Season 5 pool: - Supra-Mini-v5-8M, Supra1.5-50M-Base-exp, SupraGDN-5M, LowOnMind-1M, BananaMind-2.1-Unified, BananaMind-2-Mini, DynamicMind-MoE; + SmolLM-135M, SmolLM2-135M, Rose-1.5-Medium, Rose-Mini, Boris-1.3-75M

app.py

@@ -1 +1,384 @@
-{{FILE:/home/claude/space/app.py}}
+"""SLM Tetris Arena — decoder-only LMs play Tetris zero-shot and earn Elo."""
+import os
+
+# Keep the write token out of the environment before any model code runs:
+# only the results store receives it (custom model code runs in this process).
+_TOKEN = os.environ.pop("HF_TOKEN", None) or os.environ.pop("HUGGING_FACE_HUB_TOKEN", None)
+os.environ.setdefault("HF_HUB_DISABLE_PROGRESS_BARS", "1")
+
+import html
+import inspect
+import random
+import time
+
+import gradio as gr
+import torch
+
+from arena import MAX_PIECES, SEASON, ResultsStore, choose, rank_games
+from leaderboard import LB_CSS, leaderboard_html
+from leaderboard import fmt_params as short_params
+from players import (BASELINES, MAX_PARAMS, MIN_PARAMS, ORACLE_ID, PROMPTS, RANDOM_ID, ModelRejected, OracleReaderPlayer,
+                     RandomPlayer, fmt_params, load_player, precheck)
+from ranked import Pool, RankedRunner
+from render import CSS, arena_html, empty_html, results_html
+from tetris import TetrisGame
+
+# cpu-basic Spaces have 2 vCPUs; os.cpu_count() reports the host, which oversubscribes threads
+torch.set_num_threads(int(os.environ.get("TORCH_THREADS", 2)))
+
+RESULTS_REPO = os.environ.get("RESULTS_REPO", "DedeProGames/lm-tetris-arena-results")
+MAX_MODELS = int(os.environ.get("MAX_MODELS", 4))
+# Optional size limit for the random ranked pick (0 = none). Without it every model can meet every other, so all
+# ratings sit on one comparable scale; Elo already weighs each win by the opponent's rating, so beating a much
+# weaker model earns almost nothing once ratings have settled.
+MAX_PARAM_GAP = int(os.environ.get("MAX_PARAM_GAP", 0))
+# ...with a gap, models this size or bigger can still all play each other
+LARGE_FROM = int(os.environ.get("LARGE_FROM", 100_000_000))
+if MAX_PARAM_GAP:
+    SIZE_RULE = (f"all within ±{short_params(MAX_PARAM_GAP)} parameters of each other "
+                 f"(models with {short_params(LARGE_FROM)}+ parameters can all play each other)")
+else:
+    SIZE_RULE = "of any size (Elo weighs every win by the opponent's rating, so all models share one scale)"
+
+SUGGESTED = [
+    'AxiomicLabs/GPT-S-1.4M',
+    'AxiomicLabs/GPT-S2-5M',
+    'AxiomicLabs/GPT-X2.5-135M',
+    'BananaMind/BananaMind-2-Medium',
+    'BananaMind/BananaMind-2-Nano',
+    'BananaMind/BananaMind-2-Pro',
+    'BananaMind/BananaMind-2.1-Pico-Preview',
+    'DedeBckp/BackKiyo-10M',
+    'DedeProGames/DynamicMind-Mini',
+    'DedeProGames/GPT-U-20M',
+    'DedeProGames/Kiyo-230M-Preview',
+    'DedeProGames/Kiyo-65M',
+    'DedeProGames/LowOnMind-5M',
+    'DedeProGames/Overaddicted-500K',
+    'DedeProGames/Wisp-15M',
+    'DedeProGames/Wisp-5M',
+    'GODELEV/Rose-1.5-Medium',
+    'GODELEV/Rose-Mini',
+    'HuggingFaceTB/SmolLM-135M',
+    'HuggingFaceTB/SmolLM2-135M',
+    'openai-community/gpt2',
+    'opencerebral/Boris-1.3-75M',
+    'SupraLabs/Supra2-100M-Base',
+    'SupraLabs/Supra2-Medium-Base',
+    'SupraLabs/SupraNeo-4M',
+    'veyra-ai/Veyra2-Apricot-50M-Base',
+    'veyra-ai/Veyra2-Blueberry-5M-Base',
+    'veyra-ai/Veyra2-Mango-30M-Base',
+]
+DEFAULT_MODELS = ["DedeProGames/Kiyo-65M", "BananaMind/BananaMind-2-Medium", "HuggingFaceTB/SmolLM2-135M", "SupraLabs/Supra2-Medium-Base"]
+# The leaderboard only keeps pool models: removing a model from SUGGESTED also removes it from the leaderboard.
