OpenMementos – Derived Subsets

This repository is a duplicate of Microsoft’s microsoft/OpenMementos dataset.

It exists for one specific purpose:

to provide a working copy of OpenMementos from which additional, domain-specific and otherwise derived subsets can be generated and published.

The original dataset remains the source dataset. This repository adds a processing layer on top of it.

---

Origin

The underlying dataset is Microsoft’s microsoft/OpenMementos.

OpenMementos contains 228,557 reasoning traces with Memento-style block segmentation and compressed summaries. The original dataset provides the default training-ready configuration and the full configuration with the intermediate pipeline components.

For the complete description of the dataset, its construction, data sources, Memento format, statistics, training information, license, and citation, please refer to the original dataset:

microsoft/OpenMementos

This repository does not attempt to replace or redefine that original dataset description.

---

Why a duplicate?

The duplicate exists to make it possible to generate derived subsets without changing the original dataset.

The central idea is:

microsoft/OpenMementos
        │
        │ duplicate / working copy
        ▼
JensOtte83/OpenMementos
        │
        │ subset-generation pipelines
        ▼
     layers/
        │
        ├── domain/
        │   ├── code.parquet
        │   ├── math.parquet
        │   └── science.parquet
        │
        └── future layers

The original default and full configurations are retained.

The additional configurations such as code, math, and science are derived from the same underlying OpenMementos data.

---

Derived Subsets

Domain layer

The first derived layer is the domain layer.

It separates the original dataset into domain-specific subsets:

Subset	Examples	Share
math	123,333	~54%
science	61,485	~27%
code	43,739	~19%
Total	228,557	100%

The three subsets together cover the complete 228,557 examples of the original dataset.

The resulting files are published under:

layers/domain/

Currently:

layers/
└── domain/
    ├── code.parquet
    ├── math.parquet
    └── science.parquet

The files currently published in this layer are:

• code.parquet
• math.parquet
• science.parquet

---

Loading the derived subsets

The derived subsets are exposed as normal Hugging Face Dataset configurations.

For example:

from datasets import load_dataset

code = load_dataset(
    "JensOtte83/OpenMementos",
    "code",
    split="train"
)

math = load_dataset(
    "JensOtte83/OpenMementos",
    "math",
    split="train"
)

science = load_dataset(
    "JensOtte83/OpenMementos",
    "science",
    split="train"
)

The original training-ready configuration remains available as the default:

from datasets import load_dataset

ds = load_dataset(
    "JensOtte83/OpenMementos",
    split="train"
)

The original full configuration is also retained:

ds = load_dataset(
    "JensOtte83/OpenMementos",
    "full",
    split="train"
)

This means that existing code using the standard default configuration continues to work while the derived configurations can be accessed explicitly.

---

Subset Generation

The derived subsets are generated using Python data-processing pipelines.

The processing scripts follow the naming convention:

split_by_[...].py

For example:

split_by_domain.py

A typical pipeline follows this structure:

OpenMementos
     │
     ▼
split_by_[...].py
     │
     ├── selection / filtering
     │
     ├── validation
     │
     └── Parquet output
             │
             ▼
         layers/[layer]/

The purpose of this structure is to make the creation of additional derived layers reproducible and extensible.

Future processing pipelines may therefore create additional structures such as:

layers/
├── domain/
│   ├── code.parquet
│   ├── math.parquet
│   └── science.parquet
│
├── [future layer]/
│   └── ...
│
└── [future layer]/
    └── ...

The layers/ directory is intentionally treated as the derived-data processing layer of this repository.

---

Repository Structure

The relevant structure is:

OpenMementos/
│
├── data/
│   └── train-*
│
├── full/
│   └── train-*
│
├── layers/
│   └── domain/
│       ├── code.parquet
│       ├── math.parquet
│       └── science.parquet
│
└── split_by_[...].py

The data/ and full/ directories represent the original OpenMementos dataset configurations.

The layers/ directory contains data derived from those original data.

The split_by_[...].py scripts provide the processing logic used to create the derived layers.

---

Original Dataset vs. Derived Data

It is important to distinguish between the original dataset and the work performed in this repository.

