logicjitter 0.1.0

Algorithmically generated logic games for LLM fine-tuning and misinformation detection.

LogicJitter

LogicJitter: Let LLMs play and uncover misinformation

Python package for automatically generating logic-based training data to improve LLM reasoning and misinformation detection.

Authors: Luca Herranz-Celotti · Marco Viviani
Published in: Computers and Electrical Engineering 136 (2026) 111215

Read the paper: LogicJitter on ScienceDirect
DOI: 10.1016/j.compeleceng.2026.111215

---

Overview

LogicJitter generates structured, rule-based logic games in natural language. Each sample includes valid and fallacious reasoning patterns plus optional distractors (random characters, controlled errors, problem revisions) designed to counter cognitive biases and logical fallacies.

The dataset is virtually unbounded, perfectly balanced (50/50 true/false), and requires no human or LLM labeling.

Logic games

Game	Description
Guided Maths	Scratchpad-style arithmetic (sum, product, polynomial)
Causal Clauses	Causal graphs with connectivity and fork/collider questions
Context-Free Grammars	String membership and grammar comparison
CLEVR	Textual spatial-reasoning scenes from CLEVR-style templates

Augmentation presets

Preset	Components
full	Games + Errors + Characters + Revisions (GECR)
g	Games only
ge	Games + Errors
gec	Games + Errors + Characters

---

Installation

pip install logicjitter

Optional: set a custom data directory for generated datasets and CLEVR downloads:

export LOGICJITTER_DATA_DIR=/path/to/data

---

Quick start

Generate a single logic game

from logicjitter import LogicJitter

lj = LogicJitter(multimodal=True, n_characters=3, seed=42)
print(lj.get_logic_game())
print(lj.get_logic_game(game_type="causal"))
print(lj.get_logic_game(game_type="maths"))

Build a HuggingFace dataset

from logicjitter import create_logicjitter_dataset

ds = create_logicjitter_dataset(
    n_samples=1000,
    preset="full",
    seed=42,
    output_dir="./data",
)
print(len(ds), ds[0]["ntp"])

Command line

After pip install logicjitter, use the logicjitter-generate command. While developing from a local clone, run pip install -e . first so the CLI is registered.

# Print demo samples (python -m works even without the console script)
python -m logicjitter --demo --n-samples 5 --preset full --no-multimodal

# Same via the installed entry point
logicjitter-generate --demo --n-samples 5 --preset full --no-multimodal

# Generate and save a dataset
logicjitter-generate --n-samples 1000 --preset full --output-dir ./data

---

Adding your own logic game

Each built-in game is a function with this signature:

def get_my_problem(error: bool = False, game_pieces: dict | None = None) -> dict:
    return {
        "problem": "...",           # statement shown in the sample
        "right_answers": ["yes"],   # phrases a correct character may say
        "wrong_answers": ["no"],    # phrases an incorrect character may say
        "game_pieces": {...},       # optional state for problem revisions
    }

When error=True, bake a mistake into the statement. When game_pieces is passed back on a revision round, update the underlying facts (e.g. change a number) while keeping the same characters.

Register the game and generate samples:

from logicjitter import LogicJitter

lj = LogicJitter(multimodal=False, seed=0)
lj.register_game("parity", get_parity_problem)
print(lj.get_logic_game(game_type="parity"))

See examples/custom_parity_game.py for a complete, runnable toy example (even/odd sums).

---

Package layout

logicjitter/
├── pyproject.toml
├── README.md
└── src/logicjitter/
    ├── __init__.py          # LogicJitter, create_logicjitter_dataset
    ├── cli.py               # logicjitter-generate entry point
    ├── dataset.py           # HuggingFace dataset pipeline
    ├── paths.py             # Package and cache paths
    └── creation/            # Logic game generators
        ├── logicjitter_generator.py
        ├── arithmetics_dataset.py
        ├── causal_dataset.py
        ├── cfg_dataset.py
        ├── clevr_dataset.py
        ├── clevr_tools/
        └── data/

---

Citation

@article{herranzcelotti2026logicjitter,
  title   = {LogicJitter: Let LLMs play and uncover misinformation},
  author  = {Herranz-Celotti, Luca and Viviani, Marco},
  journal = {Computers and Electrical Engineering},
  volume  = {136},
  pages   = {111215},
  year    = {2026},
  doi     = {10.1016/j.compeleceng.2026.111215},
  url     = {https://www.sciencedirect.com/science/article/pii/S0045790626002879}
}

---

License

Published under CC BY 4.0.