README.md

metadata

language:
  - en
license: cc-by-4.0
task_categories:
  - text-generation
tags:
  - humor
  - benchmark
  - evaluation
pretty_name: Humor Transfer Bench (HTB)
size_categories:
  - n<1K
dataset_info:
  features:
    - name: id
      dtype: string
    - name: domain
      dtype: string
    - name: prompt
      dtype: string
  config_name: default
  splits:
    - name: test
      num_examples: 400

Humor Transfer Bench (HTB)

Full name: Humor Transfer Bench
Paper: HumorGen (Ajayi & Mitra, 2026)

---

What HTB is for

HTB is an evaluation-only set of 400 English input prompts for testing humor generation in language models.

Each row gives you an input prompt. You pass it to your model with your own system/user template and decoding setup—HTB does not include or require a fixed generation instruction.

HTB is not for training. Use it to compare models on the same inputs across eight prompt categories.

---

File: htb_dataset.tsv

Format	TSV, UTF-8, tab-separated
Rows	400 (+ header)
Columns	id, domain, prompt

id format: HTB_{A–H}_{001–050} (e.g. HTB_A_001)

---

Domains (as in the file)

50 prompts per domain. The domain column is the label stored in the dataset.

ID letter	domain value	Example prompt
A	Fun Facts	Apples float in water because they are 25 percent air.
B	Daily Life	My houseplant is judging my life choices.
C	Social Terms	Generational amnesia.
D	Object Voices	My calendar shrieked: You have no free time!
E	Fantasy Creatures	A genie grants a wish for more wishes.
F	Twisted Definitions	A game is a challenge you can accept.
G	Direct Prompts	Tell a joke about emojis.
H	News Headlines	Researchers find link between sleep quality and memory retention

The HumorGen paper appendix discusses the design rationale for these eight categories in more detail.

---

How to use

1. Load htb_dataset.tsv.
2. For each row, feed prompt to your model with whatever generation prompt / chat template you use.
3. Evaluate outputs with your chosen method (pairwise judging, human eval, etc.) and report your protocol.

The HumorGen paper reports one such setup (15 models, Llama-3.3-70B and Qwen-2.5-72B judges). That is their protocol, not a requirement of the benchmark.

---

License

Domains A–G: CC BY 4.0 (HTB-authored prompts).

Domain H (News Headlines): Headlines sourced from publicly available BBC News text, with basic content filtering applied. Used as evaluation prompts only.

---

Citation

If you use HTB, cite the HumorGen paper:

@misc{ajayi2026humorgen,
  title        = {HumorGen: Cognitive Synergy for Humor Generation in Large Language Models via Persona-Based Distillation},
  author       = {Ajayi, Edward and Mitra, Prasenjit},
  year         = {2026},
  eprint       = {2604.09629},
  archivePrefix= {arXiv},
  primaryClass = {cs.CL},
  url          = {https://arxiv.org/abs/2604.09629}
}

---

Authors

Edward Ajayi, Prasenjit Mitra — Carnegie Mellon University Africa