| import hashlib |
| import random |
|
|
| from functional_rl import v2 |
|
|
| from . import REVISION |
|
|
| ENC = v2.ENC |
|
|
|
|
| def make_episode(row: dict, level: int): |
| if type(level) is not int or not 1 <= level <= 16: |
| raise ValueError('Missing-token length must be an integer in [1, 16]') |
| demos = row.get('demonstrations', ()) |
| context = ''.join(d['reference']+'\n\n' for d in demos) |
| document = context+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) |
| protected = context+row['protected_prefix'] |
| if len(ids)>1024 or not prompt.startswith(protected) or prompt+completion != document: |
| raise ValueError('Suffix must remove only the function body, preserving its complete signature and docstring') |
| return v2.Episode('basic_python', row['source_id'], row['group'], level, |
| prompt, completion, tuple(prompt_ids), tuple(completion_ids), |
| prompt[len(context):], row['reference'], row['answer'], tuple(row['tests']), '', |
| version=REVISION, question=row['question'], provenance=row['provenance'], |
| demonstrations=tuple(dict(task='basic_python',source_id=d['source_id'],group=d['group'], |
| text=d['reference'],split='sft') for d in demos)) |
|
|
|
|
| def available_hints(row: dict) -> list[int]: |
| result = [] |
| for level in range(1,17): |
| try: |
| make_episode(row,level) |
| result.append(level) |
| except ValueError: |
| pass |
| return result |
|
|
|
|
| class Pool: |
| def __init__(self, rows: list[dict]): |
| self.by_hint = {level: {} for level in range(1,17)} |
| for row in rows: |
| for level in row['hint_levels']: |
| self.by_hint[level].setdefault(row['provenance']['family'],[]).append(row) |
| if any(not groups for groups in self.by_hint.values()): |
| raise ValueError('At least one function must support each missing-token length') |
|
|
| def sample(self, rng: random.Random): |
| level = rng.randint(1,16) |
| families = self.by_hint[level] |
| family = rng.choice(sorted(families)) |
| return make_episode(rng.choice(families[family]),level) |
|
|
|
|
| def fixed_panel(rows: list[dict], queries: int, seed: int): |
| if queries < 16 or queries % 16: |
| raise ValueError('Use a positive multiple of 16 evaluation questions') |
| pool = Pool(rows) |
| panel, counts = [], {} |
| for index in range(queries): |
| level = index%16+1 |
| families = sorted(pool.by_hint[level]) |
| family = families[(index//16)%len(families)] |
| candidates = sorted(pool.by_hint[level][family],key=lambda row: hashlib.sha256( |
| f'{seed}/{level}/{row["group"]}'.encode()).hexdigest()) |
| key=(level,family) |
| position=counts.get(key,0) |
| row=candidates[position%len(candidates)] |
| counts[key]=position+1 |
| panel.append((row,level)) |
| return panel |
|
|
