wolframko/betty-dota2

Betty Dota 2 — Decision Context Dataset

Overview

9,385 professional Dota 2 matches parsed from replay files (.dem) into a rich, per-second decision context: hero states, ability cooldowns, building HP, combat events, modifiers, ward placements, and objectives.

Built to train Transformer and RL models that understand the game state at each moment in time.

Dataset Structure

matches.parquet — 9,385 rows

One row per match. Match metadata, STRATZ player statistics, and draft data.

Column	Type	Description
match_id	int64	Dota 2 match ID
league_name	string	Tournament name
league_tier	string	PROFESSIONAL, MAJOR, etc.
start_time	int64	Unix timestamp
duration_sec	int64	Match duration in seconds
radiant_win	bool	True if Radiant won
radiant_team_id / dire_team_id	int64	Team IDs
p0..p9_hero_id	int64	Hero ID per player
p0..p9_kills/deaths/assists	int64	Final KDA per player
p0..p9_gpm/xpm/networth	int64	Final economy stats
p0..p9_position / p0..p9_role	string	Lane/role from STRATZ
p0..p9_gpm_per_min	list[int64]	Gold per minute time series
p0..p9_networth_per_min	list[int64]	Net worth per minute
p0..p9_hero_damage_per_min	list[int64]	Hero damage per minute
p0..p9_tower_damage_per_min	list[int64]	Tower damage per minute
p0..p9_last_hits_per_min	list[int64]	Last hits per minute
stratz_pb_hero_id	list[int64]	STRATZ pick/ban hero IDs
replay_pb_hero_id	list[int64]	Replay-parsed pick/ban hero IDs

---

ticks/ — ~211,000,000 rows

Per-second hero state snapshot. One row per hero per game second (~10 heroes × ~2,100 ticks per match).

Column	Type	Description
match_id	int64	Match ID
tick	int32	Raw engine tick
game_time	double	Game time in seconds
slot	int8	Player slot (0–9, 0–4 Radiant, 5–9 Dire)
hero	string	Hero internal name (e.g. npc_dota_hero_invoker)
x, y	double	Map position
hp, max_hp	int32	Current and max health
mana, max_mana	double	Current and max mana
gold	int32	Current unreliable gold
net_worth	int32	Total net worth
xp	int32	Total experience
is_alive	bool	Whether hero is alive this tick
respawn_seconds	double	Seconds until respawn (0 if alive)
kills / deaths / assists	int16	KDA at this moment
last_hits / denies	int16	CS at this moment
level	int8	Hero level
str / agi / int_	double	Current attribute values
armor	double	Physical armor
move_speed	int16	Movement speed
item_0..item_5	string	Items in inventory slots

---

abilities/ — ~720,000,000 rows

Per-second ability state for every ability on every hero. One row per ability per hero per tick (~20 abilities × 10 heroes × 2,100 ticks per match).

Column	Type	Description
match_id	int64	Match ID
tick	int32	Raw engine tick
game_time	double	Game time in seconds
slot	int8	Player slot
hero	string	Hero internal name
ability_points	int16	Unspent ability points
ability_name	string	Ability internal name
ability_level	int16	Current level (0 = unlearned)
cooldown	double	Remaining cooldown in seconds
cooldown_length	double	Full cooldown duration
charges	int16	Charges remaining (charge-based abilities)
mana_cost	int32	Mana cost to cast
activated	bool	Whether ability is toggled on

---

buildings/ — ~465,000,000 rows

Per-second state of all towers, barracks, and the ancient. One row per building per tick.

Column	Type	Description
match_id	int64	Match ID
tick	int32	Raw engine tick
game_time	double	Game time in seconds
entity_name	string	Building internal name
entity_type	string	tower, barracks, fort
team	int16	2 = Radiant, 3 = Dire
lane	string	top, mid, bot, base
tier	int16	Tower tier (1–4)
hp	int32	Current HP
max_hp	int32	Max HP
alive	bool	Whether building is alive

---

events/ — ~287,000,000 rows

Combat log events: damage, deaths, gold gains, XP, healing, purchases.

