wflying/rl-if-math-logic-61k

RL IF Math Logic 61K

This dataset is a mixed training set for reinforcement learning with verifiable rewards (RLVR), covering instruction following, mathematics, logical reasoning, and table reasoning. It contains 60,980 examples in one UTF-8 JSON Lines file, provided as a single train split. No validation or test split is included.

The dataset follows the prompt / label / metadata format documented in wflying/rl-if-math-coding-logic-60k. The composition below is measured from this release: it includes 1,000 multiple-choice puzzles from sunyiyou/math_logic_puzzles_zebralogic_level_5 and has no klear_code component.

Source composition

Category	metadata.data_type	Examples	Source / attribution
Instruction following	if	20,994	allenai/IF_multi_constraints_upto5, the instruction-following source family documented in the reference mixture
Mathematics	klear_math	20,000	Klear mathematics and additional AIME/HMMT problems; see the breakdown below
Procedural reasoning	nemotron_rl_reasoning_gym	9,486	wflying/logic-rl-24k, with upstream attribution to nvidia/Nemotron-RL-ReasoningGym-v1
Logic and table reasoning	guru_rl	9,500	wflying/logic-rl-24k, with upstream attribution to LLM360/guru-RL-92k
ZebraLogic Level 5 multiple-choice puzzles	math_logic_puzzle	1,000	sunyiyou/math_logic_puzzles_zebralogic_level_5
Total		60,980

The task counts are computed from metadata.data_type. This field routes examples to reward functions; it is not always a unique upstream dataset identifier. In particular, all mathematics examples use klear_math, including the added competition problems.

Mathematics breakdown

Component	Examples	Identification / source
Klear mathematics family	17,761	klear_math records without an additional metadata.source; source family: Kwai-Klear/KlearReasoner-MathSub-30K
AIME, 1983–2023	918	metadata.source == "aime"; question text and problem IDs match the corresponding records in di-zhang-fdu/AIME_1983_2024
HMMT problems from Omni-MATH	1,231	metadata.source == "hmmt" without sub_source; problem text matches KbsdJames/Omni-MATH
HMMT February 2023	30	metadata.sub_source == "hmmt_feb_2023"; MathArena/hmmt_feb_2023
HMMT February 2024	30	metadata.sub_source == "hmmt_feb_2024"; MathArena/hmmt_feb_2024
HMMT November 2025	30	metadata.sub_source == "hmmt_nov_2025"; MathArena/hmmt_nov_2025
Total	20,000

Records without a source identifier do not preserve a row-level provenance manifest. Source-family attribution for those records follows the reference mixture; it should not be interpreted as an independently reconstructed sampling history.

ZebraLogic Level 5 component

The 1,000 math_logic_puzzle records contain house/attribute constraint puzzles with numbered answer choices. Their prompts and reference answers match the linked ZebraLogic Level 5 source. The source's messages and ground_truth fields are represented here as prompt, label, and metadata.ground_truth.

• label is the zero-based choice index stored as a string, not the text of the chosen option. Observed values are "0" through "5".
• metadata.ground_truth equals label for every record in this component.
• The prompt asks for the selected index in LaTeX form, for example \boxed{2}.
• A reward implementation should extract the boxed index and compare it with the reference, while respecting the available choices in that prompt.

These records use a different reward route from Guru's logic__zebra_puzzle_dataset, whose labels are structured objects.

File and integrity information

Property	Value
File	rl-if-math-logic-61k.jsonl
Format	UTF-8 JSON Lines, one JSON object per line
Split	train
Records	60,980
File size	190,959,561 bytes
SHA-256	3d291012b7db6ea31b106893d9dc8fc0459c417657d1b8626f32eb68d3785943

All 60,980 records parse as JSON, contain the top-level fields prompt, label, and metadata, and have exactly one user message. Every record includes metadata.num_tokens. The checksum above matches the published data file.

Data structure

Each record has the following structure; the example is schematic:

{
  "prompt": [
    {"role": "user", "content": "<problem and answer-format instructions>"}
  ],
  "label": "2",
  "metadata": {
    "data_type": "math_logic_puzzle",
    "ground_truth": "2",
    "num_tokens": 400
  }
}

• prompt: A chat-formatted problem with one user message.
• label: A reference answer, or an empty string for instruction-following records whose rewards are determined by constraints. The release contains 58,017 string labels, 1,749 list labels, and 1,214 object labels.
• metadata: Reward-routing information, reference answers, constraints, token counts, source identifiers where available, and task-specific auxiliary fields. Its nested schema varies across records.

