File size: 11,631 Bytes
dfdaf38 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 | import ast
import asyncio
from concurrent.futures import ThreadPoolExecutor
from dataclasses import dataclass
import hashlib
import heapq
import json
from pathlib import Path
import shutil
from openai import AsyncOpenAI,DefaultAioHttpClient
from omegaconf import OmegaConf
from pydantic import BaseModel
from rich import print
from tqdm import tqdm
from functional_rl import v2
from hinted_rl.code import execute
from hinted_rl.teacher_programs import canonical
from refined_v4.data import load_rows,write_rows,atomic_json,signature
from refined_v4.fast_generate import CodeExample,code_row,Limiter
from . import ROOT,TASKS
from .data import annotate
from .envs import available_hints,Pool
@dataclass
class Config:
root: str=str(ROOT/'variants/clear-practice-rl-v1')
source: str=str(ROOT/'variants/clear-reference-v1')
candidates: int=1536
target: int=1024
concurrency: int=64
workers: int=16
seed: int=7169
class TestSet(BaseModel):
identity: str
tests: list[str]
class Batch(BaseModel):
items: list[TestSet]
def candidates(config: Config):
root,source=Path(config.root),Path(config.source)
path=root/'candidates.jsonl'
if path.exists():return [json.loads(line) for line in path.read_text().splitlines()]
root.mkdir(parents=True,exist_ok=True)
blocked_groups,blocked_code,blocked_questions=set(),set(),set()
for task in TASKS:
for split in ('sft','rl','calibration','dev','test'):
for row in load_rows(source,task,split):
blocked_groups.add(row['group'])
blocked_questions.add(row['question'])
blocked_code.add(canonical(row['reference']))
cache=Path('/nlp/scr/dhei/tiny-rl/experiments/refined-tasks-v4/fast-api/mbpp/rows.jsonl')
heap=[]
fingerprints=set()
seen=set()
for line in tqdm(cache.open(),desc='Select fresh RL contracts',colour='cyan',mininterval=10):
row=json.loads(line)
group=row['group']
if not row['source_id'].startswith('fast-api/') or group in blocked_groups or group in seen or row['question'] in blocked_questions:
continue
seen.add(group)
rank=int(hashlib.sha256((str(config.seed)+group).encode()).hexdigest()[:15],16)
if len(heap)==config.candidates and rank>=-heap[0][0]:continue
try:
tree=ast.parse(row['reference'])
functions=[n for n in tree.body if isinstance(n,ast.FunctionDef)]
fingerprint=canonical(row['reference'])
if len(functions)!=1 or ast.get_docstring(functions[0]) is not None or fingerprint in blocked_code or fingerprint in fingerprints:continue
if not row['reference'].isascii() or len(v2.ENC.encode_ordinary(row['question']+row['reference']))>550:continue
key=hashlib.sha256((group+row['reference']).encode()).hexdigest()
value=(-rank,key,dict(row,identity=key),fingerprint)
if len(heap)<config.candidates:heapq.heappush(heap,value)
else:
removed=heapq.heapreplace(heap,value)
fingerprints.remove(removed[3])
fingerprints.add(fingerprint)
except (SyntaxError,ValueError):continue
rows=[]
used=set()
for rank,key,row,fingerprint in sorted(heap,reverse=True):
fingerprint=canonical(row['reference'])
if fingerprint in used:continue
used.add(fingerprint)
rows.append(row)
path.write_text(''.join(json.dumps(row)+'\n' for row in rows))
return rows
async def generate(config: Config,rows: list[dict]):
root=Path(config.root)
path=root/'tests.jsonl'
saved=[json.loads(line) for line in path.read_text().splitlines()] if path.exists() else []
done={r['identity']:r for r in saved}
pending=[r for r in rows if r['identity'] not in done]
limiter=Limiter(6000,3000000)
semaphore=asyncio.Semaphore(config.concurrency)
stream=path.open('a',buffering=1)
progress=tqdm(total=len(pending),desc='Generate independent RL tests',colour='cyan',mininterval=5)
async with AsyncOpenAI(http_client=DefaultAioHttpClient(),max_retries=0,timeout=90.) as client:
async def one(batch):
text=('For each Python exercise, write five distinct direct assert tests for its stated contract. '
'Use the exact function name, deterministic literal arguments and literal expected values. '
'Cover ordinary cases and valid empty, boundary or tie cases. Do not reuse the given example. '
'Do not use imports, helper functions, comprehensions, random inputs, oracle calls or reference implementations inside tests. '
'Each assert must call the target function exactly once. Return each identity unchanged.\n'+
json.dumps([dict(identity=str(index),question=r['question'],reference=r['reference'],example=r['tests'][0])
for index,r in enumerate(batch)]))
async with semaphore:
for attempt in range(3):
ticket=await limiter.acquire(len(v2.ENC.encode_ordinary(text))+1400)
try:
response=await client.responses.parse(model='gpt-6-luna',input=text,text_format=Batch,
reasoning={'effort':'none'},max_output_tokens=2200,store=False)
usage=response.usage.model_dump()
await limiter.settle(ticket,usage['total_tokens'])
if response.output_parsed is None:raise ValueError('Missing structured test cases')
lookup={str(index):r for index,r in enumerate(batch)}
