speedrunbench / README.md
akkikiki's picture
STK license: CC-BY-SA assets
7b00875 verified
|
Raw History Blame Contribute Delete
3.4 kB
---
license: other
tags:
- speedrun
- llm-agents
- benchmark
extra_gated_heading: "Access SpeedrunBench"
extra_gated_button_content: "Request access"
extra_gated_prompt: >-
SpeedrunBench measures LLM agents optimizing speedruns. This dataset contains input tapes
and silent video renders of individual measured runs, across both open-source games and
commercially published ones.
Copyright in the underlying games remains with their respective rights-holders. We publish
these excerpts as research evidence and make no claim to them. Access is granted for
research and evaluation purpose only.
extra_gated_fields:
I will use this data for research or evaluation purpose only: checkbox
I understand copyright in the underlying games belongs to their rights-holders: checkbox
configs:
- config_name: supertux
default: true
data_files:
- split: test
path: supertux/metadata.jsonl
- config_name: astray
data_files:
- split: test
path: astray/metadata.jsonl
- config_name: tuxemon
data_files:
- split: test
path: tuxemon/metadata.jsonl
- config_name: pokemon_blue
data_files:
- split: test
path: pokemon/metadata.jsonl
- config_name: sml
data_files:
- split: test
path: sml/metadata.jsonl
- config_name: smb
data_files:
- split: test
path: smb/metadata.jsonl
- config_name: wolf3d
data_files:
- split: test
path: wolf3d/metadata.jsonl
- config_name: civ
data_files:
- split: test
path: civ/metadata.jsonl
- config_name: mk64
data_files:
- split: test
path: mk64/metadata.jsonl
- config_name: stk
data_files:
- split: test
path: stk/metadata.jsonl
---
# SpeedrunBench
LLM agents optimizing speedruns across multiple games. Each run ships as a silent video of the run that
was scored, and (except for Tuxemon) the input tape that produced it: the frame to goal and lower
is always better.
| config | game | metric |
|---|---|---|
| `supertux` | SuperTux | frames to goal |
| `astray` | Astray | frames to maze exit |
| `tuxemon` | Tuxemon | frames to first gym |
| `pokemon_blue` | Pokémon Blue | frames to first badge |
| `sml` | Super Mario Land | frames to clear 1-1 |
| `smb` | Super Mario Bros | frames to 1-1 flagpole |
| `wolf3d` | Wolfenstein 3D | frames to boss kill (E3L9) |
| `civ` | Sid Meier's Civilization | frames to turn 10 |
| `mk64` | Mario Kart 64 | frames to finish 1 lap (Wario Stadium) |
| `stk` | SuperTuxKart | seconds to finish 1 lap (Black Forest) |
```python
load_dataset("PatronusAI/speedrunbench", "supertux", split="test")
```
## Licenses of the games
Subset of the titles are open source and carry their own terms, which continue to apply to any
footage of them:
| game | license |
|---|---|
| SuperTux | GPL-3.0 (code), CC-BY-SA (assets) |
| SuperTuxKart | GPL-3.0-or-later (code), CC-BY-SA (assets) |
| Tuxemon | GPL-3.0 (code), CC-BY-SA (assets) |
The remaining titles are commercially published, and rights in them stay with their rights-holders.
## Disclaimer
SpeedrunBench is for research and benchmarking purposes only. This project conveys no
right to access, reproduce, distribute, modify, or commercially exploit any third-party title or
its assets beyond what applicable licenses, terms of service, and law already permit. Users are
responsible for lawfully obtaining the titles they evaluate on and for securing any permissions
their intended use requires.