Datasets:

batrugg
/

dota-2-pro-tournament-drafts

Every Captains-Mode draft from four top-tier Dota 2 tournaments, in clean long/wide tables — plus a column nobody else publishes: the archived pre-series win probability of a calibrated prediction model, recorded before each TI 2026 series was played.

Curated by batru.gg — calibrated Dota 2 / Deadlock / Marvel Rivals win prediction and meta analytics.

Coverage

tournament	event	games
`ti-2025`	The International 2025 (Hamburg)	144
`ewc-2025`	Esports World Cup 2025 (Riyadh)	89
`ewc-2026`	Esports World Cup 2026 (Paris)	157
`ti-2026`	The International 2026 (Shanghai)	148 (complete — won by Team Spirit, 3–2 over TEAM VISION off a full lower-bracket run)

538 games, 12,912 draft actions total — every Captains-Mode game of all four events. Completeness was verified against both OpenDota league records and Liquipedia's per-event game counts; two drafts OpenDota lost were recovered from Liquipedia and cross-validated (see notes below). The only official result not included is EWC 2026's group-D 1v1 Shadow Fiend tiebreaker (TaiLung vs Abed) — a solo mid duel, not a Captains-Mode game.

Files

games.parquet / games.csv — one row per game: tournament, phase, date, series id, best-of, teams, sides, first-pick side, winner, and model_pre_series_p_team1 (TI 2026 only; see below).
draft_steps.parquet / draft_steps.csv — one row per Captains-Mode action (24 per game): step order (1–24), pick/ban, side, team, hero.
hero_stats.csv — per-tournament per-hero aggregates: picks, bans, contests, contest rate, W–L when picked.

The model column (and its honesty rules)

model_pre_series_p_team1 is the probability (0–1) that series_team1 wins a game, produced by batru.gg's pro-team model and archived before the series started (TI 2026 was tracked live on a 30-minute cron; 86 of the 148 games carry an archived probability — series that started before tracking began carry null, not a backfilled guess). Retro-compiled tournaments (2025 events, EWC 2026) deliberately carry null — computing a "prediction" after the result is known would be a fabricated track record. The model's public evaluation ledger (every weekly champion-vs-challenger test, wins and losses) is at batru.gg/dota2/benchmark.

Provenance & licensing

Match facts derive from official match data via the OpenDota API; draft sequences are normalized to the standard Captains-Mode step pattern (24 steps: 14 bans, 10 picks). Two games OpenDota lost (8891145641, 8959362208) were recovered from Liquipedia match records: the per-team draft order was reconstructed and the method validated by reproducing the sibling games' known sequences 24/24. For 8891145641 the result was additionally verified byte-for-byte against Valve's own record of the game's crashed first instance (the game died at 48 minutes and was resumed under a new match id, losing its draft metadata upstream); first-pick sides were confirmed from that record and from the official broadcast VOD respectively. Compilation and model outputs © batru.gg, released CC BY 4.0 — use it for anything, just credit batru.gg. Dota 2 is a trademark of Valve Corporation; this dataset is not affiliated with or endorsed by Valve.

Analysis only — not betting advice.

Quick start

import pandas as pd
games = pd.read_parquet("games.parquet")
steps = pd.read_parquet("draft_steps.parquet")

# Most contested heroes at TI 2026
ti = steps[steps.tournament == "ti-2026"]
print(ti.groupby("hero").size().sort_values(ascending=False).head(10))

# First-pick win rate per tournament
g = games.assign(fp_won=(games.first_pick_side == "radiant") == games.radiant_win)
print(g.groupby("tournament").fp_won.mean().round(3))

Live versions of this data

Auto-refreshing JSON: https://batru.gg/data/events/ti-2026.json (and siblings) — same numbers, updated continuously during events.
MCP server for AI assistants: https://batru.gg/mcp (get_dota_tournament_stats).

Downloads last month: 28