metadata
license: mit
task_categories:
- text-generation
tags:
- code
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: val
path: data/val-*
dataset_info:
features:
- name: prompt
dtype: string
- name: output
sequence: string
splits:
- name: train
num_bytes: 1252371
num_examples: 6015
- name: val
num_bytes: 65996
num_examples: 317
download_size: 50645
dataset_size: 1318367
This is a dataset for CFG-based GflowNet development. The ebnf style CFG for numbers are:
root ::= number_list
number_list ::= number (number_list)?
number ::= odd even | even odd
even ::= "0," | "2," | "4," | "6," | "8," | "10," | "12," | "14," | "16," | "18," | "20,"
odd ::= "1," | "3," | "5," | "7," | "9," | "11," | "13," | "15," | "17," | "19,"
However, this CFG cannot satisfy all the requirments. The CFG give a chunk of numbers where the transition between chunks always failed to satisfy our requirment:
Total correct ≤ Total grammar Agreed
~6332/100000 (how many correct numbers in CFG generated numbers)
So we need RL or Gflow to bridge the gap between CFG and the targets.