Common metadata fields include:

Field	Meaning
data_type	Top-level task/reward family
ground_truth	Reference answer; equals label for all 39,986 non-IF records
instruction_id_list, instruction_kwargs	Constraint checker identifiers and parameters for IF records
num_tokens	Precomputed token count; the tokenizer is not identified in the release
source, sub_source	Additional mathematics provenance where retained
source_dataset	Procedural task generator for Nemotron records
data_source, ability, style, extra_info	Guru task family, reward style, and auxiliary information

All 20,994 IF records have an empty label. Some mathematics metadata retains source solutions; training code should send only prompt to the model and keep labels and reference metadata available to the reward function.

Loading

Because label mixes strings, lists, and objects, and metadata contains heterogeneous nested structures, automatic Arrow schema inference may fail. Reading the original JSONL preserves the field values:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="wflying/rl-if-math-logic-61k",
    repo_type="dataset",
    filename="rl-if-math-logic-61k.jsonl",
)

with open(path, encoding="utf-8") as f:
    for line in f:
        example = json.loads(line)
        prompt = example["prompt"]
        label = example["label"]
        metadata = example["metadata"]
        # Route to the appropriate reward using metadata["data_type"].

To construct a Hugging Face datasets.Dataset with a consistent schema, JSON-encode the heterogeneous fields first. This example continues from the path above:

from datasets import Dataset

def normalized_examples():
    with open(path, encoding="utf-8") as f:
        for line in f:
            example = json.loads(line)
            yield {
                "prompt": example["prompt"],
                "label_json": json.dumps(example["label"], ensure_ascii=False),
                "metadata_json": json.dumps(example["metadata"], ensure_ascii=False),
            }

dataset = Dataset.from_generator(normalized_examples)

# Recover the original JSON types before reward computation.
label = json.loads(dataset[0]["label_json"])
metadata = json.loads(dataset[0]["metadata_json"])

Intended use

The release supports mixed-domain RLVR training, verifiable instruction following, mathematical and logical reasoning, table question answering, and reward-routing experiments. Route examples by metadata.data_type, and use metadata.data_source to distinguish Guru task families where needed.

Reward functions must follow each prompt's expected output format, including boxed mathematics answers, boxed choice indices, <answer>...</answer> tags, and verifiable instruction constraints.

Limitations

• The release contains a training split only. It is not an independent benchmark.
• Competition problems and other source tasks can overlap with evaluation sets. In particular, the mixture includes AIME and HMMT questions; check exact and semantic overlap before reporting results on related benchmarks.
• Task families, answer formats, and difficulty levels are not uniformly distributed.
• Some records omit exact source identifiers; the release does not document a complete sampling seed or construction history.
• Syntax and structural checks do not establish the correctness or unambiguity of every prompt and reference answer.
• The dataset inherits errors, biases, and usage conditions from its sources. IF data may include output generated by third-party models.

Licensing and attribution

The repository uses the Hub's other license tag because the sources have different terms and some terms are unspecified. This tag does not assign a new license to the underlying records.

Source component	Published source terms / notes
AllenAI IF data	ODC-BY-1.0; the source card also refers to Ai2's Responsible Use Guidelines and third-party model terms
Klear mathematics	Apache-2.0
AIME source release linked above	MIT metadata; its card describes the data as a benchmark and asks that it not be used for training
Omni-MATH	Apache-2.0
MathArena HMMT releases	CC BY-NC-SA 4.0
Nemotron RL Reasoning Gym	CC BY 4.0; retained in metadata.license for all 9,486 records
Guru RL	MIT, as documented by the upstream collection
sunyiyou/math_logic_puzzles_zebralogic_level_5	No README, license file, or license metadata was present in the source repository when this card was prepared

Consult the linked source repositories for applicable terms and attribution requirements. The unspecified ZebraLogic source terms should not be interpreted as a permissive license, and the MathArena component retains its noncommercial and share-alike terms.

Source citations

Please attribute this compilation and the source datasets used in your work. The linked source cards provide their current citation entries, including:

• Generalizing Verifiable Instruction Following — AllenAI source and citation.
• Klear-Reasoner — Klear mathematics source and citations.
• Omni-MATH — dataset and citation.
• AIME and MathArena — the competition-data releases linked in the mathematics breakdown.
• Guru RL and Nemotron RL Reasoning Gym — logic mixture attribution and its linked upstream releases.
• ZebraLogic Level 5 — sunyiyou/math_logic_puzzles_zebralogic_level_5.