danielpedrozo's picture
upload from cluster
5d69d91 verified
|
Raw
History Blame Contribute Delete
482 Bytes
metadata
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-3B-Instruct
tags:
  - text-to-sql
  - self-play
  - grpo
  - sql-zero

sql-zero-3b-challenger-iter2

Challenger 3B iter2 c/ repetition penalty lambda=0.5. Dup 66%->57%; gen 36k pares.

Parte do projeto SQL-Zero (self-play zero-anotacao para Text-to-SQL, adaptacao do Dr. Zero). Protocolo de avaliacao: greedy@1, schema completo (DDL+3 rows), execution accuracy. Treinado com verl (GRPO); reward de execucao 3-tier (1.0/0.1/0.0).