Datasets:

Siewart
/

sage

@@ -1,7 +1,7 @@
 ---
 language:
 - en
-license: cc0-1.0
 task_categories:
 - other
 tags:
@@ -83,9 +83,8 @@ There are some notes regarding the data that users should be aware of when makin
 - The official AoE2 API does not retain all recorded games indefinitely; only games from the active patch at the time of collection were sampled, and not all were still available on the server.
 - Since the SAGE project was long-running, it was not fully updated throughout. Newer game mechanics may not be properly captured in the `.caderec`s and are excluded from the Arrow and Parquet exports.
 - Some recordings failed to parse for a variety of reasons. An effort was made to maximise coverage, but not all files could be recovered. The estimated `.aoe2record` → `.caderec` → Arrow success rate is roughly 80–90%. Incompatibilities may be feature-consistent and could introduce bias depending on the study.
-- All games were validated to have a declared winner and to run for at least 10 minutes; shorter games were excluded from selection outright rather than merely flagged.
   - For KotD4, this includes "restores" where a game was unexpectedly terminated and then restored from the last backup checkpoint.
-- Each match was additionally checked for player activity: every player must issue at least one human command in every full minute of the game (the final, possibly partial, minute is excluded from this check). Matches with idle stretches are not dropped outright, since some flagged cases are legitimate competitive behaviour (e.g. a player deliberately stalling in protest of an opponent's conduct) rather than a data defect; these are instead retained with a `quality_flags` field on the affected game's `index.json` entry so downstream users can filter or weight them as needed.
 - If the metadata in the `.caderec`s (which contains: players, winner, map, civ IDs, and game length) is insufficient, the Arrow or Parquet data alongside `descriptor.json` can help validate suitability for a specific goal.
 - Despite being a newer feature of CaptureAge, the KotD4 `.caderec` files are present. They lack some CA:DE features due to being recorded earlier, notably fast-forward and some statistics.
 - **Filenames (may2026):** Each file is named `PlayerName1_(Civilization1)_Rating1_vs_PlayerName2_(Civilization2)_Rating2_UniqueGameId`. The player and game ID can be used to query the official API to expand and validate metadata.
@@ -112,19 +111,6 @@ Ranked 1v1 Random Map ladder games collected in May 2026.
 ---
-## Anonymization
-The `may2026` dataset is anonymized. `kotd4` (King of the Desert 4, a public tournament whose participants are already publicly known) is not.
-- **Player names** are replaced with a stable pseudonym of the form `Player<6-char base62 hash>`, derived via PBKDF2-HMAC-SHA256 over the player's official Relic `profileId` and a fixed, private per-dataset salt (not published). The same real player always receives the same pseudonym across every match they appear in, so cross-match player behaviour can still be studied without exposing identity. Players whose `profileId` could not be resolved (e.g. unranked slots, or matches sourced outside the official match-history API) fall back to a generic `Player1` / `Player2` label instead of a hash. This fallback label carries no identity information and is **not** unique across games, do not treat it as a stable pseudonym.
-- **Relic profile IDs** are zeroed everywhere they appear: `.aoe2record` header, `.caderec` metadata, Arrow/Parquet `profileId` columns, and the `index.json` / JSONL `players[].profileId` field.
-- **Match/game UUIDs** are zeroed in the `.caderec` metadata and the `gameId` field.
-- **ELO ratings** are floored to reduce re-identification risk from precise values: ratings ≥2500 floor to 2500; 2200–2499 floor to the nearest 100; below 2200 floor to the nearest 50.
-- **In-game chat** (age-up announcements, lobby chat, and player-to-player messages) is replaced with a generic placeholder describing the event type, rather than stripped outright, since chat cadence and timing can themselves be a useful signal.