Column	Type	Description
match_id	int64	Match ID
tick	int32	Raw engine tick
game_time	double	Game time in seconds
timestamp	double	Exact event timestamp
event_type	string	DOTA_COMBATLOG_DAMAGE, _DEATH, _GOLD, _HEAL, _PURCHASE, _XP
source	string	Source entity
target	string	Target entity
value	int32	Amount (damage/gold/xp/heal)
inflictor	string	Ability or item that caused the event
gold_reason / xp_reason	int32	Reason code for gold/XP events
item	string	Item name (PURCHASE events)

---

modifiers/ — ~196,000,000 rows

Modifier (buff/debuff) add and remove events. One row per modifier lifecycle event.

Column	Type	Description
match_id	int64	Match ID
timestamp	double	Event timestamp
modifier_type	string	add, remove, stack
slot	int8	Affected player slot
actor	string	Caster entity
target	string	Target entity
modifier_name	string	Modifier internal name
modifier_category	string	stun, silence, slow, buff, debuff, etc.
duration	double	Total duration in seconds
elapsed_duration	double	Time elapsed when event fired
stack_count	int32	Stack count at event time
hidden / invisibility / silence / root / aura / armor_debuff	bool	Effect flags

---

objectives/ — ~176,000 rows

Objective events: tower kills, rune pickups, aegis, barracks, Roshan.

Column	Type	Description
match_id	int64	Match ID
timestamp	double	Event timestamp
objective_type	string	building_kill, rune_pickup, aegis, roshan_kill, etc.
slot	int8	Player slot
actor / target	string	Entities involved
team	int16	Acting team
entity_type / entity_subtype	string	Building or rune type
lane / tier	string/int16	For building kills
rune_type	string	For rune pickups
value	int32	Gold bounty or other numeric value
x, y	double	Map position

---

wards/ — ~971,000 rows

Ward placement and destruction events.

Column	Type	Description
match_id	int64	Match ID
timestamp	double	Event timestamp
ward_event_type	string	placed, destroyed, expired
slot	int8	Placing player slot
actor / target	string	Entities involved
team	int16	Placing team
ward_type	string	observer, sentry
owner_slot	int8	Original owner slot
x, y	double	Map position

---

Statistics

Table	Rows	Size
matches	9,385	—
ticks	~211,000,000	4.1 GB
abilities	~720,000,000	503 MB
buildings	~465,000,000	332 MB
events	~287,000,000	2.3 GB
modifiers	~196,000,000	1.6 GB
actions	~29,000,000	271 MB
objectives	~176,000	7.4 MB
wards	~971,000	14 MB
Total	~1.9 billion	~9.1 GB

---

Usage

import pandas as pd

# Match metadata
matches = pd.read_parquet("hf://datasets/wolframko/betty-dota2/matches.parquet")
print(f"Matches: {len(matches)}, Radiant winrate: {matches.radiant_win.mean():.1%}")

# Hero states (streaming — large table)
from datasets import load_dataset
ticks = load_dataset(
    "wolframko/betty-dota2",
    data_files="ticks/*.parquet",
    split="train",
    streaming=True,
)

# Ability cooldowns for a single match
import pyarrow.parquet as pq
import pyarrow.dataset as ds

abilities = ds.dataset("hf://datasets/wolframko/betty-dota2/abilities/", format="parquet")
match_abilities = abilities.to_table(
    filter=ds.field("match_id") == 8106964131
).to_pandas()

---

Data Pipeline

1. Fetch — Match IDs from STRATZ GraphQL API (professional leagues 2025)
2. Download — Replay files (.dem.bz2) from Valve CDN
3. Parse — Per-second hero states, ability states, building HP, combat log, modifiers, wards, objectives via manta (Go)
4. Enrich — Player/team/draft stats from STRATZ API
5. Convert — JSON → Parquet (snappy compression, 10 matches/shard)

Source code: github.com/wolframko/betty — branch decision-context-v1

---

License

MIT