File size: 7,344 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 | import ast
from collections import Counter
from dataclasses import dataclass
import hashlib
import json
from pathlib import Path
import random
import shutil
from omegaconf import OmegaConf
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,atomic_json,signature
from refined_v4.fast_generate import CodeExample,code_row
from . import ROOT,TASKS
from .audit import authentic
from .data import annotate,documents
@dataclass
class Config:
root: str=str(ROOT/'variants/fresh-native-full-v1')
source: str=str(ROOT/'variants/clear-practice-rl-v1')
tokens: int=16777216
benchmark_fraction: float=.1
seed: int=7169
def convert(raw: dict,task: str):
tree=ast.parse(raw['reference'])
functions=[n for n in tree.body if isinstance(n,ast.FunctionDef)]
if len(functions)!=1:raise ValueError('One entry-point function required')
function=functions[0]
if ast.get_docstring(function) is not None:raise ValueError('Expected an implementation without a documentation wrapper')
lines=raw['reference'].splitlines(keepends=True)
start=function.body[0].lineno-1
item=CodeExample(index=0,instruction=raw['question'],signature=''.join(lines[:start]).strip(),
body=''.join(lines[start:]),example=raw['tests'][0])
row=code_row(task,item,raw['provenance'].get('response_id',''))
if row is None:raise ValueError('Native formatting rejected the source')
if task.startswith('humaneval'):
assertion=ast.parse(item.example).body
if len(assertion)!=1 or not isinstance(assertion[0],ast.Assert):raise ValueError('Expected a single assertion example')
documentation=item.instruction+'\n\n>>> '+ast.unparse(assertion[0].test)+'\nTrue'
if task.endswith('_docstring'):
literal=documentation.replace('\\','\\\\').replace('"""','\\"\\"\\"')
prefix=item.signature+'\n """'+literal.replace('\n','\n ')+'\n """\n'
else:
prefix=''.join('# '+line+'\n' if line else '#\n' for line in documentation.splitlines())+item.signature+'\n'
body=raw['reference'][len(''.join(lines[:start])):]
row.update(reference=prefix+body,answer=prefix+body,protected_prefix=prefix)
identity=hashlib.sha256((raw['group']+raw['reference']).encode()).hexdigest()
row.update(source_id='fresh-contract/'+identity,group=raw['group'],tests=raw['tests'],setup=raw.get('setup',''))
row['provenance']=dict(raw['provenance'],sft_origin='cached_api_contract',format_revision='native-code-doctest-v1')
row=annotate(row)
if len(v2.ENC.encode_ordinary(v2.render(row)))>900:raise ValueError('Source too long for complete few-shot packing')
return row
def build(config: Config):
root,source=Path(config.root),Path(config.source)
if (root/'dataset.json').exists():raise ValueError('Fresh dataset already frozen')
root.mkdir(parents=True,exist_ok=True)
original={task:load_rows(source,task,'sft') for task in TASKS}
blocked_groups,blocked_code,blocked_questions=set(),set(),set()
for task in TASKS:
for split in ('rl','calibration','dev','test'):
for row in load_rows(source,task,split):
blocked_groups.add(row['group']);blocked_code.add(canonical(row['reference']))
if row['question']:blocked_questions.add(row['question'])
base=original['mbpp']
retained=[row for row in base if not authentic(row)]
seen_groups={r['group'] for r in base}|blocked_groups
seen_code={canonical(r['reference']) for r in retained}|blocked_code
selected=list(retained)
counts=Counter()
cache=Path('/nlp/scr/dhei/tiny-rl/experiments/refined-tasks-v4/fast-api/mbpp/rows.jsonl')
for line in tqdm(cache.open(),desc='Expand the native SFT source bank',colour='cyan',mininterval=10):
row=json.loads(line)
if not row['source_id'].startswith('fast-api/') or row['group'] in seen_groups or row['question'] in blocked_questions:continue
try:
fingerprint=canonical(row['reference'])
if fingerprint in seen_code:continue
converted=convert(row,'mbpp')
except (SyntaxError,ValueError,KeyError):
counts['format_or_context']+=1
continue
selected.append(row)
seen_groups.add(row['group']);seen_code.add(fingerprint)
counts['additional_contracts']+=1
report=json.loads((source/'dataset.json').read_text())
report.update(revision='fresh-native-code-v1',source=str(source),sft_tokens=config.tokens,seed=config.seed,
benchmark_draw_probability=config.benchmark_fraction,source_expansion=dict(counts),
sft_policy='Use the existing clear source bank plus unused cached API contracts; deduplicate source groups and canonical code against RL and held-out pools. Uniform source-origin selection with the declared benchmark probability. Full native programs, 2–4 independent examples/document.',
humaneval_synthetic_format='Native comments or docstrings with valid doctest-style boolean examples; no wrapper tokens.',
validation='SFT syntax/context checks only. The existing independently tested RL bank and original held-out bytes stay unchanged.')
for task in TASKS:
rows=[r for r in original[task] if authentic(r)]
seen={canonical(r['reference']) for r in rows}|blocked_code
for raw in tqdm(selected,desc=task+' native sources',colour='cyan',mininterval=10):
try:
row=convert(raw,task)
fingerprint=canonical(row['reference'])
if fingerprint in seen:continue
seen.add(fingerprint)
rows.append(row)
except (SyntaxError,ValueError,KeyError):continue
shutil.copytree(source/'data'/task,root/'data'/task,dirs_exist_ok=True)
write_rows(root/'data'/task/'sft.jsonl.gz',rows)
corpus=documents(rows,config.tokens,random.Random(config.seed),config.benchmark_fraction)
target=root/'sft'/task/'train.jsonl.gz';write_rows(target,corpus)
item=report['tasks'][task]
item.update(rows=dict(item['rows'],sft=len(rows)),documents=len(corpus),source_groups=len({r['group'] for r in rows}),
document_example_counts=dict(Counter(len(d['examples']) for d in corpus)),sft_sha256=signature(target),
unique_sft_programs_used=len({e['source_id'] for d in corpus for e in d['examples']}),
source_bank_rendered_tokens=sum(len(v2.ENC.encode_ordinary(v2.render(r))) for r in rows))
item['assets']['data/'+task+'/sft.jsonl.gz']=signature(root/'data'/task/'sft.jsonl.gz')
item.setdefault('hint_coverage',{})['sft']={str(n):sum(n in r['hint_levels'] for r in rows) for n in range(1,17)}
for split in ('rl','calibration','dev','test'):
assert signature(root/'data'/task/(split+'.jsonl.gz'))==signature(source/'data'/task/(split+'.jsonl.gz'))
atomic_json(root/'build-progress.json',report)
atomic_json(root/'dataset.json',report)
print('[green]Larger native SFT bank built; RL and held-out data unchanged[/green]',flush=True)
if __name__=='__main__':
build(OmegaConf.to_object(OmegaConf.merge(OmegaConf.structured(Config),OmegaConf.from_cli())))
|