HF_TOKEN secret, so results are not saved."
else:
note = "Unranked match: Elo not changed."
note += f" Ranking: score, then lines, then pieces survived. ✓ = still alive at the {MAX_PIECES}-piece cap."
results = results_html(order, ranks, players, games, elos, note)
yield _status(f"Match finished · seed {seed} · {protocol} · {mode}.", "ok"), arena_html(games, players, ranks, elos), results
def stop_status(current):
# only claim a stop when a match was actually running
if (current or "").startswith("⏳"):
return _status("Match stopped. Nothing was recorded.", "warn")
return current
def leaderboard_view(protocol):
protocol = protocol or "guided"
entries = [e for e in STORE.rows(protocol) if e["model"] not in BASELINES]
return leaderboard_html(entries, protocol, MAX_PARAM_GAP, short_params(MAX_PARAM_GAP))
def refresh_leaderboard(protocol):
STORE.reload()
return leaderboard_view(protocol)
INTRO = f"""
SMALL MODELS. ZERO-SHOT TETRIS.
Decoder-only language models (≤ {short_params(MAX_PARAMS)} parameters, custom architectures welcome) play Tetris zero-shot: no fine-tuning, no game data, only what they learned from pre-training on text. Every player gets the same piece sequence, and ranked matches update a public Elo leaderboard.
HF_TOKEN secret with write access '
'to the results dataset.