rasdani/smol-RLVR-IFeval

dataset_info:
features:

• name: messages
  list:
  • name: content
    dtype: string
  • name: role
    dtype: string
• name: ground_truth
  dtype: string
• name: dataset
  dtype: string
• name: constraint_type
  dtype: string
• name: constraint
  dtype: string
  splits:
• name: train
  num_bytes: 26448373
  num_examples: 14973
  download_size: 11524278
  dataset_size: 26448373
  configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*

import copy
import json
import random
import re
import datasets


ds = datasets.load_dataset("allenai/RLVR-IFeval", split="train")

def replace_numeric_constraint(example, transform_func=lambda x: x):
    ret = copy.deepcopy(example)
    ground_truth_dict = json.loads(ret["ground_truth"])
    old_N = ground_truth_dict["N"]
    new_N = transform_func(old_N)
    ground_truth_dict["N"] = new_N

    # First replace in example["constraint"]
    if ret["constraint_type"] == "Number Paragraphs + First Word in i-th Paragraph":
        old_i = ground_truth_dict["i"]
        new_i = random.randint(1, new_N)
        ground_truth_dict["i"] = new_i
        first_word = ground_truth_dict["first_word"]
        old_constraint_str = ret["constraint"].replace("{N}", str(old_N)).replace("{i}", str(old_i)).replace("{first word}", first_word)
        new_constraint_str = ret["constraint"].replace("{N}", str(new_N)).replace("{i}", str(new_i)).replace("{first word}", first_word)
    else:
        old_constraint_str = ret["constraint"].replace("{N}", str(old_N))
        new_constraint_str = ret["constraint"].replace("{N}", str(new_N))

    ret["ground_truth"] = json.dumps(ground_truth_dict)
    
    # Then replace in messages content
    if ret["constraint_type"] == "Number Sentences":
        # special constraint_type
        # "Answer with at least / around / at most {N} sentences."

        pattern = r"(Answer with (at least|around|at most)) (\d+)( sentences\.)"

        def replace_number_only(match):
            prefix = match.group(1)  # "Answer with around"
            suffix = match.group(4)  # " sentences."
            return f"{prefix} {new_N}{suffix}"

        ret["messages"][0]["content"] = re.sub(pattern, replace_number_only, ret["messages"][0]["content"])
    else: 
        ret["messages"][0]["content"] = ret["messages"][0]["content"].replace(old_constraint_str, new_constraint_str)
    return ret


numerical_constraints = [
    "Number Paragraphs",
    "Keyword Frequency",
    "Minimum Number Highlighted Section",
    "Number Paragraphs + First Word in i-th Paragraph",
    "Letter Frequency",
    "Number Sentences",
    "Multiple Sections",
    "Number Placeholder",
    "Number Bullets"
]
transform_func = lambda x: random.randint(1, 5)

ds_reduced = ds.map(
    lambda x: replace_numeric_constraint(x, transform_func=transform_func) if x["constraint_type"] in numerical_constraints else x, 
    num_proc=8
)

ds_reduced.push_to_hub("rasdani/smol-RLVR-IFeval")