┌──────────────────────────────────────┐
│ Microsoft OpenMementos                      │
│                                             │
│ 228,557 original examples                   │
│ default + full                              │
└──────────────────┬───────────────────┘
                      │
                      │ duplicate
                      ▼
┌──────────────────────────────────────┐
│ JensOtte83/OpenMementos                     │
│                                             │
│ preserved original configurations           │
└──────────────────┬───────────────────┘
                      │
                      │ Python pipelines
                      │ split_by_[...].py
                      ▼
┌──────────────────────────────────────┐
│ Derived layers                              │
│                                             │
│ domain/                                     │
│    ├── code                                 │
│    ├── math                                 │
│    └── science                              │
└──────────────────────────────────────┘

The distinction is intentional:

• OpenMementos is the original dataset.
• This repository is a duplicate/working copy.
• layers/ contains derived data.
• split_by_[...].py contains the processing logic used to create it.

---

Parquet Conversion Branch

There is also an interesting piece of infrastructure associated with this repository.

Hugging Face created the branch:

refs/convert/parquet

as part of its Parquet conversion workflow.

The branch contains the converted dataset structure:

refs/convert/parquet
├── code
├── default
├── full
├── math
└── science

This is particularly useful because the resulting structure allows the dataset to be consumed directly in the same general train-oriented Parquet form used by the original dataset.

The branch currently shows contributions from:

• parquet-converter
• JensOtte83
• vkontonis

and the relevant commit is:

Update parquet files

with commit:

5f7f626

See:

refs/convert/parquet

---

A note of appreciation

The Parquet conversion branch provides a convenient structure that makes the resulting data directly usable in the expected training-oriented format.

So, to the people maintaining the original OpenMementos dataset:

Thank you very much ✌️

Jens Otte

---

Data Format

The derived domain subsets retain the core OpenMementos schema:

Column	Type	Description
problem	string	Problem statement
response	string	Memento-formatted response
domain	string	Domain classification
source	string	Original dataset source
difficulty	int	Difficulty rating

The full configuration additionally contains the intermediate pipeline components:

Column	Type	Description
sentences	list[string]	Individual reasoning sentences
blocks	list[list[int]]	Block boundaries
block_summaries	list[string]	Compressed summaries for each block

For the authoritative description of these fields and the Memento response format, see the original dataset:

microsoft/OpenMementos

---

Provenance

The provenance chain of this repository is deliberately kept explicit:

OpenThoughts-v3
      │
      ▼
microsoft/OpenMementos
      │
      ▼
JensOtte83/OpenMementos
      │
      ├── preserved original configurations
      │
      └── derived subset generation
              │
              ▼
          layers/

The underlying data therefore originates from the original OpenMementos release and its documented upstream sources.

This repository adds processing and organization of derived subsets; it does not claim to be the original source of the underlying reasoning traces.

---

License

This repository uses the MIT License, consistent with the original OpenMementos dataset.

For the authoritative license and attribution information, please refer to the original dataset:

microsoft/OpenMementos

---

Citation

If you use the underlying OpenMementos dataset, please cite the original Memento work and dataset.

If you use one of the derived subsets generated in this repository, please also reference:

JensOtte83/OpenMementos

and, where relevant, the corresponding subset-generation pipeline.

The original citation is reproduced here for convenience:

@article{memento2026,
  author={Vasilis Kontonis and Yuchen Zeng and Shivam Garg and Lingjiao Chen and Hao Tang and Ziyan Wang and Ahmed Awadallah and Eric Horvitz and John Langford and Dimitris Papailiopoulos},
  title={Memento: Teaching LLMs to Manage Their Own Context},
  year={2026},
}

---

Summary

This repository is intentionally simple in concept:

DUPLICATE
   │
   ▼
PROCESS
   │
   ▼
SUBSET
   │
   ▼
PUBLISH

It preserves the original OpenMementos dataset while providing a reproducible place to generate and distribute derived subsets.

The first layer is domain-based:

code
math
science

Additional layers can be added without changing the original dataset.

---

Original dataset:
microsoft/OpenMementos

Derived dataset repository:
JensOtte83/OpenMementos

Domain layer:
layers/domain

Parquet conversion branch:
refs/convert/parquet