gemDataFrames
Converts a ParsedMatch into a dictionary of pandas DataFrames, one per table.
This is the primary interface for ML/data-science workflows.
Available tables
Every table starts with a match_id column (0 when the replay has no match ID).
Tables hold only primitive cells, and every column has a fixed pandas nullable
dtype (Int64, Float64, boolean, or string). Missing values are pd.NA.
A table has the same columns and Parquet schema in every match, even when the
match has no rows for it, so per-replay files concatenate cleanly.
Core tables (always returned)
| Key | Contents |
|---|---|
"match" | Single-row match metadata (id, mode, winner, ticks, final status bitmasks) |
"player_summary" | One row per player: identity, K/D/A, final net worth/LH/DN, GPM/XPM, damage/healing totals, lane stats, damage-type split, largest hero hit, consumed Aghanim's/Moon Shard flags |
"player_timeseries" | Per-player sampled gold, total_earned_gold, total_earned_xp, net_worth, lh, dn, xp, hp, max_hp, mana, max_mana by tick |
"players_minute" | Per-player series resampled to game minutes |
"player_breakdowns" | Long-form per-player dict stats as (player_id, stat, key, subkey, value) |
"positions" | Per-player sampled world (x, y) positions |
"radiant_advantage" | Per-minute gold and XP advantage curves |
"combat_log" | Every damage, kill, heal, and modifier event |
"wards" | Ward placements with exact map coordinates |
"objectives" | Tower, barracks, Roshan, Tormentor, shrine, Aegis, courier, and banner rows |
"chat" | All-chat and team-chat messages |
"draft" | Pick and ban events in order |
"fights" | Detected fight windows, one row per fight with a fight_index |
"fight_players" | Per-fight, per-player deaths, buybacks, damage, healing, gold/XP delta |
"smoke_events" | Smoke activations, ";"-joined smoked heroes, and centroids |
"smoke_members" | Flat per-hero smoke application/removal ticks, durations, and sampled positions |
"courier_snapshots" | Courier state samples |
"neutral_item_finds" | Neutral item finds with item and enhancement IDs/keys |
"hero_visibility_events" | Change-only Radiant/Dire visibility state for canonical player heroes |
"entity_visibility" | Change-only packet-boundary visibility and lifecycle for networked Dota NPC entities |
"vision_modifiers" | Flat vision-relevant modifier application lifecycles and evidence |
"vision_modifier_pairing_issues" | Flat ambiguous/orphan modifier removal evidence |
"player_kills_log", "player_purchase_log", "player_runes_log", "player_buyback_log" | Per-player combat-log projections |
Optional groups
Pass include=["analysis"] and/or include=["opendota"] to add these tables.
| Group | Key | Contents |
|---|---|---|
analysis | "fight_positioning" | Flat per-fight, per-snapshot, per-hero positioning evidence with freshness, geometry, and visibility |
analysis | "roshan_conversions" | One flat row per Roshan with attribution/lifecycle provenance, raw differential evidence and availability |
analysis | "roshan_conversion_fights" | One row per Roshan-associated fight with engagement provenance, relation, winner, and participant IDs by side |
analysis | "farming_routes" | One row per player with route availability, camp-catalog metadata, and segment/point counts |
analysis | "farming_route_segments" | Camp-local route windows with boundary provenance, sampled coverage, neutral/resource support, evidence strength, and gaps |
analysis | "farming_route_points" | Sampled path points with deterministic camp membership and optional segment membership |
opendota | "opendota_objectives" | OpenDota-shaped unified objective timeline |
opendota | "opendota_teamfights" | OpenDota-compatible 3+ death temporal teamfight windows; the per-player breakdown is a JSON string in players |
Generated API
Module gem.results.dataframes
DataFrame conversion for :class:ParsedMatch output.
Source: src/gem/results/dataframes.py
Top-level functions
build_dataframes
def build_dataframes(match: ParsedMatch, *, include: Iterable[str] = ()) -> dict[str, pd.DataFrame]Convert a :class:ParsedMatch into a dict of flat pandas DataFrames.