Hugging Face's logo

Datasets:
miceworks
/
Mark-Ultra-Curriculum

Tasks:
Text Generation
Question Answering
Modalities:
Text
Formats:
parquet
Languages:
English
Portuguese
Size:
1K - 10K
Tags:
mark-ai
code
systems-engineering
kernel
rust
cpp
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
Mark-Ultra-Curriculum / README.md
miceworks's picture
miceworks
docs: publish Mark Ultra-Curriculum dataset card
81ba514 verified
preview code
|
Raw History Blame Contribute Delete
3.58 kB
metadata 
language:
  - en
  - pt
license: agpl-3.0
task_categories:
  - text-generation
  - question-answering
tags:
  - mark-ai
  - code
  - systems-engineering
  - kernel
  - rust
  - cpp
  - ebpf
  - robotics
  - mechatronics
  - synthetic-dpo
  - reasoning
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.parquet

Mark Ultra-Curriculum: Systems, Kernel, Robotics & Epistemic Reasoning Dataset

The Mark Ultra-Curriculum is an open-source training dataset containing 4,821 high-density instruction-completion pairs designed for systems programming, low-level concurrency, robotics, STEM reasoning, and epistemic calibration.

  • License: GNU Affero General Public License v3.0 (AGPL-3.0)
  • Primary Formats: Apache Parquet (data/train.parquet) and JSON Lines (data/train.jsonl)
  • Origin: The foundational training corpus of the Mark AI model ecosystem (Mark Engineer 1.5B/8B, Mark General 1.5B, MarkVL 2.2B).

Dataset Taxonomy & Domain Breakdown

Domain Identifier Sample Count
general_systems_engineering 3736
systems_rust_concurrency 722
academic_stem_reasoning 137
senior_tech_lead_dialogue 70
systems_cpp20_atomics 45
multimodal_dense_vision 42
cybersecurity_and_reversing 28
systems_linux_ebpf 21
robotics_and_mechatronics 14
epistemic_self_rag 5
systems_windows_kmdf 1
Total Unique Samples 4,821

Key Domain Descriptions:

  1. systems_linux_ebpf: BPF/XDP packet filters, kprobes, verifier bounded-loop invariants, zero-copy socket handling.
  2. systems_windows_kmdf: Kernel-Mode Driver Framework, IRQL separation (PASSIVE vs DISPATCH), IRP queuing, pool allocation safety.
  3. systems_rust_concurrency: Asynchronous Tokio runtimes, pin projection, cross-await deadlock avoidance, Send/Sync formal properties.
  4. systems_cpp20_atomics: C++20/C++23 coroutines, atomic memory orderings (acquire/release/seq_cst), lock-free ringbuffers.
  5. robotics_and_mechatronics: PID motor control, forward/inverse kinematics, CAN bus communication, PCB differential routing in KiCad.
  6. academic_stem_reasoning: Quantum mechanics (Hamiltonian systems), fluid dynamics (Navier-Stokes), chemical kinetics, clinical pharmacokinetics (ADME).
  7. epistemic_self_rag: Built-in [UNKNOWN] and [WEB_SEARCH] calibration pairs for identifying false premises and unanswerable queries.
  8. cybersecurity_and_reversing: Memory corruption mitigation (ASLR, DEP, CFI), ROP chain mechanics, ELF/PE parsing, disassembly analysis.
  9. senior_tech_lead_dialogue: Pragmatic engineering trade-offs, legacy system refactoring, empathetic architectural leadership.

Data Structure

Each record contains the following fields:

{
  "id": "mark_sample_000123",
  "domain": "systems_linux_ebpf",
  "source_curriculum": "v2_8b_extended_master",
  "prompt": "<think>\n...\n</think>\nUser query...",
  "completion": "<think>\n1. Technical step breakdown...\n</think>\nImplementation..."
}

How to Load in Python 

from datasets import load_dataset

# Load directly from Hugging Face Hub
dataset = load_dataset("miceworks/Mark-Ultra-Curriculum", split="train")

print(dataset[0])

License & Distribution

Released under the GNU Affero General Public License v3.0 (AGPL-3.0) to guarantee open accessibility, reproducibility, and copyleft sharing across the open-source machine learning research community.