---
license: mit
task_categories:
- question-answering
language:
- en
tags:
- reasoning
- math
- science
- commonsense
size_categories:
- 1K<n<10K
---

# Reasoning Spectrum QA Dataset

This is a unified dataset of exactly 1,000 examples designed for evaluating reasoning capabilities across different dimensions (factual, commonsense, science, arithmetic, multi-hop, and extractive span).

## Dataset Purpose
The dataset brings together questions of varying difficulty and reasoning family types under a single unified schema to facilitate standardized testing and evaluation of large language models.

## Source Datasets Used and Distributions
- **Factual**: `stanfordnlp/web_questions` (120 selected)
- **Commonsense**: `tau/commonsense_qa` (120 selected)
- **Science**: `allenai/ai2_arc` (Easy & Challenge) and `allenai/sciq` (250 selected)
- **Arithmetic**: `openai/gsm8k` (200 selected)
- **Multi-hop**: `hotpotqa/hotpot_qa` (220 selected)
- **Extractive Control**: `rajpurkar/squad` (90 selected)

## Schema
Each row in the dataset matches the following schema:
- `id`: A unique ID formatted as `qa_mix_XXXXXX`.
- `source_dataset`: The name of the source dataset.
- `source_split`: The split of the source dataset.
- `source_id`: The ID of the question in the source dataset.
- `question`: The cleaned, normalized text of the question.
- `context`: Supporting context or text if needed (e.g. SQuAD, HotpotQA, SciQ support).
- `choices`: Shuffled multiple choice option dictionary (labels and text) or null.
- `gold_answer`: The final target answer text.
- `gold_answer_label`: The correct choice option letter (e.g., "A", "B") or null.
- `reference_solution`: Reasoning explanation/solutions (e.g., from GSM8K) or null.
- `supporting_facts`: Original supporting facts (e.g., from HotpotQA) or null.
- `input_mode`: Input format representation (`question_only`, `question_with_choices`, `question_with_context`, or `question_with_context_and_choices`).
- `answer_type`: The answer format category (`boolean`, `number`, `entity`, `short_phrase`, `multiple_choice`, or `extractive_span`).
- `reasoning_family`: The category of reasoning needed (`factual`, `commonsense`, `science`, `arithmetic`, `multihop`, or `extractive_control`).
- `question_token_count`: Number of tokens in the question computed using the embedding tokenizer.
- `selection_similarity`: Cosine similarity to the closest selected question in the final dataset.
- `closest_selected_id`: ID of the closest selected question in the final dataset.

## Selection and Deduplication Method
1. **Candidate Collection**: Streaming up to 4x the target quota per reasoning family.
2. **Cheap Filters**:
   - Rejection of malformed, non-English, and image/table-based questions.
   - Exact global normalized duplicate rejection.
   - Arithmetic template normalization (placeholder replacement for names, numbers, and pronouns) to reject repeated structural question templates.
   - Frequency limits on named entities in factual questions (max 3 per entity).
3. **Semantic Diversity Selection**:
   - Embeddings computed via `all-MiniLM-L6-v2`.
   - Greedy global max-min cosine similarity selection.
   - Hard rejection of cosine similarity >= 0.90.
   - Heuristic review for cosine similarity between 0.84 and 0.90 (retained only when answers/reasoning differ).

## Limitations
- Source datasets are restricted to English.
- Extractive span boundaries might be source-dependent.
- Simple heuristic parser was used for template and entity matching.

## Source License Notes
- `web_questions`: CC-BY 4.0
- `commonsense_qa`: MIT License
- `ai2_arc`: CC-BY-SA 4.0
- `sciq`: CC-BY-NC 4.0
- `gsm8k`: MIT License
- `hotpot_qa`: CC-BY-SA 4.0
- `squad`: CC-BY-SA 4.0