HalluWorld is a frozen benchmark. Its value depends on the items staying out of training corpora, so the bank is distributed behind this gate rather than from the public source repository. The data is licensed CC BY 4.0; by requesting access you acknowledge the requests below, which are asked of you as a benchmark user rather than imposed as license terms.

HalluWorld

The frozen question bank behind HalluWorld: A Controlled Benchmark for Hallucination via Reference World Models. 1,439 items across four tracks, plus 294 navigation trajectories for the in-navigation track. Code, docs, and the evaluation harness live at github.com/DegenAI-Labs/HalluWorld.

Track	Items	Levels / tasks	Cognitive tiers
grid	443	33	C:78 M:51 P:39 U:76 X:199
chess	350	7	C:100 M:50 P:150 X:50
innav	117	33	C:36 M:18 P:20 U:15 X:28
terminal	529	110	C:106 M:108 P:105 U:104 X:106

Tiers: P perceptual, M memory, C causal, U uncertainty, X compound. The chess battery has no U-tier probe by design.

Using it with the harness

The harness fetches this repository on first use. Accept the gate, log in once, and every halluworld command finds the bank on its own:

pip install halluworld
hf auth login                    # or export HF_TOKEN=...
halluworld questions verify      # downloads v0.1/questions/ and checks every checksum
halluworld eval terminal --provider openai --model gpt-4o-mini --out results/terminal

Air-gapped machines can point the harness at a local copy instead:

hf download DegenAI-Labs/HalluWorld --repo-type dataset --local-dir ./halluworld-hf
export HALLUWORLD_QUESTIONS_DIR=./halluworld-hf

Loading it as a dataset

from datasets import load_dataset

terminal = load_dataset("DegenAI-Labs/HalluWorld", "terminal", split="test")
everything = load_dataset("DegenAI-Labs/HalluWorld", split="test")   # the `all` config

Layout

Each version is one self-contained folder:

README.md                  this card
v0.1/
  README.md
  questions/               manifest.json + the frozen <track>.jsonl.gz files + trajectories
  data/                    flattened views of the same records, for load_dataset and the viewer
  configs/chess/           the chess run configs behind the published results
  raw/                     generation inputs the terminal and chess banks are rebuilt from
v0.2/                      held out (see below)
  README.md                how to rebuild v0.2 from the public repo
  questions/               the frozen v0.2 bank, both chess conditions
  configs/                 chess generation configs
  raw/grid/                the authored grid questions
  raw/terminal/            the terminal source probes

The v0.1/questions/ files are the release artifact: their digests are what halluworld questions verify checks and what published numbers trace back to. The v0.1/data/ views hold identical content but JSON-encode the fields whose type varies between rows (ground_truth, probe_kwargs, provenance, extra; messages and trajectory for traces) so every column has one Arrow type. ground_truth_type records what to decode it back to.

File	Records	SHA-256 (uncompressed)
`v0.1/questions/grid.jsonl.gz`	443	`23ca2082417aa5be10a1cf45f2aadb377339e1b431b3bd3ca3286366f44d6fad`
`v0.1/questions/chess.jsonl.gz`	350	`d56ba542c6176f113b2358d6720996fb241c042fc5a029323f085bf509a248f2`
`v0.1/questions/innav.jsonl.gz`	117	`66a8932bf10fa8d11614064e531f7d3e91daffd7cdb69e14be876c0dccffb0de`
`v0.1/questions/terminal.jsonl.gz`	529	`4092f850ef267e39d8cd8ffa1c339df2d300f6aafed8e47cd66994b581d7a102`
`v0.1/questions/trajectories.jsonl.gz`	294	`a5ef562a93d1ea17b294c59edfe78cf8ec70f70aa0dd697cdc8dadcf0196a89b`

Held-out v0.2

The held-out set lives in this gated dataset too, but not in the public GitHub repo, which holds v0.1 only. v0.2/questions/ is the frozen bank (1125 records): chess in two reported conditions, No-FEN (chess.jsonl.gz) and incorrect-FEN (chess_fen_transpose.jsonl.gz, the same questions with a corrupted FEN line), plus grid and terminal, all under one manifest.

halluworld eval chess --provider openai --model gpt-4o-mini --version v0.2 --out results/chess_v0.2
halluworld eval chess --provider openai --model gpt-4o-mini --version v0.2 --fen-mode transpose \
    --out results/chess_v0.2_transpose

The generation configs (v0.2/configs/), the authored grid source (v0.2/raw/grid/) and the terminal source probes (v0.2/raw/terminal/) are here as well. v0.2/README.md explains how to copy them into a checkout of the public repo and rebuild v0.2 from scratch. Please report v0.2 results without reposting its items.

Record schema

Every item shares one schema, defined in halluworld/questions.py:

Field	Meaning
`question_id`	stable, namespaced: `terminal/000498`, `grid/P4_harder_array/q07`
`track`, `suite`, `task_or_level`	where the item comes from
`kind`	`fixed`: question and ground truth are frozen literals. `generated`: rebuilt at run time from `probe_class` + `probe_kwargs` under a fixed seed (all InNav items, 9 grid items)
`question`, `ground_truth`, `answer_schema`	the item; `ground_truth` is a bool, int, str, list, or null depending on the probe
`context`	the observation the model is shown, when it is frozen with the item (terminal, chess)
`cognitive_tier`	P / M / C / U / X, the cross-track comparison axis
`failure_mode_target`	terminal-only fine-grained label; `unclassified` elsewhere
`difficulty`, `answerability`	1-5 generator scores where available
`provenance`, `extra`	track-specific payload

Terminal context fields are verbatim tmux pane captures from agent runs inside Terminal-Bench task containers, so they contain fragments of those tasks' files and command output. HalluWorld's CC BY 4.0 grant covers its questions, answers, labels, and curation; it does not relicense third-party material appearing inside captured contexts.

Citation

@article{liu2026halluworld,
  title   = {HalluWorld: A Controlled Benchmark for Hallucination via Reference World Models},
  author  = {Liu, Emmy and Gangal, Varun and Yu, Michael and Tao, Zhuofu and
             Singh, Karan and Kumar, Sachin and Feng, Steven Y.},
  journal = {arXiv preprint arXiv:2605.19341},
  year    = {2026}
}

Downloads last month: 7

Total file size:

102 MB

Paper for DegenAI-Labs/HalluWorld

HalluWorld: A Controlled Benchmark for Hallucination via Reference World Models

Paper • 2605.19341 • Published May 19