accepted=[]
for value in response.output_parsed.items:
if value.identity not in lookup:raise ValueError('Unknown test identity')
tests=list(dict.fromkeys(value.tests))
if len(tests)<4:raise ValueError('At least four distinct tests required')
function=next(n for n in ast.parse(lookup[value.identity]['reference']).body if isinstance(n,ast.FunctionDef))
for test in tests:
parsed=ast.parse(test)
if len(parsed.body)!=1 or not isinstance(parsed.body[0],ast.Assert):raise ValueError('Expected a direct assert')
calls=[n for n in ast.walk(parsed) if isinstance(n,ast.Call) and isinstance(n.func,ast.Name) and n.func.id==function.name]
if len(calls)!=1:raise ValueError('Each assert must call its target once')
accepted.append(dict(identity=lookup[value.identity]['identity'],tests=tests,response_id=response.id))
if {r['identity'] for r in accepted}!={r['identity'] for r in batch}:raise ValueError('Missing test identities')
for record in accepted:
record['usage_batch']=usage
stream.write(json.dumps(record)+'\n')
done[record['identity']]=record
break
except Exception as error:
await limiter.settle(ticket,0)
if attempt==2:print('[yellow]Test generation failed[/yellow]',type(error).__name__,flush=True)
else:await asyncio.sleep(attempt+1)
progress.update(len(batch))
try:await asyncio.gather(*(one(pending[i:i+4]) for i in range(0,len(pending),4)))
finally:stream.close();progress.close()
return done
def checked(row: dict,testset: dict):
tests=list(dict.fromkeys(row['tests']+testset['tests']))
return execute(row['reference'],tuple(tests),row.get('setup',''))
def build(config: Config,rows: list[dict],tests: dict):
root,source=Path(config.root),Path(config.source)
if (root/'dataset.json').exists():raise ValueError('RL environment is already frozen')
pending=[r for r in rows if r['identity'] in tests]
with ThreadPoolExecutor(max_workers=config.workers) as pool:
valid=list(tqdm(pool.map(checked,pending,[tests[r['identity']] for r in pending]),total=len(pending),
desc='Verify RL reference and independent tests once',colour='green',mininterval=5))
selected=[r for r,ok in zip(pending,valid,strict=True) if ok][:config.target]
if len(selected)<config.target:raise ValueError(f'Only {len(selected)} reward contracts verified; need {config.target}')
report=json.loads((source/'dataset.json').read_text())
report.update(revision='clear-practice-rl-v1',source=str(source),practice_sources=len(selected),
practice_candidate_sources=len(rows),practice_tested=len(pending),practice_passed=sum(valid),
rl_policy='Uniform missing length, then uniform benchmark/practice origin, then eligible contract. Fresh practice groups/questions/canonical code excluded from all existing SFT/RL/dev/test pools. SFT and held-out bytes unchanged.')
for task in TASKS:
shutil.copytree(source/'data'/task,root/'data'/task,dirs_exist_ok=True)
base=load_rows(source,task,'rl')
practice=[]
for row in selected:
tree=ast.parse(row['reference'])
function=next(n for n in tree.body if isinstance(n,ast.FunctionDef))
lines=row['reference'].splitlines(keepends=True)
start=function.body[0].lineno-1
item=CodeExample(index=0,instruction=row['question'],signature=''.join(lines[:start]).strip(),
body=''.join(lines[start:]),example=row['tests'][0])
converted=code_row(task,item,row['provenance']['response_id'])
if converted is None:raise ValueError('Fresh RL contract failed native formatting')
converted.update(source_id='rl-contract/'+row['identity'],group=row['group'],
tests=list(dict.fromkeys(row['tests']+tests[row['identity']]['tests'])),
demonstrations=base[0]['demonstrations'],sampling_origin='practice')
converted['provenance'].update(tests_origin='Original example plus independent gpt-6-luna edge-case assertions; all passed unchanged reference.',
tests_response_id=tests[row['identity']]['response_id'],source='Fresh synthetic RL contract',source_family='synthetic_practice')
converted=annotate(converted)
converted['hint_levels']=available_hints(converted)
if converted['hint_levels']:practice.append(converted)
combined=[dict(r,sampling_origin='benchmark') for r in base]+practice
Pool(combined)
target=root/'data'/task/'rl.jsonl.gz'
write_rows(target,combined)
report['tasks'][task]['rows']['rl']=len(combined)
report['tasks'][task]['assets'][str(target.relative_to(root))]=signature(target)
for split in ('sft','calibration','dev','test'):
assert signature(root/'data'/task/(split+'.jsonl.gz'))==signature(source/'data'/task/(split+'.jsonl.gz'))
(root/'sft').symlink_to(source/'sft',target_is_directory=True)
atomic_json(root/'dataset.json',report)
print('[green]Fresh, verified practice environments built with unchanged SFT and held-out data[/green]',flush=True)
def main(config: Config):
rows=candidates(config)
tests=asyncio.run(generate(config,rows))
build(config,rows,tests)
if __name__=='__main__':
main(OmegaConf.to_object(OmegaConf.merge(OmegaConf.structured(Config),OmegaConf.from_cli())))
|