You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

bcenv-replays-v1

Every ranked ladder game of Cambridge Battlecode 2026a (event cambc2026a) on engine 1.0.2 that we downloaded: 8850 replays (281 MB) from 34 teams, created between 2026-10-21T05:20 and 2026-10-22T05:00 UTC. It is the data the Intercontinental Fish rule book and the fishlike bot were reverse-engineered from. It is packaged so any teammate, or their agent, can do the same for another team.

Getting it in place

The layout mirrors battlecode26a-rl-env/data/delta/, which is where the reverse-engineering tools read it:

cd battlecode26a-rl-env
hf download derekz5/bcenv-replays-v1 --repo-type dataset --local-dir data/delta      # into an empty data/delta/

This gives data/delta/matches.json and data/delta/replays/<matchId>.repl26a. The maps come from the official pool, so also run cambc maps pull in the arena project the tools point at (ARENA in the tools).

Starting on a team

  1. Read the work order battlecode26a-rl-env/reverse_engineering/README.md and the finished example, reverse_engineering/reports/fish.md. Then read reverse_engineering/reports/diagnostics.md: which diagnostics helped, and the traps.
  2. Pick one bot version. Find the team in teams.csv, and take its games from one window, normally the latest. Mixing versions mixes policies. Use versions-openings.json where the team has one. Otherwise use versions.json, and check the openings yourself: draft-based windows miss bot changes that keep the draft (Fish's 11:10 change is one).
  3. Walk its replays once into per-turn records, then mine and test: python reverse_engineering/tools/walk_replays.py --team "<name>" --since <window start> --engine 1.0.2. Each record holds what the team's bot saw (the API view) and what it did.

Files

  • matches.json: one entry per game, in the downloader's format:
    • matchId, createdAt, map, kind (ranked), engine, endReason, minRating, bytes;
    • teams: for each side, teamId, name, side (a / b), result, draftPicks, rank and rating at the time.
  • replays/<matchId>.repl26a: the replay bundle exactly as downloaded (a zip). Open it with bcenv.open_replay, which picks the engine build that recorded it (an engine only loads replays it recorded). bcenv.resim replays it turn by turn in a live engine.
  • teams.csv: one row per team over these games:
    • games, wins, win_rate;
    • best_rank (the best ladder rank seen in its games), latest_rating;
    • first_game, last_game;
    • versions and the latest window's start, games and win rate.
  • versions.json: per team (with 20+ games), its bot-version windows: from, to, games, win_rate, most_common_draft. A new window starts in an hour where the team's draft changed on 3+ maps. A bot change that keeps the draft is invisible here.
  • versions-openings.json: finer windows from changes in opening play, for five teams (Intercontinental Fish, Univardisty of Shingtonwash, prism best girl, policy wonk, procless). Fish's last version starts at 2026-10-21T11:10.
  • lists/re-intercontinental-fish.txt: the 535 Fish games behind its rule book: its last version, created at or after 2026-10-21T11:10. The tools hold out games with sha1(matchId) % 5 == 0 for validation.

Top teams in this set

Team Games Win rate Best rank Versions Latest version from
Intercontinental Fish 705 0.601 1 1 2026-10-21T05:20 (705 games)
prism best girl 180 0.778 1 1 2026-10-21T23:00 (180 games)
procless 705 0.591 2 2 2026-10-21T07:00 (655 games)
EMPEROR Tung 700 0.514 2 2 2026-10-21T09:00 (595 games)
:3 700 0.519 4 2 2026-10-21T15:00 (420 games)
Sponge(Albert and Bob) 705 0.511 5 3 2026-10-22T00:00 (150 games)
Univardisty of Shingtonwash 325 0.594 5 2 2026-10-21T18:10 (315 games)
policy wonk 695 0.524 6 2 2026-10-21T06:00 (680 games)
Pantheon 710 0.479 7 1 2026-10-21T05:20 (710 games)
Nanyang Technological University 700 0.493 8 3 2026-10-21T17:00 (365 games)
Zarzorki 700 0.501 9 4 2026-10-21T12:00 (510 games)
hello 700 0.481 9 2 2026-10-21T13:00 (480 games)
Trzy Kwaterki 710 0.48 10 3 2026-10-21T11:00 (545 games)
Let's LARP again 685 0.502 11 4 2026-10-21T21:00 (240 games)
NEETCoders 710 0.496 12 3 2026-10-21T09:00 (600 games)
new team(4) 710 0.487 12 3 2026-10-21T23:00 (185 games)
Sneaky Bush Campers 690 0.477 14 4 2026-10-21T17:00 (365 games)
FürFortnite 695 0.462 15 2 2026-10-21T16:00 (390 games)
conqueror_of_openai 605 0.511 15 4 2026-10-22T00:00 (155 games)
Tartarus 670 0.49 16 6 2026-10-22T05:00 (5 games)

Things the replays do not contain

  • Team shared memory: the 8 bytes teammates share are recorded as zeros. Coordination has to be inferred from behaviour, for example targets shared between teammates.
  • Cast slots, purchases and priority values are not stored directly. The walker recovers them from the state around each step.
  • Games after 2026-10-22T05:00 UTC. Newer games need another download (bcenv-ladder, with a logged-in cambc).

Where they come from

  • Download: the official cambc CLI with a logged-in session, through bcenv-ladder (battlecode26a-rl-env), on the laptop and on NCSA Delta.
  • Merge: Delta's data directory merged both downloads: 15,930 games on engines 1.0.0–1.0.2, of which these are the 8850 on 1.0.2.
  • Converted rows: the critic dataset derekz5/bcenv-critic-v1 holds rows converted from the same games.

Private: these are other teams' games, downloaded through a logged-in ladder session.

Downloads last month
27