-- Every processed match was validated with an automated pass confirming no original player name, un-zeroed profile ID, or un-zeroed UUID leaks into any output file (`.aoe2record`, `.caderec`, Arrow, Parquet, or the JSONL/`index.json` metadata), and that the pseudonym used is consistent across all file types for the same match.
----
 ## Legal and Attribution
 ### Microsoft Game Content
@@ -141,7 +127,7 @@ All files were processed with written permission from CaptureAge and verbal perm
 ### License
-The non-string Arrow, Parquet, and `.caderec` components of this dataset are released under [Creative Commons Zero](https://creativecommons.org/publicdomain/zero/1.0/). The game string assets (`strings.arrow`, `strings.parquet`) and the original `.aoe2record` files are subject to their respective terms described above.
 ### Citation
@@ -156,7 +142,7 @@ If you use this dataset, please cite the accompanying thesis:
   address  = {Enschede},
   type     = {Thesis},
   title    = {One Thousand and One Tales: Automated Age of Empires {II} eSports Reports from Raw Real-time Game Data},
-  url      = {https://purl.utwente.nl/essays/109934},
 }
 @dataset{VanWingerden2026sage,
@@ -169,7 +155,7 @@ If you use this dataset, please cite the accompanying thesis:
 ```
 **APA:**
-van Wingerden, S. (2026). *One Thousand and One Tales: Automated Age of Empires II eSports Reports from Raw Real-time Game Data* (Master's thesis, University of Twente, Enschede). https://purl.utwente.nl/essays/109934
 ---

 ---
 language:
 - en
+license: apache-2.0
 task_categories:
 - other
 tags:
 - The official AoE2 API does not retain all recorded games indefinitely; only games from the active patch at the time of collection were sampled, and not all were still available on the server.
 - Since the SAGE project was long-running, it was not fully updated throughout. Newer game mechanics may not be properly captured in the `.caderec`s and are excluded from the Arrow and Parquet exports.
 - Some recordings failed to parse for a variety of reasons. An effort was made to maximise coverage, but not all files could be recovered. The estimated `.aoe2record` → `.caderec` → Arrow success rate is roughly 80–90%. Incompatibilities may be feature-consistent and could introduce bias depending on the study.
+- All games were validated to have a declared winner at the end. Very short games where a winner was declared early were kept intentionally, but may warrant scrutiny depending on the goals of a specific study. These games are not included in the train, eval or test sets.
   - For KotD4, this includes "restores" where a game was unexpectedly terminated and then restored from the last backup checkpoint.
 - If the metadata in the `.caderec`s (which contains: players, winner, map, civ IDs, and game length) is insufficient, the Arrow or Parquet data alongside `descriptor.json` can help validate suitability for a specific goal.
 - Despite being a newer feature of CaptureAge, the KotD4 `.caderec` files are present. They lack some CA:DE features due to being recorded earlier, notably fast-forward and some statistics.
 - **Filenames (may2026):** Each file is named `PlayerName1_(Civilization1)_Rating1_vs_PlayerName2_(Civilization2)_Rating2_UniqueGameId`. The player and game ID can be used to query the official API to expand and validate metadata.
 ---
 ## Legal and Attribution
 ### Microsoft Game Content
 ### License
+The non-string Arrow, Parquet, and `.caderec` components of this dataset are released under the [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0). The game string assets (`strings.arrow`, `strings.parquet`) and the original `.aoe2record` files are subject to their respective terms described above.
 ### Citation
   address  = {Enschede},
   type     = {Thesis},
   title    = {One Thousand and One Tales: Automated Age of Empires {II} eSports Reports from Raw Real-time Game Data},
+  url      = {https://purl.utwente.nl/essays/<TODO>},
 }
 @dataset{VanWingerden2026sage,
 ```
 **APA:**
+van Wingerden, S. (2026). *One Thousand and One Tales: Automated Age of Empires II eSports Reports from Raw Real-time Game Data* (Master's thesis, University of Twente, Enschede). https://purl.utwente.nl/essays/<TODO>
 ---

