"""One deterministic generation per task, followed by isolated execution.""" import argparse import hashlib import json from pathlib import Path import sys from validators.dataset import load from runners import get_runner,LANGUAGES from runners.common import extract_code from runners.sandbox import Sandbox from metrics import aggregate ROOT=Path(__file__).resolve().parent PROMPT_VERSION='smolscript-completion-v1' def prompt(row): # Only these public fields reach the model. No tests, answers or references. fixture=row['metadata'].get('fixture','');base=row['metadata'].get('base_css','') return ('Write code for the following task. Use the standard library only. Return code only.\n' +row['prompt']+'\n'+row['output_format']+'\n'+(row['metadata'].get('input_domain') or '') +('\nHTML fixture:\n'+fixture if fixture else '')+('\nExisting CSS:\n'+base if base else '')+'\nCode:\n') def save(path,value): path.parent.mkdir(parents=True,exist_ok=True);temp=path.with_suffix('.tmp');temp.write_text(json.dumps(value,indent=2,allow_nan=False)+'\n');temp.replace(path) def main(): p=argparse.ArgumentParser(description=__doc__);p.add_argument('--model');p.add_argument('--revision',default='main');p.add_argument('--output',type=Path,required=True) p.add_argument('--language',nargs='+',choices=LANGUAGES);p.add_argument('--difficulty',nargs='+',choices=('easy','medium','moderate')) p.add_argument('--resume',action='store_true');p.add_argument('--predictions',type=Path,help='JSONL with id and code; bypass model inference') p.add_argument('--image',default='smolscript-runtime:0.1.0');p.add_argument('--engine',default='docker');p.add_argument('--max-new-tokens',type=int,default=384);p.add_argument('--device',default='auto');a=p.parse_args() if not a.model and not a.predictions:p.error('--model or --predictions required') if a.max_new_tokens<1:p.error('--max-new-tokens must be positive') sandbox=Sandbox(a.image,a.engine) rows=[r for r in load(ROOT/'data/smolscript.jsonl') if (not a.language or r['language'] in a.language) and (not a.difficulty or r['difficulty'] in a.difficulty)] hidden={r['id']:r for r in load(ROOT/'private_tests/hidden.jsonl')} files=['data/smolscript.jsonl','private_tests/hidden.jsonl','evaluate.py','metrics.py']+[str(p.relative_to(ROOT)) for p in sorted((ROOT/'runners').glob('*.py'))]+['sandbox/worker.py'] fingerprint={f:hashlib.sha256((ROOT/f).read_bytes()).hexdigest() for f in files} settings={'model':a.model,'requested_revision':a.revision,'language':a.language,'difficulty':a.difficulty, 'max_new_tokens':a.max_new_tokens,'prompt_version':PROMPT_VERSION,'dataset':fingerprint,'sandbox':sandbox.identity, 'predictions_sha':hashlib.sha256(a.predictions.read_bytes()).hexdigest() if a.predictions else None} if a.output.exists() and not a.resume:p.error('Output exists; use --resume or a new path') state=json.loads(a.output.read_text()) if a.output.exists() else {'settings':settings,'results':[],'complete':False} if state['settings']!=settings:p.error('Resume settings mismatch') if a.resume and not a.output.exists():p.error('Resume output missing') # Lock the output across processes to prevent concurrent checkpoint overwrites. import fcntl lock=a.output.with_suffix(a.output.suffix+'.lock');lock.parent.mkdir(parents=True,exist_ok=True) with lock.open('a') as stream: fcntl.flock(stream,fcntl.LOCK_EX|fcntl.LOCK_NB) codes={} if a.predictions:codes={r['id']:r['code'] for r in load(a.predictions)} else: import torch from transformers import AutoTokenizer,AutoModelForCausalLM,AutoConfig from huggingface_hub import snapshot_download snapshot=snapshot_download(a.model,revision=state.get('resolved_revision') or a.revision) revision=Path(snapshot).name state['resolved_revision']=revision tokenizer=AutoTokenizer.from_pretrained(snapshot) model=AutoModelForCausalLM.from_pretrained(snapshot,torch_dtype='auto',device_map=a.device) model.eval();torch.manual_seed(0) import importlib.metadata state['environment']={'python':sys.version,'packages':{n:importlib.metadata.version(n) for n in ('torch','transformers','huggingface-hub')},'device':str(model.device)} done={r['id'] for r in state['results']} if len(done)!=len(state['results']) or done-set(r['id'] for r in rows):raise ValueError('Corrupt resume IDs') save(a.output,state) for row in rows: if row['id'] in done:continue if state.get('pending',{}).get('id')==row['id']: text=state['pending']['generation'] elif a.predictions: if row['id'] not in codes:raise ValueError('Missing prediction '+row['id']) text=codes[row['id']] else: encoded=tokenizer(prompt(row),return_tensors='pt',return_token_type_ids=False).to(model.device) native=getattr(model.config,'max_position_embeddings',getattr(model.config,'n_positions',4096)) if encoded['input_ids'].shape[1]+a.max_new_tokens>native: raise ValueError('Insufficient context for fixed generation settings: '+row['id']) with torch.inference_mode():out=model.generate(**encoded,max_new_tokens=a.max_new_tokens,do_sample=False,num_beams=1,pad_token_id=tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id) text=tokenizer.decode(out[0,encoded['input_ids'].shape[1]:],skip_special_tokens=True) state['pending']={'id':row['id'],'generation':text};save(a.output,state) try:code=extract_code(text) except ValueError as e:result={'passed':False,'syntax_success':False,'tests':[],'error':str(e)} else:result=get_runner(row['language']).evaluate(row,code,hidden[row['id']],sandbox) state['results'].append({'id':row['id'],'language':row['language'],'difficulty':row['difficulty'],'generation':text,**result}) state.pop('pending',None) save(a.output,state);print(f"{len(state['results'])}/{len(rows)} {row['id']} {'PASS' if result['passed'] else 'FAIL'}",flush=True) state['metrics']=aggregate(state['results'],[r['id'] for r in rows],full_run=not(a.language or a.difficulty));state['complete']=True;save(a.output,state) print(json.dumps(state['metrics'],indent=2)) if __name__=='__main__':main()