt_type = 'plain' # trained with data already in human bot form
train_on_inputs = True
add_eos_token = False
cutoff_len = 512 # can choose 2048
generate_and_tokenize_prompt_fun = partial(generate_and_tokenize_prompt, prompt_type=prompt_type,
train_on_inputs=train_on_inputs, add_eos_token=add_eos_token,
cutoff_len=cutoff_len, tokenizer=tokenizer)
from datasets import load_dataset
data = load_dataset("json", data_files={"train": filename})
val_set_size = 0.90
train_val = data["train"].train_test_split(
test_size=val_set_size, shuffle=True, seed=42
)
train_data = train_val["train"]
train_data = train_data.shuffle().map(generate_and_tokenize_prompt_fun, num_proc=os.cpu_count())
df_tokens = pd.DataFrame([len(x) for x in train_data['input_ids']], columns=['token_count'])
plt.figure(figsize=(10, 10))
plt.hist(df_tokens['token_count'], bins=100)
token_avg = np.mean(df_tokens['token_count'])
token_median = np.median(df_tokens['token_count'])
plt.title("token_count with cutoff=%s avg: %s median: %s" % (cutoff_len, token_avg, token_median))
plt.savefig('token_hist_%s.png' % cutoff_len)
plt.close()
def get_unhelpful_list():
# base versions
unhelpful = ["I'm sorry, I didn't quite understand your question, could you please rephrase it?",
"I'm sorry, but I don't understand your question. Could you please rephrase it?",
"I'm sorry, I don't quite understand your question",
"I'm sorry, I don't know",
"I'm sorry, but I don't know",
"I don't know anything",
"I do not know",
"I don't know",
"I don't know how",
"I do not know how",
"Can you please explain what you mean",
"please explain what you mean",
"please explain",
"I'm sorry, but I don't know how to tell a story. Can you please explain what you mean by",
"I'm sorry but I don't understand what you mean",
"I don't understand",
"I don't have the ability",
"I do not have the ability",
"I do not have",
"I am a language model,",
"I am a large language model,",
"I do not understand your question. Can you please try to make it clearer?",
"I'm sorry, but as an AI language model",
"I apologize, but I cannot rephrase text that I cannot understand. Your post is difficult to read and follow.",
"I apologize, but I am not h2oGPT. I am a language model developed by H2O.ai. How may I help you?",
"Sorry, but I am not an actual Linux shell, nor am I capable of emulating one. I am an open source chat assistant and would be glad t",
"I apologize, but I cannot perform the task you have requested.",
"I'm sorry, I cannot perform this task as I am an AI language model and do not have access",
"I'm sorry, I'm not sure what you're asking for here.",
"I'm not sure what you are asking",
"You need to provide more context",
]
# reduced versions, with redundant parts, just to give context for where they came from
unhelpful += ["sorry, I didn't quite understand your question",
"I didn't quite understand your question",
"I didn't understand your question",
"I did not understand your question",
"I did not understand the question",
"could you please rephrase"
"could you rephrase"
"I do not understand your question.",
"I do not understand the question.",
"I do not understand that question.",
"Can you please try to make it clearer",
"Can you try to make it clearer",
"sorry, but as an AI language model",
"as an AI language model",
"I apologize, but I cannot",
"I cannot rephrase text",
"I cannot understand. Your post is difficult to read and follow."
"Your post is difficult to read and follow."
