wolframko/betty-dota2-raw-v2

Betty Dota 2 Raw Dataset v2

Lossless delta-encoded extraction of Dota 2 professional match replays.

Structure

betty/
├── README.md              — this file
├── replays/               — original .dem.bz2 replay files (symlink)
├── raw/{match_id}/        — parsed data per match
│   ├── entities.parquet       — entity property deltas (all 200+ entity classes)
│   ├── entity_classes.parquet — class_id → class_name mapping
│   ├── property_dict.parquet  — prop_id → property_name mapping
│   ├── combat_log.parquet     — parsed combat log (all CMsgDOTACombatLogEntry fields)
│   ├── messages.parquet       — protobuf messages (55+ types, raw bytes)
│   └── string_tables.parquet  — CombatLogNames, instancebaseline
└── tools/
    └── clarity-parser.jar     — Java parser (requires JDK 17+)

Stats

• Matches: 9,388 professional Dota 2 matches (2025)
• Parser: skadistats/clarity 3.1.1 (Java) with native delta tracking
• Format: Apache Parquet + ZSTD compression
• Size ratio: ~1.07x of original replay size
• Parse speed: ~12 sec/match single-threaded

Entity Changes Schema (entities.parquet)

Column	Type	Description
tick	int32	Game tick
entity_id	int32	Entity index
entity_serial	int32	Entity serial (handle = serial << 14
class_id	int32	→ entity_classes.parquet
op	int32	0=create, 1=enter, 2=update, 3=leave, 4=delete
prop_id	int32?	→ property_dict.parquet
val_int	int64?	Integer/boolean values
val_uint64	int64?	Large unsigned values (handles, steam IDs)
val_float	float?	Float values
val_string	string?	String values
val_vec_x/y/z	float?	Vector components

Usage (Python)

import pyarrow.parquet as pq
import pandas as pd

match_id = "7596541883"
entities = pq.read_table(f"raw/{match_id}/entities.parquet")
classes = pq.read_table(f"raw/{match_id}/entity_classes.parquet")
props = pq.read_table(f"raw/{match_id}/property_dict.parquet")
combat_log = pq.read_table(f"raw/{match_id}/combat_log.parquet")

# Reconstruct hero HP timeline
class_map = {r["class_id"]: r["class_name"] for r in classes.to_pylist()}
prop_map = {r["prop_id"]: r["prop_name"] for r in props.to_pylist()}

Parsing a match

java -jar tools/clarity-parser.jar --replay replays/7596541883.dem.bz2 --output raw/7596541883

Batch parsing

java -Xmx512m -jar tools/clarity-parser.jar \
  --replay-dir replays/ --output raw/ --workers 32

Verification

java -jar tools/clarity-parser.jar \
  --verify --replay replays/7596541883.dem.bz2 --dataset raw/7596541883