{
 "body": "\n## What it measures\n\nPROST is a zero-shot English physical-reasoning quiz. Each item sets a tiny scene (\"a person drops a glass, a pillow, a coin, and a ball\") and asks which object is most likely to break, roll, or otherwise behave as named. Ten concepts and fourteen hand-written templates expand to 18,736 four-choice questions. Objects were chosen as single tokens for masked LMs. The authors' point is that this knowledge should come from pretraining, not from a PROST training split \u2014 there is none.\n\n## How it is scored\n\nlm-eval reports accuracy and length-normalised accuracy over A\u2013D. The paper's Table 5 is a macro average over the ten concepts, with some T5 and UnifiedQA rows dropping Slide because of tokenisation. Those two summaries are not the same number. Few-shot PROST would violate the paper's stated protocol (section 7 in the lm-eval README's words). Chance is 25% if labels are balanced; this page did not re-count label balance in the JSONL.\n\n## Dataset and licence\n\n18,736 test rows on `corypaik/prost`, Apache-2.0 on the Hub card, the nala-cub README, and the dataset_info licence field. Each row has cloze `question` and multiple-choice `ex_question`; lm-eval uses the latter. Answers are public.\n\n## Who publishes it\n\nSt\u00e9phane Aroca-Ouellette, Cory Paik, Alessandro Roncone, and Katharina Kann at the University of Colorado Boulder. Findings of ACL-IJCNLP 2021; arXiv:2106.03634 (2021-06-07). The GitHub repo is unmaintained and points to the Hub copy.\n\n## Lineage\n\nThe paper contrasts PROST with [piqa](piqa.md): PIQA covers many how-to situations, PROST covers a small set of objective physical concepts. Bangla PIQA is a translation of PIQA, not of PROST. No successor page is in this repository.\n\n## Saturation and contamination\n\nUnknown for current models. The 2021 table tops out at 46.7% macro for UnifiedQA-3B, well above 25% and well below 100%. Templates are public. Zero-shot intent does not stop a later corpus from including the JSONL.\n\n## How to run it\n\n`lm_eval --tasks prost`. Keep zero-shot. Do not train on the test file. Compare acc to acc_norm before mixing papers. There is no inspect_evals, HELM, OpenCompass, or BIG-bench task with this name in the paths checked.\n\n## Reading the numbers\n\nA 40% score means the model beats chance on these templated object questions, not that it has a physics engine. Superlative flips (most vs least) and option order are known failure modes in the paper. PIQA numbers are a different task. Treat any few-shot PROST figure as off-protocol unless the reporter says why.\n",
 "build": {
  "built_at": "2026-09-09T16:56:50+00:00",
  "commit": "0a599558854c0e238c03a0f0d725239cb28f9d11",
  "eligibility_as_of": "2026-09-09"
 },
 "disposition": {
  "canonical_id": "prost",
  "reasons": [],
  "status": "unassessed",
  "verified_results": []
 },
 "models_covered": [],
 "page": {
  "aliases": [
   "PROST",
   "Physical Reasoning about Objects Through Space and Time",
   "corypaik/prost"
  ],
  "category": "reasoning",
  "contamination": {
   "note": "Templates were written for this set and answers are public on Hugging Face (Apache-2.0). The paper argues zero-shot use so models are not trained on these templates. Everyday object facts may still appear in pretraining. lm-eval sets should_decontaminate on context plus question.\n",
   "risk": "medium"
  },
  "dataset": {
   "languages": [
    "en"
   ],
   "license": "Apache-2.0",
   "modalities": [
    "text"
   ],
   "public_test_set": true,
   "size": 18736,
   "size_note": "Paper, Hub card, datasets-server, and lm-eval README all give 18,736 test questions from 14 templates covering 10 concepts. No train split by design.\n",
   "splits": "test only (18,736 rows)",
   "url": "https://huggingface.co/datasets/corypaik/prost"
  },
  "freshness": {
   "researched": "2026-09-08",
   "researched_by": "Grok Build, batch-066 (Codex coordinated)",
   "reviewed": "2026-09-08",
   "reviewed_by": "Grok Build independent review, batch-066"
  },
  "harness": {
   "bigbench": "",
   "helm": "",
   "inspect_evals": "",
   "lm_eval": "prost",
   "opencompass": "",
   "other": "YAML file corypaik_prost.yaml; task name prost; dataset json from corypaik/prost"
  },
  "id": "prost",
  "last_updated": "2022-10",
  "leaderboard_url": "",
  "lineage": {
