bertybaums/marc2

---
license: mit
task_categories:
- text-generation
- visual-question-answering
language:
- en
tags:
- arc-agi
- abstract-reasoning
- metaphor
- figurative-language
- marc
pretty_name: "MARC2: Metaphor Abstraction and Reasoning Corpus v2"
size_categories:
- 1K<n<10K
---

# MARC2: Metaphor Abstraction and Reasoning Corpus v2

MARC2 extends the [MARC-from-LARC](https://huggingface.co/datasets/bertybaums/marc-from-larc) methodology to the [ARC-AGI2](https://github.com/arcprize/ARC-AGI-2) dataset. It provides a corpus of figurative language puzzles where metaphorical descriptions help AI models solve abstract reasoning tasks they cannot solve from examples alone.

## The MARC Property

A task has the **MARC property** (for a given model) when:
1. **Examples alone fail** — the model cannot solve the task from input/output examples
2. **Figurative description alone fails** — the metaphor is too ambiguous without examples
3. **Figurative + examples succeeds** — the metaphor triggers an "aha" moment when combined with examples

## Pipeline

| Phase | Result |
|-------|--------|
| Claude Opus 4.6 solves ARC-AGI2 training tasks | 865/1000 (86.5%) |
| Distill reasoning into language-complete descriptions | 865/865 (100%) |
| Validate descriptions (fresh solver, no examples) | 791/865 (91.4%) |
| Baseline testing on gpt-oss-120b (3 conditions) | 2,373 trials |
| Task classification | 350 MARC-eligible |
| Generate figurative descriptions | 350 original + 1,560 alternatives |
| MARC verification | **104 puzzles, 824 MARC-valid clues** |

## Dataset Configs

| Config | Rows | Description |
|--------|------|-------------|
| `tasks` | 1,120 | ARC-AGI2 task metadata |
| `solve_trials` | 1,007 | Claude's solving attempts with reasoning traces |
| `descriptions` | 865 | Language-complete see/do/grid descriptions |
| `task_subsets` | 791 | Per-model task classification |
| `figurative_descriptions` | 1,910 | Figurative clues (original + 15 domain alternatives) |
| `baseline_trials` | 2,373 | Subject model baseline results |
| `figurative_trials` | 8,645 | Subject model figurative trial results |

## Source Domains (15)

biology, cooking, music, sports, weather, architecture, warfare, theater, gardening, astronomy, ocean/sailing, electronics, mythology, dance, geology

## Key Findings

- **Language descriptions dramatically outperform examples**: 58.2% vs 25.8% accuracy for gpt-oss-120b
- **Opacity-guided metaphor generation** improved MARC yield from 29.7% to 47.6%
- **824 MARC-valid figurative clues** across 104 puzzles and 15 source domains
- Average 6.9 MARC-valid variants per puzzle

## Links

- **Code**: [github.com/bertybaums/marc2](https://github.com/bertybaums/marc2)
- **Parent project**: [MARC-from-LARC](https://github.com/bertybaums/marc-from-larc)
- **ARC-AGI2**: [github.com/arcprize/ARC-AGI-2](https://github.com/arcprize/ARC-AGI-2)

## Citation

```bibtex
@dataset{baum2026marc2,
title={MARC2: Metaphor Abstraction and Reasoning Corpus v2},
author={Baum, Bert},
year={2026},
url={https://huggingface.co/datasets/bertybaums/marc2},
doi={10.5281/zenodo.19241782}
}
```

## Date

March 26, 2026