+STORE = ResultsStore(RESULTS_REPO, _TOKEN, min_params=MIN_PARAMS, allowed=SUGGESTED)
+# Ranked: the arena picks the players at random from the suggested models (nobody chooses who plays)
+POOL = Pool(SUGGESTED, MAX_PARAM_GAP, MAX_MODELS, LARGE_FROM)
+RANKED = RankedRunner(POOL, STORE)
+PROTOCOL_CHOICES = [("Guided: the rules are in the prompt", "guided"), ("Blind: no rules, only pre-training knowledge", "blind")]
+BASELINE_CHOICES = [(label, key) for key, label in BASELINES.items()]
+
+# ---- Look & feel: BananaMind SLM Leaderboard palette and type (dark + light), light-blue accent ----
+_D = dict(bg="#0b0e0d", surface="#111613", surface2="#191e1b", text="#f0f1ec", muted="#929b93", line="#29312c", accent="#4d9fff")
+_L = dict(bg="#f4f5f1", surface="#ffffff", surface2="#edf0e9", text="#17221b", muted="#626e64", line="#d7ded5", accent="#1b64c4")
+_CHECK = ("url(\"data:image/svg+xml,%3csvg viewBox='0 0 16 16' fill='%2307182e' xmlns='http://www.w3.org/2000/svg'%3e"
+          "%3cpath d='M12.207 4.793a1 1 0 010 1.414l-5 5a1 1 0 01-1.414 0l-2-2a1 1 0 011.414-1.414L6.5 9.086l4.293-4.293a1 1 0 011.414 0z'/%3e%3c/svg%3e\")")
+
+
+_THEME_KEYS = set(inspect.signature(gr.themes.Base.set).parameters)
+
+
+def _both(**pairs):
+    """name=(light, dark) -> theme kwargs for both modes (skips variables this Gradio version lacks)."""
+    out = {}
+    for k, (light, dark) in pairs.items():
+        if k in _THEME_KEYS:
+            out[k] = light
+        if k + "_dark" in _THEME_KEYS:
+            out[k + "_dark"] = dark
+    return out
+
+
+THEME = gr.themes.Base(
+    primary_hue=gr.themes.colors.blue, secondary_hue=gr.themes.colors.blue, neutral_hue=gr.themes.colors.stone,
+    font=[gr.themes.GoogleFont("DM Sans"), "Arial", "sans-serif"], font_mono=["ui-monospace", "SFMono-Regular", "monospace"],
+).set(
+    block_border_width="1px", block_radius="13px", block_label_border_width="0px", block_title_text_weight="500",
+    input_radius="8px", checkbox_check=_CHECK,
+    **_both(
+        body_background_fill=(_L["bg"], _D["bg"]), body_text_color=(_L["text"], _D["text"]),
+        body_text_color_subdued=(_L["muted"], _D["muted"]),
+        background_fill_primary=(_L["surface"], _D["surface"]), background_fill_secondary=(_L["surface2"], _D["surface2"]),
+        block_background_fill=(_L["surface"], _D["surface"]), block_border_color=(_L["line"], _D["line"]),
+        block_shadow=("none", "none"), block_label_background_fill=("transparent", "transparent"),
+        block_label_text_color=(_L["muted"], _D["muted"]), block_label_shadow=("none", "none"),
+        block_title_background_fill=("transparent", "transparent"), block_title_text_color=(_L["muted"], _D["muted"]),
+        block_info_text_color=(_L["muted"], _D["muted"]),
+        border_color_primary=(_L["line"], _D["line"]), border_color_accent=("#4d9fff", "#4d9fff"),
+        color_accent=("#4d9fff", "#4d9fff"), color_accent_soft=("#4d9fff26", "#4d9fff26"),
+        input_background_fill=(_L["surface2"], _D["surface2"]), input_border_color=(_L["line"], _D["line"]),
+        input_border_color_focus=("#4d9fff", "#4d9fff"),
+        button_primary_background_fill=("#4d9fff", "#4d9fff"), button_primary_background_fill_hover=("#74b4ff", "#74b4ff"),
+        button_primary_text_color=("#07182e", "#07182e"), button_primary_border_color=("#4d9fff", "#4d9fff"),
+        button_secondary_background_fill=(_L["surface2"], _D["surface2"]),
+        button_secondary_background_fill_hover=(_L["line"], _D["line"]),
+        button_secondary_text_color=(_L["text"], _D["text"]), button_secondary_border_color=(_L["line"], _D["line"]),
+        checkbox_background_color_selected=("#4d9fff", "#4d9fff"), checkbox_border_color_selected=("#4d9fff", "#4d9fff"),
