scripts/ — recompute every number in this repository
Standard library only; nothing to install. Run from the repository root.
Quick start
git clone https://huggingface.co/datasets/Gyubeum/AndroidFlux_Evaluation_Output
cd AndroidFlux_Evaluation_Output
python3 scripts/reproduce.py --check # recount every cell and compare with its header
python3 scripts/score_tables.py --root . --all # regenerate the paper tables
reproduce.py — a small, readable reproducer
It answers one question: does the number in a cell's stored header come out of the per-episode results?
python3 scripts/reproduce.py # recompute every cell
python3 scripts/reproduce.py --check # ... and compare with the headers (exit 1 on any mismatch)
python3 scripts/reproduce.py --csv out.csv # one line per cell
python3 scripts/reproduce.py --cell human_verified/replay_65/t/ma3_7b
Example output:
human_verified/replay_65/t/ma3_7b 23/65 35.38% valid=True issues=0
Cells whose denominator differs from the expected one, and cells that have no per-episode rows and therefore fall back to the header, are flagged in place. The header is never trusted silently.
score_tables.py — regenerate the paper tables
Holds the model registry, the AndroidWorld axis, and all diagnostics.
python3 scripts/score_tables.py --root . --all # all four tables
python3 scripts/score_tables.py --root . --verify # recompute every cell; exit != 0 on a fatal defect
python3 scripts/score_tables.py --root . --table tab:hvlvcomparison
python3 scripts/score_tables.py --root . --cell human_verified/replay_65/t_plus_1_pos/ma3_32b --explain
python3 scripts/score_tables.py --root . --aw-root <AW tree> --table aw --show-counts
--verify does not check the AndroidWorld axis. It only recomputes the axes that are
re-scored per episode (human-verified and LLM-verified). Running --verify with only an
AndroidWorld tree prints 0 cell(s) recomputed, 0 fatal and exits 0, which means
not checked, not passed. For AndroidWorld, confirm that --table aw --show-counts
produces values.
scratch_80_tasks.txt · scratch_65_tasks.txt · snapshot_15_tasks.txt
scratch_80 is the union of the 65 tasks of replay_65 and the 15 tasks of
snapshot_15. The two sets have an empty intersection and the union is exactly 80.
The list was derived independently in three places and matched byte for byte.
Scoring rules — the two families differ
validation.json · per_episode.json success = score > 0 (0.5 counts as success)
sr_summary.json (AndroidWorld) success = score >= 1 (0.5 counts as failure)
read from ["total"]["successes"]
(not the top level, not score_points)
The fields of an episode record differ between cells. The fallback order is
is_successful → success → score, and is_successful may be null (the task died
with an exception before it was judged). null is a failure, not an exclusion: do not
shrink the denominator.
The difference matters on AndroidWorld. One tree's per-task score distribution was
0.0 → 4939 · 1.0 → 4730 · null → 155 · 0.5 → 36; counting with > 0 would flip those
36 to successes.
Denominators are fixed
replay_65 clean · t_minus_5 · t_minus_3 · t_minus_1 · t 65
t_plus_1_pos · t_plus_1_neg 52 each
snapshot_15 clean · t_minus_1 · t 15
replay_312 clean · t 312 (104 tasks × 3 source trajectories)
androidworld 116 (including the ~6 tasks that die with an exception)
t_minus_3 and t_minus_5 do not drop episodes whose prefix is too short. They are
clamped and stay in the denominator of 65: t−3 has 39 feasible + 26 clamped, t−5 has 21 + 44.
A run that loaded fewer tasks does not get a smaller denominator. One AndroidWorld run
loaded 115 of 116 (episodes=115 != expected_tasks=116); the missing task counts as a
failure. Do not pad per_task.json to 116 rows.