lara-martin
commited on
Commit
•
c9ea0be
1
Parent(s):
ed87604
initial information
Browse files
README.md
CHANGED
@@ -1,3 +1,151 @@
|
|
1 |
-
---
|
2 |
-
license: cc-by-4.0
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: cc-by-4.0
|
3 |
+
task_categories:
|
4 |
+
- text-generation
|
5 |
+
- text2text-generation
|
6 |
+
language:
|
7 |
+
- en
|
8 |
+
tags:
|
9 |
+
- story
|
10 |
+
- storytelling
|
11 |
+
- story generation
|
12 |
+
- dnd
|
13 |
+
- creative generation
|
14 |
+
- command generation
|
15 |
+
- dungeons and dragons
|
16 |
+
- ttrpg
|
17 |
+
- dungeon master
|
18 |
+
pretty_name: FIREBALL
|
19 |
+
configs:
|
20 |
+
- config_name: default
|
21 |
+
data_files:
|
22 |
+
- split: filtered
|
23 |
+
path: "filtered/*.jsonl.gz"
|
24 |
+
dataset_info:
|
25 |
+
features:
|
26 |
+
- name: combat_id
|
27 |
+
dtype: string
|
28 |
+
- name: event_type
|
29 |
+
dtype: string
|
30 |
+
- name: timestamp
|
31 |
+
dtype: float
|
32 |
+
- name: message_id
|
33 |
+
dtype: string
|
34 |
+
- name: author_id
|
35 |
+
dtype: string
|
36 |
+
- name: author_name
|
37 |
+
dtype: string
|
38 |
+
- name: created_at
|
39 |
+
dtype: float
|
40 |
+
- name: content
|
41 |
+
dtype: string
|
42 |
+
- name: embeds
|
43 |
+
dtype: list
|
44 |
+
- name: proxy_url
|
45 |
+
dtype: string
|
46 |
+
- name: fields
|
47 |
+
dtype: list
|
48 |
+
- name: components
|
49 |
+
dtype: list
|
50 |
+
language_creators:
|
51 |
+
- crowdsourced
|
52 |
+
---
|
53 |
+
|
54 |
+
# Dataset Card for FIREBALL
|
55 |
+
|
56 |
+
## Table of Contents
|
57 |
+
- [Data Description](#data-description)
|
58 |
+
- [Filtered Triplets Schema](#filtered-triplets-schema)
|
59 |
+
- [Normalized Actor State](#normalized-actor-state)
|
60 |
+
- [Additional Information](#additional-information)
|
61 |
+
- [Citation](#citation)
|
62 |
+
- [Licensing](#licensing)
|
63 |
+
|
64 |
+
|
65 |
+
---
|
66 |
+
## Data Description
|
67 |
+
**FIREBALL: A Dataset of Dungeons and Dragons Actual-Play with Structured Game State Information**
|
68 |
+
|
69 |
+
FIREBALL is a large crowdsourced dataset of people playing Dungeons and Dragons on Discord. In addition to playing the game using natural language (primarily English), players also used a bot called [Avrae](https://avrae.io/). Avrae enables players to keep track of the state of the game by writing commands, which we collected.
|
70 |
+
|
71 |
+
The data contains nearly 25,000 unique sessions of gameplay.
|
72 |
+
|
73 |
+
|
74 |
+
* [Published paper](https://aclanthology.org/2023.acl-long.229/)
|
75 |
+
* [Paper on arXiv](https://arxiv.org/abs/2305.01528)
|
76 |
+
|
77 |
+
**Abstract**
|
78 |
+
> Dungeons & Dragons (D&D) is a tabletop roleplaying game with complex natural language interactions between players and
|
79 |
+
> hidden state information. Recent work has shown that large language models (LLMs) that have access to state
|
80 |
+
> information can generate higher quality game turns than LLMs that use dialog history alone. However, previous work
|
81 |
+
> used game state information that was heuristically created and was not a true gold standard game state. We present
|
82 |
+
> FIREBALL, a large dataset containing nearly 25,000 unique sessions from real D&D gameplay on Discord with true game
|
83 |
+
> state info. We recorded game play sessions of players who used the Avrae bot, which was developed to aid people in
|
84 |
+
> playing D&D online, capturing language, game commands and underlying game state information. We demonstrate that
|
85 |
+
> FIREBALL can improve natural language generation (NLG) by using Avrae state information, improving both automated
|
86 |
+
> metrics and human judgments of quality. Additionally, we show that LLMs can generate executable Avrae commands,
|
87 |
+
> particularly after finetuning.
