| """Offline native-token and prompt checks before freezing the pilot.""" |
| import json |
| from encoding_tax.common import ROOT,load_config,write_json,digest,now,READERS |
| from encoding_tax.data import load_rows |
| from encoding_tax.tokenization import tokenizer |
| from encoding_tax.prompts import serialize,continue_answer |
| from .design import HERE,cases,truth,body,case_id |
|
|
| def main(): |
| cfg=load_config(str(HERE/'config.json'));base=load_config() |
| rows=load_rows({'datasets':{'synthetic':cfg['dataset']}},'synthetic','pilot') |
| results={} |
| for reader in READERS: |
| p=base['models'][reader];tok=tokenizer(p['tokenizer_repo'],p['tokenizer_revision']) |
| maxanswer=0;maxprompt=0 |
| for m in cases(rows,reader): |
| text,_=body(m,base);raw=serialize(p,text,'scratchpad');count=tok.count(raw) |
| maxprompt=max(maxprompt,count) |
| assert count+m['budget']+48+64<p['context_length'],case_id(m) |
| if m['mode']=='diagnostic': |
| n=tok.count(json.dumps(truth(m),separators=(',',':')));maxanswer=max(maxanswer,n) |
| assert n<=48,case_id(m) |
| results[reader]={'samples':1824,'maximum_initial_prompt_tokens':maxprompt,'maximum_canonical_diagnostic_tokens':maxanswer,'answer_cap':48,'context_verified':True} |
| print(reader,results[reader],flush=True) |
| write_json(HERE/'validation.json',{'recorded_at':now(),'offline_only':True,'config_sha256':digest(HERE/'config.json'),'readers':results}) |
| if __name__=='__main__':main() |
|
|
