fchaubard/funny_bench

---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- comedy
- stand-up
- humor
- dpo
- sft
- funny
size_categories:
- 10K<n<100K
dataset_info:
features:
- name: text
dtype: string
- name: source
dtype: string
splits:
- name: train
num_bytes: 64671620
num_examples: 369940
- name: test
num_bytes: 1321826
num_examples: 7550
download_size: 37816655
dataset_size: 65993446
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
---

# FunnyBench: Stand-Up Comedy Dataset for LLM Training

LLMs aren't funny. This initiative tries to solve that. A curated dataset of **24,438 stand-up comedy transcripts** with engagement metrics, designed for teaching LLMs to generate funny content. Shout out to Shofo (https://www.shofo.ai/) for providing the dataset for free for public use!

## Dataset Splits

### SFT (Supervised Fine-Tuning)
- **23,216 train** / **1,222 test** examples
- Chat format with quality-tier conditioning
- Fields: `messages`, `tier`, `engagement_rate`, `like_count`, `play_count`, `duration_seconds`, `video_id`, `author`

```python
from datasets import load_dataset
ds = load_dataset("fchaubard/funny_bench", "sft")
```

### DPO (Direct Preference Optimization)
- **11,607 train** / **611 test** preference pairs
- Pairs matched by duration bucket
- "Chosen" = higher engagement rate, "Rejected" = lower engagement rate
- Fields: `prompt`, `chosen`, `rejected`, `chosen_engagement`, `rejected_engagement`

```python
from datasets import load_dataset
ds = load_dataset("fchaubard/funny_bench", "dpo")
```

## Source Data

- **29,729 TikTok stand-up comedy clips** (pre-filtered: 10,000+ likes, English, standup hashtags)
- Transcribed using NVIDIA Canary-Qwen 2.5B
- Speaker diarization via NVIDIA NeMo MSDD
- Labels: `[COMEDIAN]`, `[AUDIENCE]`, `[LAUGHTER]`

## Cleaning Pipeline

| Filter | Threshold | Dropped |
|--------|-----------|---------|
| Transcript length | 80-12,000 chars | 1,163 |
| Duration | 15-300 seconds | 2,098 |
| Word repetition score | <= 0.55 | 1,954 |
| Unique word count | >= 15 | 54 |
| ASR garbage detection | trigram loops | 22 |
| **Total removed** | | **5,291 (17.8%)** |
| **Clean dataset** | | **24,438 (82.2%)** |

## Quality Tiers (SFT)

Each SFT example has a quality tier based on engagement rate (likes/views):

| Tier | Percentile | Engagement Rate | Count |
|------|-----------|----------------|-------|
| `[LEGENDARY]` | Top 5% | > 21.8% | 1,222 |
| `[KILLER]` | 75-95th | 14.7-21.8% | 4,888 |
| `[SOLID]` | 50-75th | 10.8-14.7% | 6,109 |
| `[WARMING_UP]` | Bottom 50% | < 10.8% | 12,219 |

At inference, prompt with `[LEGENDARY]` to generate top-tier comedy.

## Why Engagement Rate?

Raw like counts are dominated by virality and follower counts. The engagement rate (likes/views) better captures per-viewer funniness. A clip with 1M views and 200K likes (20%) is funnier per-viewer than one with 100M views and 5M likes (5%).

## SFT Format

```json
{
"messages": [
{"role": "system", "content": "You are a stand-up comedian performing a live set..."},
{"role": "user", "content": "[LEGENDARY] Perform a stand-up comedy bit."},
{"role": "assistant", "content": "[COMEDIAN]: So where are you from?\n[AUDIENCE]: Texas!\n[COMEDIAN]: Texas? Oh man...\n[LAUGHTER]"}
],
"tier": "LEGENDARY",
"engagement_rate": 0.22,
"like_count": 500000,
"play_count": 2200000
}
```

## DPO Format

```json
{
"prompt": [
{"role": "system", "content": "You are a stand-up comedian..."},
{"role": "user", "content": "Perform a stand-up comedy bit."}
],
"chosen": [{"role": "assistant", "content": "...funnier transcript..."}],
"rejected": [{"role": "assistant", "content": "...less funny transcript..."}],
"chosen_engagement": 0.18,
"rejected_engagement": 0.05
}
```

## Limitations

- ASR artifacts from NVIDIA Canary-Qwen 2.5B transcription
- Comedy depends heavily on delivery and timing that text can't capture
- TikTok bias toward short-form, punchy comedy
- Engagement != funny (controversy and relatability also drive engagement)

## Citation

If you use this dataset, please cite:
```
@misc{funnybench2026,
title={FunnyBench: Teaching LLMs Stand-Up Comedy with Engagement-Based Preference Learning},
year={2026}
}
```