File size: 11,249 Bytes
5d32603 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 | import ast
from collections import Counter
from concurrent.futures import ThreadPoolExecutor
from copy import deepcopy
from dataclasses import dataclass
import hashlib
import json
from pathlib import Path
import random
import re
import shutil
from huggingface_hub import hf_hub_download
from omegaconf import OmegaConf
import pyarrow.parquet as pq
from rich import print
from tqdm import tqdm
from functional_rl import v2
from hinted_rl.teacher_programs import canonical
from refined_v4.data import load_rows,write_rows,signature,atomic_json
from . import ROOT,TASKS
from .audit import authentic
from .data import annotate,documents,gold_passes
from .envs import Pool
from .named_data import styles_for
REPO='google-research-datasets/mbpp'
REVISION='4bb6404fdc6cacfda99d4ac4205087b89d32030c'
@dataclass
class Config:
source: str=str(ROOT/'variants/consistent-natural-names-pilot-v1')
root: str=str(ROOT/'variants/full-mbpp-seeds-pilot-v2')
tokens: int=2097152
benchmark_fraction: float=.5
seed: int=6719
workers: int=8
def native(raw: dict,task: str):
question=raw['text'].strip()
tree=ast.parse(raw['code'])
functions=[n for n in tree.body if isinstance(n,ast.FunctionDef)]
if len(functions)!=1 or any(not isinstance(n,(ast.Import,ast.ImportFrom,ast.FunctionDef)) for n in tree.body):
raise ValueError('Require one standalone entry function and imports')
setup=raw.get('test_setup_code') or ''
setup_tree=ast.parse(setup)
if any(not isinstance(n,(ast.Import,ast.ImportFrom)) for n in setup_tree.body):
raise ValueError('Setup must contain imports only')
tests=list(raw['test_list'])+list(raw.get('challenge_test_list') or [])
if len(tests)<3 or not question:raise ValueError('Require instruction and three published tests')
function=deepcopy(functions[0])
if ast.get_docstring(function) is not None:function.body.pop(0)
example=ast.parse(tests[0]).body
if not example or not isinstance(example[0],ast.Assert):raise ValueError('Require assertion examples')
expression=example[0].test
documentation=question+'\n\nExample:\n '+tests[0]
if task.startswith('humaneval'):
left,expected=ast.unparse(expression),'True'
if isinstance(expression,ast.Compare) and len(expression.ops)==1 and isinstance(expression.ops[0],ast.Eq):
try:
expected=repr(ast.literal_eval(expression.comparators[0]));left=ast.unparse(expression.left)
except (ValueError,TypeError):pass
documentation=question+'\n\n>>> '+left+'\n'+expected
if task.endswith('_docstring'):function.body.insert(0,ast.Expr(value=ast.Constant(value=documentation)))
imports=[];seen=set()
for node in setup_tree.body+[n for n in tree.body if isinstance(n,(ast.Import,ast.ImportFrom))]:
key=ast.dump(node)
if key not in seen:imports.append(deepcopy(node));seen.add(key)
reference=ast.unparse(ast.fix_missing_locations(ast.Module(body=imports+[function],type_ignores=[])))
if task=='humaneval':reference=''.join('# '+line+'\n' if line else '#\n' for line in documentation.splitlines())+reference
identity=str(raw['task_id'])
row=annotate(dict(task=task,source_id='mbpp-extra-'+identity,
group=hashlib.sha256(('full-mbpp/'+identity).encode()).hexdigest(),question=question,
reference=reference,answer=reference,tests=tests,setup=setup,
provenance=dict(source=REPO,revision=REVISION,upstream_config='full',upstream_split=raw['upstream_split'],
upstream_task_id=raw['task_id'],license='cc-by-4.0',sft_origin='additional_original_mbpp')))
if len(v2.ENC.encode_ordinary(v2.render(row)))>850:raise ValueError('Too long for few-shot packing')
return row
def build(config: Config):
source,root=Path(config.source),Path(config.root)
if (root/'dataset.json').exists():raise ValueError('Full-source dataset already frozen')
cache=ROOT/'inputs/mbpp-full';sanitized=set();raw=[]
for split in tqdm(('train','validation','test','prompt'),desc='Read pinned original MBPP files',colour='cyan'):
def download(flavor):
return hf_hub_download(REPO,flavor+'/'+split+'-00000-of-00001.parquet',repo_type='dataset',revision=REVISION,cache_dir=cache)
sanitized.update(pq.read_table(download('sanitized'),columns=['task_id'])['task_id'].to_pylist())
raw.extend(dict(row,upstream_split=split) for row in pq.read_table(download('full')).to_pylist())
blocked=set();questions=set();known_ids=set();stats=Counter()
verified=ROOT/'variants/verified-reference-pilot-v1'
practice=ROOT/'variants/clear-practice-rl-v1'
for task in tqdm(TASKS,desc='Index existing training and held-out solutions',colour='cyan'):
for split in ('sft','rl','calibration','dev','test'):
rows=load_rows(verified,task,split)
if split=='rl':rows+=load_rows(practice,task,'rl')
for row in rows:
match=re.match(r'mbpp-(\d+)(?:\D|$)',row['source_id'])
if match:known_ids.add(int(match.group(1)))
blocked.add(canonical(row['reference']))
if row['question']:questions.add(' '.join(row['question'].casefold().split()))
for item in raw:
if item['task_id'] not in sanitized|known_ids:continue
questions.add(' '.join(item['text'].casefold().split()))
try:blocked.add(canonical(item['code']))
except (SyntaxError,ValueError):pass
try:blocked.add(canonical(native(item,'mbpp')['reference']))
