Reasoning Core ◉

Procedural reasoning data for language-model pre-training, post-training, evaluation, and RL.

Reasoning Core generates verifiable textual tasks across first-order logic, formal mathematics with Lean and TPTP, planning, algorithms, syntax, and more. Use it as a Python library, generate datasets at scale, or plug it into modern reinforcement-learning environments.

More than 10B tokens of pre-generated data are available in the 🤗 Reasoning Core dataset collection.

Tasks target compact, canonical answers and expose task-native scorers—clean targets for supervised fine-tuning (SFT), with deterministic rewards for RL and evaluation. The public training and influence API provides reproducible paired baseline/treatment experiments. Follow the task influence guide to measure a new or changed task, and compare it with the reference results.

Quickstart

uv pip install "reasoning-core[gen]"   # generate and score
uv pip install reasoning-core          # score only, e.g. as an RL reward over the procedural pile

from reasoning_core import get_task, score_answer

task = get_task("arithmetics")
example = task.generate_example()

print(example.prompt)
assert score_answer(example.answer, example) == 1

Everyday workflows

Python 3.10+. From a checkout, use python -m pip install -e '.[gen]' or bash scripts/install_authoring.sh for lightweight task authoring.

Repository layout explains package boundaries and compatibility. Workflow recipes cover prerequisites, commands, output files, and local checks. Catalogue queries read source metadata without importing tasks.

Representative example

Reasoning Core includes compilation-checked formal reasoning tasks such as lean_candidate_compilation:

Prompt

Does this Lean 4 tactic body close the theorem?
The answer is True or False.

THEOREM:
theorem ex (p2 p4 : Prop) : p2 → (p2 ∨ p4) := by
  ?

CANDIDATE:
linarith

Answer: False

Browse all 65 task examples.

Task catalogue

GALLERY (names link to gallery examples)

arithmetics · math_word_problem · equation_system · combinatorics_formula · function_manipulation · lean_missing_line · lean_candidate_compilation · planar_geometry_relations · metamath_entailment · metamath_core_select · lambda_reduction · rewrite_system · unification_entailment · most_probable_evidence · most_probable_outcome · multistep_nli · defeasible_nli · multistep_evidence_retrieval · multistep_abduction · logic_qa · logic_derivation · planning · set_missing_element · set_expression · sequential_induction · qualitative_reasoning · grid_navigation · reference_tracking · belief_tracking · coreference · constraint_satisfaction · graph_pathfinding · graph_successors · regex_following · regex_reasoning · analogical_case_matching · parsing_derivation · syntax_error_detection · constrained_continuation · table_qa · table_equivalence · table_statistics · string_transduction · game_best_move · game_forced_win · qualitative_causal_reasoning · code_analysis · code_runnability · code_execution · program_synthesis · backtracking_search · boolean_propagation_search · controlled_code_execution · dynamic_programming · fixpoint_iteration · matrix_induction · pattern_induction · schema_bound_query · conditional_response_contract · protected_span_transformation · rule_switching · shift_reduce_parsing · spatial_folding · typed_relation_extraction · variable_elimination

Task authoring guidelines

A task authoring guide describes the interface and guidelines.
TASK_AUTHORING_GUIDE TASK_MUTATION_GUIDE TRAINING_AND_INFLUENCE TASK_INFLUENCE

Ecosystem and integrations

• Prime Intellect — install Reasoning Core from the Environments Hub for evaluation and RL workflows.
• OpenReward — run Reasoning Core as an OpenReward-compatible environment.
• OpenEnv — explore the interactive Reasoning Core OpenEnv on Hugging Face Spaces.
• reasoning-gym — mix Reasoning Core and reasoning-gym tasks through either library's interface.
• SynLogic — generate SynLogic games through the same task API, backed by SynLogic's native verifiers.

See the integration guide for runnable examples.

Generate datasets at scale

Go from a single example to large pre-training, post-training, and evaluation corpora with balanced difficulty, token budgets, and verifiable answers. The generation pipeline supports parallel workers, resumable jobs, JSONL shards, and postprocessing for Hugging Face Datasets. Start with the sampling and collection recipe.

Citation and paper

@article{sileo2026reasoning,
  title={Reasoning Core: Designing Broad Procedural Data for Completion-Supervised Reasoning Training},
  author={Sileo, Damien and Lacombe, Valentin and Kachler, Dimitri},
  journal={arXiv preprint arXiv:2608.05148},
  year={2026},
  url={https://arxiv.org/abs/2608.05148}
}

https://arxiv.org/abs/2608.05148
Contact: damien.sileo@inria.fr