File size: 3,499 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 | from concurrent.futures import ThreadPoolExecutor
from dataclasses import dataclass
import json
from pathlib import Path
import shutil
from omegaconf import OmegaConf
from rich import print
from tqdm import tqdm
from refined_v4.data import load_rows,write_rows,atomic_json,signature
from hinted_rl.teacher_programs import canonical
from . import ROOT,TASKS
from .audit import authentic
from .data import annotate,gold_passes
from .envs import Pool
@dataclass
class Config:
root: str=str(ROOT/'variants/joint-train-rl-v1')
source: str=str(ROOT/'variants/clear-full-practice-rl-v1')
workers: int=8
def build(config: Config):
source,root=Path(config.source),Path(config.root)
if (root/'dataset.json').exists():raise ValueError('Joint training environment already frozen')
report=json.loads((source/'dataset.json').read_text())
report.update(revision='joint-benchmark-train-rl-v1',source=str(source),
training_overlap='SFT and RL deliberately share original benchmark TRAINING problems. Calibration/development/test remain held out and byte-identical.',
rl_policy='Uniform missing length 1..16, then uniform origin: original RL benchmark, SFT-training benchmark, or independently tested fresh practice. Then sample an eligible source.')
for task in TASKS:
shutil.copytree(source/'data'/task,root/'data'/task,dirs_exist_ok=True)
base=load_rows(source,task,'rl')
groups={row['group'] for row in base}
demos=base[0]['demonstrations']
demo_groups={row['group'] for row in demos}
heldout=[row for split in ('calibration','dev','test') for row in load_rows(source,task,split)]
blocked={canonical(r['reference']) for r in heldout}
blocked_groups={r['group'] for r in heldout}
candidates=[]
for row in load_rows(source,task,'sft'):
if not authentic(row) or row['group'] in groups|demo_groups|blocked_groups or canonical(row['reference']) in blocked:continue
try:
candidates.append(annotate(dict(row,demonstrations=demos,sampling_origin='sft_benchmark')))
except ValueError:continue
with ThreadPoolExecutor(max_workers=config.workers) as pool:
valid=list(tqdm(pool.map(gold_passes,candidates),total=len(candidates),
desc=task+' training benchmark reward checks',colour='green',mininterval=5))
admitted=[row for row,ok in zip(candidates,valid,strict=True) if ok]
if len(admitted)<20:raise ValueError('Too few additional benchmark training problems')
rows=base+admitted
assert not {r['group'] for r in rows}&blocked_groups
Pool(rows)
target=root/'data'/task/'rl.jsonl.gz';write_rows(target,rows)
report['tasks'][task]['rows']['rl']=len(rows)
report['tasks'][task]['sft_benchmark_practice']=len(admitted)
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]Complete benchmark TRAIN pool assembled for RL; held-out data unchanged[/green]',flush=True)
if __name__=='__main__':
build(OmegaConf.to_object(OmegaConf.merge(OmegaConf.structured(Config),OmegaConf.from_cli())))
|