|
88 |
+
|
89 |
+
|
90 |
+
|
91 |
+
### Filtered Triplets Schema
|
92 |
+
All user IDs and usernames have been randomized (by way of a hash function) to preserve anonymity.
|
93 |
+
Each line contains a filtered triple, each of which includes the following keys:
|
94 |
+
```
|
95 |
+
{
|
96 |
+
"speaker_id": The anonymized user ID of the user who sent the commands in the triple.
|
97 |
+
"before_utterances": A list of strings corresponding to the "preceding" utterances in the triple.
|
98 |
+
"combat_state_before": A list of normalized actor states (see below) for each actor in the combat instance at the instant before the command was run.
|
99 |
+
"current_actor": (nullable) The normalized actor state of the actor whose turn it currently is.
|
100 |
+
"commands_norm": A list of strings corresponding to the "commands" portion of the triple.
|
101 |
+
"automation_results": A mechanically generated list of strings representing the results of running the action in the Avrae engine.
|
102 |
+
"caster_after": The normalized actor state of the actor who ran the action(s), which may or may not be the current actor.
|
103 |
+
"targets_after": A list of normalized actor states for each actor who was targeted by the action.
|
104 |
+
"combat_state_after": A list of normalized actor states for each actor in the combat instance at the instant after the command was run.
|
105 |
+
"after_utterances": A list of strings corresponding to the "following" utterances in the triple.
|
106 |
+
"utterance_history": The last 5 messages in the chat history before the command was run.
|
107 |
+
"before_idxs": A list of integers corresponding to the index of the "message" events containing the "preceding" utterances in the raw event file.
|
108 |
+
"before_state_idx": The index of the "combat_state_update" event in the raw event file that was used to derive "combat_state_before".
|
109 |
+
"command_idxs": The indexes of the "command" events corresponding to the "commands_norm" key.
|
110 |
+
"after_state_idx": The index of the "combat_state_update" event corresponding to the "combat_state_after" key.
|
111 |
+
"after_idxs": The indexes of the "message" events corresponding to the "after_utterances" key.
|
112 |
+
"embed_idxs": (nullable, same length as "automation_results") The indexes of "message" events corresponding to rich results shown to players on Discord for each result in the "automation_results" key.
|
113 |
+
}
|
114 |
+
```
|
115 |
+
|
116 |
+
#### Normalized Actor State
|
117 |
+
The normalized actor state is only a subset of the available actor information, corresponding to the information we used for our engineering experiments for the FIREBALL paper. For a full list of available actor information, see table 6 in the [FIREBALL paper](https://aclanthology.org/2023.acl-long.229/).
|
118 |
+
```
|
119 |
+
{
|
120 |
+
"name": The name of the actor.
|
121 |
+
"hp": The numerical and narrative hit points (e.g. "<12/34; Bloodied>").
|
122 |
+
"class": The actor's class(es) and level(s), if applicable (e.g. "Fighter 3")
|
123 |
+
"race": The actor's race, if applicable (e.g. "Mountain Dwarf", "Adult Red Dragon").
|
124 |
+
"attacks": A list of the actor's available attack names.
|
125 |
+
"spells": A list of the actor's available spells.
|
126 |
+
"actions": A list of the actor's available special abilities.
|
127 |
+
"effects": A list of any temporary effects on the actor (e.g. "Stunned").
|
128 |
+
"description": The actor's narrative description (if available).
|
129 |
+
"controller_id": The anonymized user ID of this actor's controller.
|
130 |
+
}
|
131 |
+
```
|
132 |
+
|
133 |
+
## Additional Information
|
134 |
+
### Citation
|
135 |
+
```
|
136 |
+
@inproceedings{Zhu2023FIREBALL,
|
137 |
+
title={{FIREBALL: A Dataset of Dungeons and Dragons Actual-Play with Structured Game State Information}},
|
138 |
+
author={Zhu, Andrew and Aggarwal, Karmanya and Feng, Alexander and Martin, Lara J. and Callison-Burch, Chris},
|
139 |
+
year={2023},
|
140 |
+
booktitle={Annual Meeting of the Association for Computational Linguistics (ACL)},
|
141 |
+
month={7},
|
142 |
+
url={https://aclanthology.org/2023.acl-long.229/},
|
143 |
+
address={Toronto, Canada},
|
144 |
+
pages={4171--4193},
|
145 |
+
publisher={ACL},
|
146 |
+
doi={10.18653/v1/2023.acl-long.229}
|
147 |
+
}
|
148 |
+
```
|
149 |
+
---
|
150 |
+
### Licensing
|
151 |
+
The Creative Commons Attribution 4.0 International License. https://creativecommons.org/licenses/by/4.0/
|