except (SyntaxError,ValueError,TypeError,KeyError):pass
candidates=[];seen=set(blocked)
for item in tqdm(sorted(raw,key=lambda r:r['task_id']),desc='Exclude original IDs and known solutions',colour='cyan',mininterval=5):
if item['task_id'] in sanitized:stats['original_sanitized_id']+=1;continue
if item['task_id'] in known_ids:stats['existing_task_id']+=1;continue
if ' '.join(item['text'].casefold().split()) in questions:stats['known_question']+=1;continue
try:
row=native(item,'mbpp');keys={canonical(item['code']),canonical(row['reference'])}
if keys&seen:stats['known_solution']+=1;continue
native_rows={task:native(item,task) for task in TASKS}
except (ValueError,SyntaxError,TypeError,KeyError):stats['format_or_context']+=1;continue
seen.update(keys);candidates.append((item,native_rows))
with ThreadPoolExecutor(max_workers=config.workers) as pool:
passed=list(tqdm(pool.map(gold_passes,[rows['mbpp'] for item,rows in candidates]),total=len(candidates),
desc='Check additional published MBPP tests once',colour='green',mininterval=5))
accepted=[item for item,ok in zip(candidates,passed,strict=True) if ok]
stats['published_test_failure']=len(candidates)-len(accepted)
if len(accepted)<100:raise ValueError('Insufficient additional original problems')
is_rl=lambda item:int(hashlib.sha256(f'{config.seed}/{item["task_id"]}'.encode()).hexdigest()[:8],16)%5==0
report=json.loads((source/'dataset.json').read_text())
report.update(revision='full-mbpp-seeds-v2',source=str(source),sft_tokens=config.tokens,seed=config.seed,
benchmark_draw_probability=config.benchmark_fraction,upstream=dict(repo=REPO,revision=REVISION,license='cc-by-4.0'),
extra_source_filtering=dict(stats),extra_sft_ids=[item['task_id'] for item,rows in accepted if not is_rl(item)],
extra_rl_ids=[item['task_id'] for item,rows in accepted if is_rl(item)],
extra_source_policy='Additional original full-MBPP problems only. Exclude EVERY original sanitized task ID and every known existing upstream task ID. Also exclude original RAW upstream solutions/questions for those IDs, alongside current rewritten references. Deterministic80/20 SFT/RL split. Every admitted reference passes its published tests. No rewritten expected values.',
sft_policy='Sample body band from ORIGINAL authentic SFT-training frequencies, then source origin at the declared benchmark fraction, then source group, then a natural local-name variant. Two to four complete native examples per document.',
rl_policy='Uniform missing length1..16 first, then uniform original-benchmark/additional-MBPP origin, then eligible source. No synthetic practice contracts in this variant. Calibration/development/test stay byte-identical.')
for task in TASKS:
shutil.copytree(source/'data'/task,root/'data'/task,dirs_exist_ok=True)
rows=load_rows(source,task,'sft')
old_auth={r['group']:r for r in load_rows(verified,task,'sft') if authentic(r)}
bands=Counter(r['provenance']['difficulty'] for r in old_auth.values())
weights={band:count/sum(bands.values()) for band,count in bands.items()}
original=load_rows(verified,task,'rl')
rl=[dict(row,sampling_origin='benchmark') for row in original]
for item,native_rows in tqdm(accepted,desc=task+' original-source styles',colour='cyan',mininterval=5):
row=native_rows[task]
if is_rl(item):
try:rl.append(annotate(dict(row,demonstrations=original[0]['demonstrations'],sampling_origin='full_mbpp')))
except ValueError:stats['rl_context_rejected']+=1
else:
variants,filtered=styles_for(row,4,config.seed);rows.extend(variants)
assert not {r['group'] for r in rows}&{r['group'] for r in rl}
Pool(rl)
write_rows(root/'data'/task/'sft.jsonl.gz',rows);write_rows(root/'data'/task/'rl.jsonl.gz',rl)
corpus=documents(rows,config.tokens,random.Random(config.seed),config.benchmark_fraction,True,weights)
target=root/'sft'/task/'train.jsonl.gz';write_rows(target,corpus)
info=report['tasks'][task]
info.update(rows=dict(info['rows'],sft=len(rows),rl=len(rl)),documents=len(corpus),sft_sha256=signature(target),
source_groups=len({r['group'] for r in rows}),sft_body_band_target=weights,
admitted_extra_rl_ids=[r['provenance']['upstream_task_id'] for r in rl if r['sampling_origin']=='full_mbpp'],
unique_sft_programs_used=len({r['source_id'] for doc in corpus for r in doc['examples']}),
document_example_counts=dict(Counter(len(d['examples']) for d in corpus)))
for fold in ('sft','rl'):
relative='data/'+task+'/'+fold+'.jsonl.gz';info['assets'][relative]=signature(root/relative)
pool_rows=rows if fold=='sft' else rl
info.setdefault('hint_coverage',{})[fold]={str(n):sum(n in r['hint_levels'] for r in pool_rows) for n in range(1,17)}
for fold in ('calibration','dev','test'):
assert signature(root/'data'/task/(fold+'.jsonl.gz'))==signature(source/'data'/task/(fold+'.jsonl.gz'))
atomic_json(root/'build-progress.json',report)
report['extra_source_filtering']=dict(stats)
atomic_json(root/'dataset.json',report)
print('[green]Additional original MBPP SFT/RL sources ready[/green]',len(report['extra_sft_ids']),len(report['extra_rl_ids']),flush=True)
if __name__=='__main__':
build(OmegaConf.to_object(OmegaConf.merge(OmegaConf.structured(Config),OmegaConf.from_cli())))
|