| """Static extraction only. Never executes dataset code on the host.""" |
| import ast, gzip, hashlib, json, re, urllib.request |
| from collections import Counter |
| from pathlib import Path |
| import pyarrow.parquet as pq |
| from transformers import AutoTokenizer |
| ROOT=Path(__file__).resolve().parent |
| REV='356bb069eee815daa6e23e9a282eeefe1490ad44' |
| MODEL_REV='47e169305d2e8376be1d31e765533382721b2cc1' |
| def digest(x): return hashlib.sha256(x.encode()).hexdigest() |
| def write(name, rows): (ROOT/name).write_text(''.join(json.dumps(r)+'\n' for r in rows)) |
| def extract(row): |
| blocks=re.findall(r'```(?:python|py)?\s*\n(.*?)```',row['response'],re.S) |
| if len(blocks)!=1: raise ValueError('not_one_code_block') |
| try: tree=ast.parse(blocks[0]) |
| except SyntaxError: raise ValueError('syntax_error') |
| if len(tree.body)!=1 or not isinstance(tree.body[0],ast.FunctionDef): raise ValueError('not_one_standalone_function') |
| fn=tree.body[0] |
| if fn.decorator_list: raise ValueError('decorated_function') |
| if any(isinstance(n,(ast.Import,ast.ImportFrom,ast.Global,ast.Nonlocal,ast.ClassDef,ast.AsyncFunctionDef)) for n in ast.walk(fn)): raise ValueError('complex_or_external_dependencies') |
| calls={n.func.id for n in ast.walk(fn) if isinstance(n,ast.Call) and isinstance(n.func,ast.Name)} |
| if calls & {'open','exec','eval','compile','__import__','input','globals','locals'}: raise ValueError('io_or_dynamic_execution') |
| |
| if ast.get_docstring(fn) is not None: fn.body=fn.body[1:] |
| if not fn.body: raise ValueError('empty_function') |
| body='\n'.join(ast.unparse(n) for n in fn.body) |
| signature=ast.unparse(fn).splitlines()[0] |
| prompt=signature+'\n '+repr(row['instruction'])+'\n' |
| completion=''.join(' '+line+'\n' if line else '\n' for line in body.splitlines()) |
| return prompt,completion,digest(ast.dump(ast.Module(body=fn.body,type_ignores=[]),include_attributes=False)) |
|
|
| def main(): |
| tok=AutoTokenizer.from_pretrained('EleutherAI/gpt-j-6b',revision=MODEL_REV) |
| url='https://raw.githubusercontent.com/openai/human-eval/b9453e8f1865df0a0057a1383767caa400acfbcf/data/HumanEval.jsonl.gz' |
| benchmark_file=ROOT/'humaneval-reference.jsonl.gz' |
| if not benchmark_file.exists(): benchmark_file.write_bytes(urllib.request.urlopen(url).read()) |
| benchmark=[json.loads(l) for l in gzip.decompress(benchmark_file.read_bytes()).decode().splitlines()] |
| |
| eval_texts={' '.join(p['prompt'].split()).lower() for p in benchmark} |
| eval_bodies=set() |
| for p in benchmark: |
| tree=ast.parse(p['prompt']+p['canonical_solution']); fn=next(n for n in tree.body if isinstance(n,ast.FunctionDef)) |
| if ast.get_docstring(fn) is not None: fn.body=fn.body[1:] |
| eval_bodies.add(digest(ast.dump(ast.Module(body=fn.body,type_ignores=[]),include_attributes=False))) |
| rows=pq.read_table(ROOT/'source.parquet').slice(0,500).to_pylist() |
| rejected=[]; accepted=[]; seen=set() |
| for index,row in enumerate(rows): |
| try: |
| prompt,completion,body_hash=extract(row) |
| if body_hash in eval_bodies or ' '.join(row['instruction'].split()).lower() in eval_texts: raise ValueError('exact_benchmark_overlap') |
| if body_hash in seen: raise ValueError('duplicate_body') |
| if len(tok(prompt+completion)['input_ids'])+1>2048: raise ValueError('context_too_long') |
| seen.add(body_hash) |
| split='dev' if int(digest(row['sha1'])[:8],16)%10==0 else 'train' |
| accepted.append(dict(source_id=row['id'],source_row=index,source_revision=REV,source_seed_sha1=row['sha1'],source_url=f'https://huggingface.co/datasets/bigcode/self-oss-instruct-sc2-exec-filter-50k/tree/{REV}',instruction=row['instruction'],original_response=row['response'],prompt=prompt,completion=completion,token_count=len(tok(prompt+completion)['input_ids'])+1,body_sha256=body_hash,split=split,validation_status='static_candidate_not_execution_verified',loss_mask='prompt=-100; completion+EOS=supervised')) |
| except ValueError as exc: rejected.append(dict(source_id=row['id'],source_row=index,reason=str(exc))) |
| |
| selected=accepted[:100] |
| write('candidates.jsonl',selected); write('rejected.jsonl',rejected) |
| write('candidate-train.jsonl',[r for r in selected if r['split']=='train']) |
| write('candidate-dev.jsonl',[r for r in selected if r['split']=='dev']) |
| ids={165808,533341,81575} |
| write('reviewed-preview.jsonl',[dict(r,validation_status='human_reviewed_tests_pending') for r in accepted if r['source_id'] in ids]) |
| report=dict(dataset='bigcode/self-oss-instruct-sc2-exec-filter-50k',revision=REV,license='odc-by',source_file_sha256=hashlib.sha256((ROOT/'source.parquet').read_bytes()).hexdigest(),source_rows_total=pq.read_metadata(ROOT/'source.parquet').num_rows,inspected_rows=500,static_accepted=len(accepted),selected_candidates=len(selected),candidate_splits=dict(Counter(r['split'] for r in selected)),rejection_reasons=dict(Counter(r['reason'] for r in rejected)),tokenizer_revision=MODEL_REV,benchmark_sha256=hashlib.sha256(benchmark_file.read_bytes()).hexdigest(),notes=['Static candidates are not approved for training.','Exact text/AST screening does not establish absence of paraphrased benchmark contamination.','Split is grouped by source seed SHA1; near-duplicate families need further review.','No code was executed on the host.','Prompt uses the exact source instruction as a Python docstring; body is AST-normalized.']) |
| (ROOT/'manifest.json').write_text(json.dumps(report,indent=2)); print(json.dumps(report,indent=2)) |
| if __name__=='__main__': main() |
|
|
