README.md

---
pretty_name: Procedural Pile
license: cc-by-4.0
task_categories:
- text-generation
- question-answering
language:
- en
dataset_info:
  features:
  - name: task
    dtype: large_string
  - name: prompt
    dtype: large_string
  - name: answer
    dtype: large_string
  - name: metadata
    dtype: large_string
  - name: level
    dtype: int64
  - name: mode
    dtype: large_string
tags:
- reasoning
- procedural
- synthetic
- pretraining
- formal
size_categories:
- 10M<n<100M
---

<p align="center">
  <img src="https://imgshare.cc/api/image/proxy?id=443lb70v" alt="Procedural Pile: verifiable procedural data optimized for transferability" width="100%">
</p>

<p align="center">
  <a href="https://github.com/sileod/reasoning_core/blob/main/GALLERY.md">Task gallery</a> ·
  <a href="https://github.com/sileod/reasoning_core">Source</a> ·
  <a href="https://huggingface.co/papers/2509.18083">Paper</a> ·
  <a href="https://huggingface.co/datasets/reasoning-core/rc1">RLVR dataset</a>
</p>

Procedural Pile is a synthetic corpus of verifiable reasoning problems generated by [Reasoning Core](https://github.com/sileod/reasoning_core). It is intended for continued pretraining, mid-training, and supervised fine-tuning.

Answers come from procedural generators and task-specific solvers or checkers, rather than language-model generation. The corpus spans mathematics, formal logic, planning, graphs, parsing, code, structured data, and other symbolic domains. Difficulty is controlled continuously within each task.

Reasoning Core is optimized for transferability: task selection and difficulty ranges are guided by reproducible measurements of transfer, solvability, and shortcut resistance.

---

## Load

```python
from datasets import load_dataset

dataset = load_dataset("reasoning-core/procedural-pile")
```

No configuration name is required. The dataset provides `train` and `test` splits.

## Dataset structure

| Field | Description |
|---|---|
| `task` | Reasoning task identifier |
| `prompt` | Model input |
| `answer` | Canonical target answer |
| `metadata` | JSON-encoded generation and validation metadata |
| `level` | Difficulty level |
| `mode` | `instruct`, `few_shot`, or `verification` |

Most examples use direct instruction format. A smaller share adds one in-context demonstration or asks the model to verify a candidate answer.

---

## Task catalogue

The corpus currently contains 50 task families. The [task gallery](https://github.com/sileod/reasoning_core/blob/main/GALLERY.md) includes a worked example for each one.

| Area | Tasks |
|---|---|
| **Mathematics & formal methods** · 10 | `arithmetics` · `math_word_problem` · `equation_system` · `combinatorics_formula_selection` · `planar_geometry_relations` · `lean_candidate_compilation` · `lean_missing_line` · `metamath_core_select` · `metamath_entailment` · `sequential_induction` |
| **Logic & inference** · 10 | `logic_formalization` · `logic_nli` · `logic_qa` · `defeasible_nli` · `multistep_nli` · `multistep_abduction` · `multistep_evidence_retrieval` · `qualitative_reasoning` · `qualitative_causal_reasoning` · `belief_tracking` |
| **Symbolic transformations** · 7 | `lambda_reduction` · `rewrite_system` · `unification_entailment` · `set_expression` · `set_missing_element` · `string_transduction` · `analogical_case_matching` |
| **Planning, state & graphs** · 7 | `planning` · `constraint_satisfaction` · `grid_navigation` · `reference_tracking` · `coreference` · `graph_pathfinding` · `graph_successors` |
| **Language & formal languages** · 5 | `parsing_derivation` · `regex_following` · `regex_reasoning` · `constrained_continuation` · `syntax_error_detection` |
| **Structured data & code** · 7 | `table_qa` · `table_equivalence` · `table_statistics` · `code_analysis` · `code_execution` · `code_runnability` · `program_synthesis` |
| **Games & probability** · 4 | `game_best_move` · `game_forced_win` · `most_probable_evidence` · `most_probable_outcome` |

## Citation

```bibtex
@article{reasoningcore2026,
  title   = {Reasoning Core: A Scalable Procedural Data Generation Suite for Symbolic Pre-training and Post-Training},
  author  = {Lacombe, Valentin and Quesnel, Valentin and Sileo, Damien},
  journal = {arXiv preprint arXiv:2603.02208},
  year    = {2026},
  url     = {https://arxiv.org/abs/2603.02208}
}
```