File size: 4,641 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 | import ast
import hashlib
import random
from functional_rl import v2
REVISION='refined-code-v1'
ENC=v2.ENC
TASKS=('mbpp','mbpp_docstring','humaneval','humaneval_docstring')
def protected_prefix(row: dict) -> str:
if row.get('protected_prefix'):
return row['protected_prefix']
tree=ast.parse(row['reference'])
functions=[node for node in tree.body if isinstance(node,ast.FunctionDef)]
if len(functions)!=1:
raise ValueError('Expected exactly one top-level entry-point function')
function=functions[0]
statements=function.body[1:] if ast.get_docstring(function) is not None else function.body
if not statements:
raise ValueError('The entry point needs an implementation')
return ''.join(row['reference'].splitlines(keepends=True)[:statements[0].lineno-1])
def make_episode(row: dict, level: int):
if type(level) is not int or not 1<=level<=16 or row['task'] not in TASKS:
raise ValueError('Expected a code task and an integer missing length in 1..16')
groups={row['group']}
demos=row.get('demonstrations',())
for demo in demos:
if demo['task']!=row['task'] or demo['group'] in groups:
raise ValueError('Demonstrations must use distinct same-task source groups')
groups.add(demo['group'])
context=''.join(v2.render(d).rstrip()+'\n\n' for d in demos)
header=context+v2.header(row)
document=header+row['reference']
ids=ENC.encode_ordinary(document)
prompt_ids,completion_ids=ids[:-level],ids[-level:]
prompt,completion=ENC.decode(prompt_ids),ENC.decode(completion_ids)
if len(ids)>1024 or not prompt.startswith(header+row['protected_prefix']) or prompt+completion!=document:
raise ValueError('Missing tokens must stay in the implementation and preserve the instruction/signature')
return v2.Episode(row['task'],row['source_id'],row['group'],level,prompt,completion,
tuple(prompt_ids),tuple(completion_ids),prompt[len(header):],row['reference'],row['answer'],
tuple(row['tests']),row.get('setup',''),version=REVISION,question=row['question'],
provenance=(dict(row['provenance'],rl_origin=row['sampling_origin']) if row.get('sampling_origin') else row['provenance']),
demonstrations=tuple(dict(task=d['task'],source_id=d['source_id'],
group=d['group'],text=v2.render(d),split='sft') for d in demos))
def available_hints(row: dict) -> list[int]:
demos=row.get('demonstrations',())
groups=[row['group']]+[d['group'] for d in demos]
if row['task'] not in TASKS or len(set(groups))!=len(groups) or any(d['task']!=row['task'] for d in demos):
return []
context=''.join(v2.render(d).rstrip()+'\n\n' for d in demos)
header=context+v2.header(row)
document=header+row['reference']
ids=ENC.encode_ordinary(document)
if len(ids)>1024:
return []
protected=header+row['protected_prefix']
return [level for level in range(1,min(16,len(ids)-1)+1)
if ENC.decode(ids[:-level]).startswith(protected)
and ENC.decode(ids[:-level])+ENC.decode(ids[-level:])==document]
class Pool:
def __init__(self, rows: list[dict]):
self.by_hint={n:[r for r in rows if n in r['hint_levels']] for n in range(1,17)}
if any(not rows for rows in self.by_hint.values()):
raise ValueError('Each hint length requires eligible benchmark sources')
origins={row.get('sampling_origin') for row in rows}
if None in origins and len(origins)>1:
raise ValueError('An origin-balanced RL pool must label every source')
self.by_origin=({n:{origin:[r for r in eligible if r['sampling_origin']==origin]
for origin in sorted(origins) if any(r['sampling_origin']==origin for r in eligible)}
for n,eligible in self.by_hint.items()} if None not in origins else None)
def sample(self, rng: random.Random):
level=rng.randint(1,16)
eligible=self.by_hint[level]
if self.by_origin is not None:
groups=self.by_origin[level]
eligible=groups[rng.choice(list(groups))]
return make_episode(rng.choice(eligible),level)
def fixed_panel(rows: list[dict], queries: int, seed: int):
if queries<16 or queries%16:
raise ValueError('Evaluation draws must be a positive multiple of 16')
pool=Pool(rows)
ordered={n:sorted(pool.by_hint[n],key=lambda row:hashlib.sha256(
f'{seed}/{n}/{row["group"]}'.encode()).hexdigest()) for n in range(1,17)}
return [(ordered[i%16+1][(i//16)%len(ordered[i%16+1])],i%16+1) for i in range(queries)]
|