   "family": "",
   "predecessor": "",
   "successors": [],
   "variants": []
  },
  "measures": "PROST (Physical Reasoning about Objects Through Space and Time) asks which of four everyday objects best fits a short English scene. Concepts are direction, mass, height, circumference, stackable, rollable, graspable, breakable, slideable, and bounceable. Aroca-Ouellette, Paik, Roncone, and Kann wrote 14 templates and expanded them to 18,736 items. The paper and the lm-eval README require zero-shot use. It is not [piqa](piqa.md), which is two-way how-to commonsense from Instructables.\n",
  "metric": {
   "baseline_note": "Four choices, so uniform chance is 25%. Paper Table 5 macro averages (2021 models, some T5/UnifiedQA rows omit Slide): GPT 26.1, GPT-2 XL 30.1, RoBERTa-L 31.3, ALBERT-xlarge 31.8, UnifiedQA-3B 46.7 (starred). No human-accuracy figure is in the paper HTML opened here. Nine validators checked object order and affordance groups, not question accuracy.\n",
   "direction": "higher_is_better",
   "human_baseline": null,
   "max_score": 100,
   "name": "accuracy (acc); acc_norm also reported",
   "random_baseline": 25,
   "unit": "%"
  },
  "name": "PROST",
  "page_kind": "benchmark",
  "paper": {
   "arxiv": "2106.03634",
   "title": "PROST: Physical Reasoning about Objects through Space and Time",
   "url": "https://arxiv.org/abs/2106.03634",
   "year": 2021
  },
  "publisher": {
   "authors": [
    "St\u00e9phane Aroca-Ouellette",
    "Cory Paik",
    "Alessandro Roncone",
    "Katharina Kann"
   ],
   "org": "University of Colorado Boulder (NaLa lab)",
   "url": "https://github.com/nala-cub/prost"
  },
  "released": "2021-06",
  "repo_url": "https://github.com/nala-cub/prost",
  "saturation": {
   "as_of": "2021-08",
   "note": "Highest paper Table 5 macro average opened here is UnifiedQA-3B at 46.7 (starred; Slide omitted). That is a 2021 checkpoint, not a 2026 frontier figure. No later public leaderboard cell was read. Four-choice chance is 25.\n",
   "status": "unknown",
   "top_score": 46.7
  },
  "sources": [
   {
    "accessed": "2026-09-08",
    "title": "PROST arXiv abs (18,736 items, 14 templates, 10 concepts, 2021-06-07)",
    "url": "https://arxiv.org/abs/2106.03634"
   },
   {
    "accessed": "2026-09-08",
    "title": "ar5iv HTML (Table 5 macro averages, UnifiedQA-3B 46.7)",
    "url": "https://ar5iv.labs.arxiv.org/html/2106.03634"
   },
   {
    "accessed": "2026-09-08",
    "title": "ACL Anthology Findings ACL-IJCNLP 2021 landing page",
    "url": "https://aclanthology.org/2021.findings-acl.404/"
   },
   {
    "accessed": "2026-09-08",
    "title": "Hub card (18,736 test, Apache-2.0, cloze vs ex_question)",
    "url": "https://huggingface.co/datasets/corypaik/prost"
   },
   {
    "accessed": "2026-09-08",
    "title": "datasets-server: 18,736 test rows",
    "url": "https://datasets-server.huggingface.co/size?dataset=corypaik/prost"
   },
   {
    "accessed": "2026-09-08",
    "title": "nala-cub/prost README (Hub recommended; Apache-2.0; posterity repo)",
    "url": "https://raw.githubusercontent.com/nala-cub/prost/main/README.md"
   },
   {
    "accessed": "2026-09-08",
    "title": "lm-eval PROST README (zero-shot only, task name prost)",
    "url": "https://raw.githubusercontent.com/EleutherAI/lm-evaluation-harness/main/lm_eval/tasks/prost/README.md"
   },
   {
    "accessed": "2026-09-08",
    "title": "lm-eval prost YAML (acc, acc_norm, ex_question, decontaminate)",
    "url": "https://raw.githubusercontent.com/EleutherAI/lm-evaluation-harness/main/lm_eval/tasks/prost/corypaik_prost.yaml"
   }
  ],
  "status": "unknown",
  "subcategory": "zero-shot four-choice English physical reasoning over 10 object concepts",
  "summary": "18,736 four-choice English questions from 14 templates that probe physical reasoning about objects in space and time, meant to be used zero-shot.",
  "tags": [
   "physical-reasoning",
   "commonsense",
   "multiple-choice",
   "zero-shot"
  ],
  "task_format": "Four-option multiple choice. lm-eval task prost loads hf://datasets/corypaik/prost/data/default.jsonl, test split. Prompt is context, then \"Question: {ex_question}\", then \"Answer:\". Choices are fields A\u2013D. Target is the integer label. should_decontaminate true. Metrics acc and acc_norm. YAML filename corypaik_prost.yaml; runnable name is prost.\n"
 }
}