data-files/README.md DELETED Viewed

@@ -1,131 +0,0 @@
-# AoE2:DE reference data extraction — README
-Extracted directly from the game's own `empires2_x2_p1.dat` (Genie engine
-format) via a custom C++ tool built against the official `Tapsa/genieutils`. Two
-snapshots exist:
-- `may2026/` `VER 8.8`, 59 civs, 1510 techs, 1409 effects.
-- `kotd4/` — matched to the KOTD4 tournament
-  dataset, `VER 7.4`, 39 civs, 851 techs, 851 effects.
-The contents of these files are released under the Mi
-Each snapshot's display strings (`display_name`, `description`, etc.) are
-resolved against **that same install's** `key-value-strings-utf8.txt`,
-## Files
-- `aoe2_civs.json` — one entry per civilization.
-- `aoe2_techs.json` — one entry per tech (includes hidden/auto-research techs
-  used internally for civ bonuses and tech-tree gating — see `is_civ_bonus`).
-- `aoe2_units.json` — one entry per **skirmish-relevant** unit: filtered to
-  unit ids that appear in the game's own tech-tree UI (`TechTree
-  .UnitConnections` / `BuildingConnections`). This excludes scenario-editor
-  entities — corpse variants (`_D` suffix), duplicate age-specific
-  building meshes (`Shipyard2/3/4`, `Port 2/3/4`), decorative props, and
-  unused duplicate entries — that share id ranges with real units but were
-  never wired into the tech tree. So this excludes gold mines, trees, rocks,
-  and other gaia props, as well as the "dummy" units.
-- `aoe2_effects.json` — every effect group (id + name + raw command list),
-  for looking up an `effect_id` referenced from anywhere else in these files.
-IDs (`civ id`, `tech id`, `unit id`) match the raw `.dat` values and were
-cross-checked against the SAGE/KOTD4 per-match parquet export pipeline's own
-id space — they align.
-## Text fields
-- `internal_name` — old internal codename (e.g. `"British"`), not the
-  in-game display name.
-- `display_name` — resolved in-game name. For civs this isn't stored in the
-  dat file at all; it's a hardcoded engine convention (`string id = 10270 +
-  civ id`).
-- Units: `creation_text` (train-button label), `help_text`, and
-  `description` — the official train-button tooltip ("Strong vs. X, Weak
-  vs. Y" blurb) which lives at an string offset
-  (`21000`).
-- Techs: `description`, `help_text`, `tech_tree_text` (tech-tree panel text).
-- Civs: `overview_text` — the civ-selection-screen blurb (bonuses, unique
-  unit/techs, team bonus), resolved from another undocumented convention
-  (`string id = 120149 + civ id`). `team_bonus_text` is the "Team Bonus:"
-  section parsed out of that same blurb.
-## Effect commands (`type`/`a`/`b`/`c`/`d`)
-Every place an effect's command list appears, each command has been
-enriched in place with `type_name` and `meaning` (a decoded, human-readable
-one-liner) using the field semantics from `Tapsa/AGE` (the actual AoE2
-modding editor's source — the tool players use to hand-edit these exact
-values, so it's authoritative, not inferred).
-The type number encodes both a base operation and a scope:
-`type = base + scope_offset`, where `scope_offset` is `0` (self), `10`
-(team), `20` (enemy), `30` (neutral), or `40` (gaia), and `base` is one of:
-`0` set attribute, `1` set/change resource, `2` enable/disable unit,
-`3` upgrade unit, `4` change attribute (+/-), `5` multiply attribute,
-`6` multiply resource, `7` spawn unit, `8` modify tech, `9` set civ name.
-Types `101`/`102`/`103` (tech cost/disable/research-time modifiers) don't
-get scope variants.
-**Known gap:** the `a`/`b`/`c`/`d` → class-name lookup for attack/armor
-bonuses (`class N` in `meaning`) is only populated through class id 61; AoE2
-DE has since added more unit classes for newer content, so some decoded
-lines show a wrong or generic class name (e.g. "vs Tree" where the real
-target is a newer class not in the table) or fall back to `unit N` for a
-few unresolved ids in some team bonuses. Treat `meaning` as a best-effort
-gloss, not a certified translation — the raw `type`/`a`/`b`/`c`/`d` values
-are always present alongside it if you need to re-derive.
-## Civ access / tech-tree fields
-- `civs_with_access` / `civs_disabled` (on units) and `disabled_techs` (on
-  civs) / `disabled_for_civs` (on techs) are two ends of the same data,
-  built from **two independent sources**:
-  1. civ-restricted bonus techs (`Tech.Civ >= 0`) and their effects, and
-  2. each civ's own dedicated "Tech Tree" effect
-     (`Civ.TechTreeID` → `Effects`), which is **not** reachable through any
-     `Tech` row at all — it's a direct field on the civ record. This is
-     where most `type=102` "Disable Tech" commands live (e.g. Britons'
-     tech-tree effect disables `Bombard Tower`, `Stone Shaft Mining`, etc.
-     — 51 techs in the current snapshot). An earlier version of this
-     extraction missed source (2) entirely, which is also why
-     `civs_with_access`/`unique_to_civ` came back empty far more often than
-     they should have.
-- **An empty `civs_with_access` list does NOT mean "no civ has this unit."**
-  Most units have no explicit per-civ restriction at all and are available
-  to every civ by default (e.g. Archer, Militia-line) — for those, empty is
-  correct and expected. The list is only populated when a specific civ-only
-  *unlock* command was found; treat it as "explicitly restricted to these
-  civs," not as the full roster.
-- `unique_to_civ` is derived from `civs_with_access` having exactly one
-  entry — same caveat applies (a unit with no restriction at all is neither
-  "unique" nor generically "shared," it's simply unrestricted).
-These choices a reflection of the game's internal logic, not a design decision`
-of this extraction tool.
-## Team bonus (fixed)
-`team_bonus.effect_id` (`Civ.TeamBonusID`) is a **direct index into
-`Effects`**, not a tech id. An earlier version of this tool wrongly treated
-it as a tech id and looked up `Techs[TeamBonusID].EffectID`, which for
-Britons (`TeamBonusID=399`) silently landed on an unrelated tech and
-decoded as something nonsensical (Vikings' unique unit). This is now fixed
-and cross-checked against `team_bonus_text` (the string-based version) —
-they agree for all civs in both snapshots.
-## `repeatable` on techs
-`repeatable=true` does **not** mean "can be researched
-multiple times in a normal game" — the great majority of techs in the dat
-(1154 of 1510 in the current snapshot, including ordinary one-shot techs
-like Loom and Crop Rotation) carry this flag. It only takes effect under
-specific game/lobby settings that permit re-research; in a standard game
-almost everything behaves as single-research regardless of this flag.
-Don't treat it as a reliable "can be researched N times" signal on its own.
-## AI Use
-This document was generated using AI, then manually reviewed and edited for accuracy.
-The AI was used to summarize the contents of the files and provide explanations for
-AI was also used to extract the information from the game's data files and organize
-it into a readable format, following the author's direction.

data-files/kotd4/aoe2_civs.json DELETED Viewed