"I apologize, but I am",
"Sorry, but I am not ",
"nor am I capable",
"I am not capable of",
"I apologize, but I cannot perform the task you have requested",
"I cannot perform the task",
"I cannot complete the task",
"I'm sorry",
"I am sorry",
"do not have access",
"not sure what you're asking for",
"not sure what you are asking for",
"not sure what is being asked",
"I'm not sure what you are asking",
"not sure what you are asking",
"You need to provide more context",
"provide more context",
]
unhelpful += ["As a large language model",
"cannot provide any information",
"As an artificial intelligence I do not have the capability",
"As an artificial intelligence I don't have the capability",
"As an artificial intelligence I can't",
"As an artificial intelligence I cannot",
"I am sorry but I do not understand",
"Can you please explain",
"(sorry couldn't resist)",
"(sorry could not resist)",
" :)",
" ;)",
" :-)",
" ;-)",
" lol ",
"Thanks so much!!!",
"Thank You :)!!!",
"Please try not to repeat",
"I am an AI language model",
"I'm a AI assistant that",
"I'm an AI assistant that",
"I am an AI assistant that",
"etc.",
"etc.etc.",
"etc. etc.",
"etc etc",
]
return unhelpful
def test_check_unhelpful():
# file = '/home/jon/Downloads/openassistant_oasst1_h2ogpt_graded.json'
file = '/home/jon/Downloads/openassistant_oasst1_h2ogpt_grades.json'
# file = 'h2ogpt-oig-oasst1-instruct-cleaned-v2.json'
unhelpful = get_unhelpful_list()
# data = json.load(open(file, 'rt'))
df = pd.read_json(file)
use_reward_score_threshold = False
use_bleu_threshold = False
use_sentence_sim = True
from sacrebleu.metrics import BLEU
bleu = BLEU()
from nltk.translate.bleu_score import sentence_bleu
def get_bleu(actual, expected_list):
# return bleu.sentence_score(actual, expected_list).score
return sentence_bleu(expected_list, actual)
threshold = 0.0
if use_reward_score_threshold:
df = df[df['grade_deberta'] > threshold]
# back to as if original json load
data = df.to_dict(orient='records')
bads = {}
string_all = str(data)
for sub in unhelpful:
bads[sub] = string_all.count(sub)
bads = {k: v for k, v in bads.items() if v > 0}
import pprint
pp = pprint.PrettyPrinter(indent=4)
pp.pprint(bads)
total_bads = sum(list(bads.values()))
print('total_bads: %s' % total_bads, flush=True)
# check just bot
import re
convs = [[x.strip() for x in re.split(r'%s|%s' % (human, bot), y['input']) if x.strip()] for y in data]
humans = [[x for i, x in enumerate(y) if i % 2 == 0] for y in convs]
bots = [[x for i, x in enumerate(y) if i % 2 == 1] for y in convs]
# FIXME: apply back to json etc., just see for now
bleu_threshold = 0.9
if use_bleu_threshold:
bots = [[x for x in y if get_bleu(x, unhelpful) < bleu_threshold] for y in tqdm(bots)]
cosine_sim_threshold = 0.8
if use_sentence_sim:
# pip install sentence_transformers-2.2.2
from sentence_transformers import SentenceTransformer
# sent_model = 'bert-base-nli-mean-tokens'
# sent_model = 'nli-distilroberta-base-v2'
sent_model = 'all-MiniLM-L6-v2'
model = SentenceTransformer(sent_model)
sentence_embeddings = model.encode(unhelpful)
from sklearn.metrics.pairwise import cosine_similarity
bots = [x for x in tqdm(bots) if
np.max(cosine_similarity(model.encode(x), sentence_embeddings)) < cosine_sim_threshold]
bads_bots = {}
string_all = str(bots)
for sub in unhelpful:
bads_bots[sub] = string_all.count(sub)
bads_bots = {k: v for k, v in bads_bots.items() if v > 0}
import pprint
pp = pprint.PrettyPrinter(indent=4)
pp.pprint(bads_bots)
total_bads_bots = sum(list(bads_bots.values()))
print('threshold: %g use_bleu_threshold: %g total_bads_bots: %s total_bots: %s total_humans: %s' % (
threshold, use_bleu_threshold, total_bads_bots, len(bots), len(humans)), flush=True)
# assert len(bads) == 0, bads
assert len(bads_bots) == 0, bads_bots
def test_fortune2000_personalized():
row_list = []
import glob
if not os.path.isdir("wikitext"):
raise RuntimeError("download https://github.com/h2oai/h2ogpt/files/11423008/wikitext.zip and unzip")
for file in glob.glob("wikitext/*.txt"):
with open(file, "r") as f:
blob = f.read()
N = 512 * 4
row_list.extend([{'input': s, 'prompt_type': 'plain', 'source': "%s" % os.path.basename(file)}
for s in get_sentences(blob, N) if s])
personality = create_personality_data()
import copy
for i in range(10):
row_list.extend(copy.deepcopy(personality))
np.random.seed(123)
np.random.shuffle(row_list)
for i in range(len(row_list)):
row_list[i]['id'] = i
for i in range(len(row_list)):
assert row_list[i]['id'] == i
with open("h2ogpt-fortune2000-personalized.json", "w") as ff:
ff.write(json.dumps(row_list, indent=2))