Angband expert and model play records

The recorded data behind the paper Angband as an Environment for LLM Agents: A Long Game in Short Episodes: the Borg's permadeath lives and backtracking victory, the stage and diagnostic episodes of three language models and the Borg from shared saved departures, the Borg's town branches from recovered model arrivals, the HTTP cache of every model response, and the saved starting states. The paths mirror the code repository: download this dataset into the repository root and every script, configuration and reproduction command finds its inputs in place.

Path	Size	Content
final_runs/backtrack_expert_80M_f2/	35 GB	the backtracking Borg, 8 streams of 10M visited steps: raw per-key traces of every branch (attempts/*.jsonl.gz), Parquet exports, branch metadata and results; the winning stream is seed 6001000
final_runs/borg_full_f2/	9.4 GB	50 Borg lives under permadeath, traces, metadata and Parquet exports
final_runs/stages30_v1/	1.5 GB	the thirty-departure study: stages/ (one episode per player and departure at the early, middle and end stage), clock/ (the Borg under clock-input budgets), early_long/ (the Borg for 5,000 keys from the early departures), town_pool/ and town_borg/ (47 recovered town arrivals and 470 Borg branches), analysis exports and the certification manifests in reproduce_compare/
final_runs/challenges9_f2/	46 MB	the diagnostic tasks, Borg and models
final_runs/stages9_f2/, stages_clock2_f2/, town_pool9_f2/, town_borg9_f2/	1.6 GB	the earlier four-departure stage, clock and town collections
final_runs/reproduce_compare_f2/	0.2 MB	the byte-level comparison of the full-corpus rerun
final_runs/llm_cache/llm_cache.sqlite	441 MB	every recorded model response, keyed by request; a rerun of any model collection replays from it without inference
snapshot_library/v2/	10 MB	base saves of the stages and the diagnostic tasks, with the paper's four departures per stage
snapshot_library/stages50_v1/	16 MB	fifty authored departures per stage (the study uses the first thirty), with sha256sums.txt
snapshot_library/winning_run/	101 MB	804 checkpoints along the winning path
DATA_MANIFEST.sha256		SHA-256 of every file in this dataset
DATA_LICENSE.md		licence terms

Model episodes carry the request log (*.requests.jsonl.gz, the messages as sent), the transcript (*.transcript.jsonl, replies, parsed keys and per-key execution feedback) and the game trace; Borg episodes carry the trace with the Borg's own annotations. Every run folder has a run_spec.json naming the engine, game data, start save, system prompt and source hashes it was collected with.

Reproducing the paper

With the code repository checked out and this dataset in place:

repro/remake_paper.sh          # every table, figure and number, then the PDF
TAG=_r repro/reproduce_all.sh  # collect everything again into <name>_r folders
TAG=_r repro/11_certify.sh     # compare the reproduction with this data

repro/REPRODUCE.md in the code repository gives the setup, the inputs that are not regenerated and the time each step takes. Model collections replay from the cache; the Borg collections are seeded and deterministic on the released engine, and two runs differ only in the wall-clock time printed on a death screen.

Integrity

shasum -a 256 -c DATA_MANIFEST.sha256

Collection logs are not included; the data are otherwise the folders as collected.

Licence

The project's contribution is licensed under Apache 2.0 (DATA_LICENSE.md). Upstream game text and other third-party material keep their own terms.