The tree-based architectures on the leaderboard all consume the same 441-dim
vector built by features.py. The sequence models do not - they
re-parse the raw session and build their own per-step embeddings - but the
cleaning rules below decide which sessions enter any of the
pipelines.
Each row of save.ligma describes one session. A session is a
21-step elimination tournament:
Stored fields per row: an id, a UNIX timestamp, a censored IP, age (years),
mood (0–60), gender, the 21-character pick string, the 21 cumulative
millisecond timestamps, and the colour arrays
[64 offered, 16 r1 winners, 4 r2 winners, final].
A row enters the dataset only if it passes all of these checks (see
is_valid() in features.py). The same rules are
duplicated in every sequence-model script so the train sets are identical:
"g" (boy) or "j" (girl)[6, 68][15s, 600s] — throws out abandoned sessions and afk usersAbout 1.2% of raw rows fail these checks. The rest become the 6,710-row dataset every model on the leaderboard is trained on.
Jump to a block:
Mean and standard deviation of R, G, B, H, S, L (12 numbers) across each of the three colour stages: the 64 offered colours, the 16 round-1 winners, and the 4 round-2 winners. Captures things like "this user gravitates toward dark saturated colours" without committing to which dimensions actually matter.
off_*, r1_*, r2_* with suffixes r_mean g_mean b_mean r_std g_std b_std h_mean s_mean l_mean h_std s_std l_std.
The colour the user ultimately chose, encoded redundantly so the trees can split on whichever colour space happens to be informative.
final_r, final_g, final_b (RGB, 0–1) · final_h, final_s, final_l (HSL) · final_y, final_u, final_v (YUV) · final_warmth = (R−B) / 255 · final_chroma = (max−min) / 255.
How much the round-1 mean colour differs from the offered-set mean, per
channel, and how much the round-2 mean differs from the round-1 mean. A
large positive sel_r1_dr means "this user pushed the
selection toward redder colours when given the chance".
sel_r1_dr, sel_r1_dg, sel_r1_db, sel_r2_dr, sel_r2_dg, sel_r2_db.
How many distinct 8×8×8 RGB voxels the round-1 and round-2 winners spread across, normalised by 16 and 4 respectively. Close to 1.0 means the user picked all over the colour space; close to 0.0 means they stayed in one neighbourhood.
voxel_div_r1, voxel_div_r2.
Mean pairwise Euclidean distance between the 16 round-1 winners in normalised RGB. A second flavour of "did this user stay close to one colour or wander?".
r1_internal_spread.
For every one of the 16 round-1 questions the model gets a full picture of
what was on screen and what the user picked. Per question (q00
through q15):
qNN_chosen_* — the same 8-number colour block as the
"Final colour" group (r, g, b, h, s, l, warmth, chroma) but for the
colour the user picked in question NN.
qNN_dr, qNN_dg, qNN_db — the channel-wise
delta from the chosen colour to the mean of the three rejected colours.
qNN_decisive — magnitude of that delta vector.
High = the chosen colour was very different from the rejects.
qNN_pos — which of the 4 corners (0–3) the
chosen colour was in.
Mean and standard deviation of the 16 per-question decisiveness magnitudes. A blunt summary of "how strong are this user's preferences?".
mean_decisiveness, std_decisiveness.
Same colour block as round-1 plus the corner position the user picked, for each of the 4 round-2 questions. No "delta from rejects" here because by round 2 every option was already approved in round 1, so the delta is less informative.
r2qN_chosen_* (r, g, b, h, s, l, warmth, chroma), r2qN_pos.
What fraction of the 16 round-1 picks landed in each of the four corner positions, plus the Shannon entropy of that distribution. Bored kids hit the same corner; engaged users spread out.
pos_0_frac, pos_1_frac, pos_2_frac, pos_3_frac, pos_entropy.
The 16 round-1 winners binned into a 4×4×4 RGB voxel grid (so each bin has 64-cube edges), then each bin divided by 16. Trees love this kind of dense low-resolution histogram for picking out "this user prefers the dark-blue corner of colour space".
hist_v00 through hist_v63.
Euclidean distance in normalised RGB from the final-pick colour to each of 12 hand-picked reference colours (pink, red, orange, yellow, green, cyan, blue, purple, brown, gray, black, white), plus the index of the closest one. The point is to give the model a head start at carving the colour space along axes that already mean something to humans, without forcing the model to discover "near pink" from raw RGB.
final_to_pink, final_to_red, ..., final_to_white, final_closest_ref.
Distance in normalised RGB from the final pick to the round-1 winner mean, and from the final pick to the round-2 winner mean. Did the user converge throughout the tournament or pick something off-trend at the end?
final_to_r1_mean, final_to_r2_mean.
Treats the 16 round-1 winners as a path through RGB space. Mean and std of step-to-step distance, mean distance to the centroid, and the slopes of warmth, lightness, and saturation across question index. Picks up drifts like "got tired and started clicking blue".
r1_consec_mean, r1_consec_std, r1_centroid_mean_dist, r1_warmth_slope, r1_light_slope, r1_sat_slope.
For each round-1 question, was the chosen colour the warmest / coolest / lightest / darkest / most-saturated / least-saturated of the 4 offered? Each feature is the fraction of the 16 questions for which the answer is yes.
extreme_warmest_frac, extreme_coolest_frac, extreme_lightest_frac, extreme_darkest_frac, extreme_most_sat_frac, extreme_least_sat_frac.
Per-question time deltas for the 16 round-1 questions binned into five buckets: under 1s, 1–3s, 3–7s, 7–15s, 15s+. Each value is a fraction summing to ~1. Picks up "skimmer vs reader" patterns.
r1_tbucket_0 through r1_tbucket_4.
Two hand-crafted interactions that the trees might construct on their own eventually but are useful to feed them directly:
time_x_low_entropy = mean question time × (1 −
position entropy / log 4). High = slow and concentrated picking,
the "deliberate corner-banger" signal.
time_per_decisive = mean question time / (mean
decisiveness + 50). How much time the user spends per unit of preference.
Aggregate statistics over the 21 per-question time deltas plus the round means and the linear slope across the whole session.
total_sec, mean_q_ms, std_q_ms, min_q_ms, max_q_ms, median_q_ms, first5_mean_ms, last5_mean_ms, time_slope, r1_mean_ms, r2_mean_ms, final_ms.
The raw time spent on each of the 21 questions, in seconds.
t_q00 through t_q20.
Time of day the session was recorded, encoded as sine and cosine of the
fractional hour in Oslo time (so midnight and noon are not adjacent in
feature space). The hour_known flag is 0 when the timestamp
parse failed, which lets the trees treat unknown-hour sessions separately.
hour_sin, hour_cos, hour_known.
Running python features.py writes three files into
training/ that every tree-based architecture loads:
features.npy — float32, shape [6710, 441]targets.npz — gender (int8), age (int16), mood (int16)feature_names.json — the 441 column names, in order