WEBMemo

WEBMemo is a synthetic text-generation dataset designed for reasoning-heavy supervised fine-tuning. It combines instruction-following prompts, explicit reasoning traces, and math-oriented tasks in a Hugging Face-ready parquet layout.

The repository is structured to support two common workflows:

• loading sharded training data from data/
• loading a separate standalone corpus from full/

Overview

WEBMemo targets four core capabilities:

• reasoning
• instruction-following
• chain-of-thought style supervision
• math problem solving

Unlike a minimal prompt-answer corpus, WEBMemo stores a dedicated reasoning field alongside the final response, which makes it useful for experiments that separate intermediate reasoning from answer generation.

Intended Use

WEBMemo is intended for:

• supervised fine-tuning
• reasoning-oriented instruction tuning
• math and applied problem-solving training
• prompt-format and response-style experiments
• evaluation of data mixture strategies across separate parquet shards

It is best suited for training or analysis pipelines that need:

• explicit reasoning text
• unique rows across the full repository
• mixed symbolic and verbal tasks
• a stable source label for filtering and attribution

Repository Layout

WEBMemo/
  .gitattributes
  LICENSE
  README.md
  data/
    train-000.parquet
    train-001.parquet
  full/
    train.parquet
  tools/
    generate_webmemo.py

Configs

WEBMemo exposes two Hugging Face configs:

Config	File(s)	Role	Focus
data	data/train-000.parquet, data/train-001.parquet	Sharded training set	Instruction reasoning and math reasoning
full	full/train.parquet	Standalone training set	Mixed reasoning, policy logic, evidence comparison, and applied math

Size And Uniqueness

The repository contains 21,500 total rows:

• data/train-000.parquet: 7,000 rows
• data/train-001.parquet: 7,000 rows
• full/train.parquet: 7,500 rows

Uniqueness guarantees:

• no exact duplicate prompts within any file
• no exact duplicate prompt-response pairs within any file
• no exact duplicate prompts across the full repository
• no exact duplicate prompt-response pairs across the full repository

Data Composition

data/train-000.parquet

This shard is centered on instruction-heavy reasoning tasks, including:

• prioritization under constraints
• evidence comparison
• policy interpretation
• workflow diagnosis
• planning tradeoffs
• decision selection with competing goals

data/train-001.parquet

This shard is centered on math reasoning tasks, including:

• arithmetic word problems
• fractions and percentages
• algebra and equations
• geometry
• rate and distance
• conversions, averages, and patterns

full/train.parquet

This is a separate standalone corpus rather than a duplicate merge of the shard files. It combines:

• mixed reasoning
• policy reasoning
• argument evaluation
• strategic tradeoff analysis
• evidence-based judgments
• applied math problems

Schema

All parquet files share the same columns:

Field	Type	Description
id	string	Unique sample id
subset	string	Source shard name
category	string	Task family
prompt	string	User instruction or problem
reasoning	string	Structured intermediate reasoning
response	string	Final answer
difficulty	string	Relative difficulty label
source	string	Data origin label
language	string	Language code

Example Record

{
  "id": "wmr-00001",
  "subset": "data-shard-a",
  "category": "instruction_reasoning",
  "prompt": "Reason through the scenario step by step before answering. ...",
  "reasoning": "The decision should track the core requirement rather than raw headcount. ...",
  "response": "The analysts should lead the next phase.",
  "difficulty": "medium",
  "source": "Goldgolf-exchange",
  "language": "en"
}

Loading

from datasets import load_dataset

sharded = load_dataset("Surpem/WEBMemo", "data", split="train")
full = load_dataset("Surpem/WEBMemo", "full", split="train")

You can also load an individual parquet file directly:

import pandas as pd

df = pd.read_parquet("data/train-000.parquet")

Source Label

Every row uses the source value Goldgolf-exchange.

Generation Design

The generator for WEBMemo follows a few strict constraints:

• exact prompt duplication is blocked across the whole repository
• exact prompt-response duplication is blocked across the whole repository
• the full config is a standalone corpus, not a merged copy of the data config
• reasoning and math tasks are intentionally separated in the sharded config
• all data remains English-language and text-generation oriented

Limitations

WEBMemo is synthetic. That makes it useful for controllable fine-tuning, but it also means:

• realism depends on fit to your downstream use case
• chain-of-thought style supervision may not match every deployment policy
• prompt diversity does not automatically imply benchmark-level difficulty

Notes

• The two files in data/ are intentionally different in prompt mix.
• full/train.parquet is not a duplicate merge of the shard files.
• Samples are generated to avoid exact prompt duplication across all files.
• The dataset is intended as a structured training resource, not a ground-truth benchmark.