+        checkbox_label_background_fill=(_L["surface2"], _D["surface2"]),
+        checkbox_label_background_fill_selected=(_L["surface2"], _D["surface2"]),
+        checkbox_label_border_color=(_L["line"], _D["line"]), checkbox_label_border_color_selected=("#4d9fff", "#4d9fff"),
+        checkbox_label_text_color_selected=(_L["text"], _D["text"]),
+        slider_color=("#4d9fff", "#4d9fff"), loader_color=("#4d9fff", "#4d9fff"),
+        link_text_color=(_L["accent"], _D["accent"]), link_text_color_hover=(_L["accent"], _D["accent"]),
+        panel_background_fill=(_L["surface"], _D["surface"]), panel_border_color=(_L["line"], _D["line"]),
+        table_border_color=(_L["line"], _D["line"]), code_background_fill=(_L["surface2"], _D["surface2"]),
+    ),
+)
+
+APP_CSS = """
+@import url('https://fonts.googleapis.com/css2?family=DM+Sans:wght@400;500;600;700&family=Space+Grotesk:wght@400;500;600;700&family=Press+Start+2P&display=swap');
+.gradio-container{font-family:'DM Sans',Arial,sans-serif!important}
+.gradio-container h1,.gradio-container h2,.gradio-container h3{font-family:'Space Grotesk',Arial,sans-serif!important;letter-spacing:-.4px}
+.ah-heading{display:flex;justify-content:space-between;align-items:center;gap:20px;padding:10px 0 6px}
+.ah-eyebrow{font:11px/1.5 monospace!important;letter-spacing:1.9px;color:var(--body-text-color-subdued)!important;margin:0 0 10px!important}
+.ah-heading h1{font:500 clamp(32px,4vw,48px)/1.2 'Space Grotesk',Arial,sans-serif!important;letter-spacing:-2px!important;margin:0!important;
+  color:var(--body-text-color)!important}
+.ah-accent{color:#1b64c4}.dark .ah-accent{color:#4d9fff}
+.ah-intro{margin:10px 0 0!important;font-size:15px!important;color:var(--body-text-color-subdued)!important;max-width:860px}
+.ah-version{font:11px monospace;letter-spacing:1px;color:var(--body-text-color-subdued);display:flex;align-items:center;gap:10px;white-space:nowrap}
+.ah-dot{height:6px;width:6px;background:#95c79a;border-radius:50%}
+.ah-note{margin:0 0 4px!important;font-size:14px!important;line-height:1.55;color:var(--body-text-color-subdued)!important;max-width:900px}
+.ah-note b{color:var(--body-text-color)}
+.ah-warn{margin-top:14px;padding:10px 14px;border:1px solid #4d9fff66;border-radius:8px;font-size:13px;color:var(--body-text-color)}
+button[role=tab]{font-size:14px!important;color:var(--body-text-color-subdued)!important}
+button[role=tab][aria-selected=true]{color:var(--body-text-color)!important;border-color:#4d9fff!important}
+@media(max-width:550px){.ah-version{display:none}}
+"""
+
+
+def _status(text, kind="info"):
+    icon = {"info": "⏳", "ok": "✅", "err": "⛔", "warn": "⚠️"}[kind]
+    return f"{icon} {text}"
+
+
+def run_match(model_ids, baselines, protocol, seed, delay):
+    """Friendly match: any models, any seed, nothing is recorded (Elo only changes in the Ranked tab)."""
+    model_ids = [m.strip() for m in (model_ids or []) if m and m.strip()]
+    model_ids = list(dict.fromkeys(model_ids))
+    baselines = baselines or []
+    protocol = protocol or "guided"
+    if not model_ids:
+        yield _status("Pick at least one language model.", "err"), empty_html(), ""
+        return
+    if len(model_ids) > MAX_MODELS:
+        yield _status(f"At most {MAX_MODELS} language models per match on this CPU.", "err"), empty_html(), ""
+        return
+    if len(model_ids) + len(baselines) < 2:
+        yield _status("A match needs at least 2 players: add another model or a baseline.", "err"), empty_html(), ""
+        return
+
+    players = []
+    try:
+        metas = []
+        for m in model_ids:
+            yield _status(f"Checking `{m}`…"), empty_html("Checking models…"), ""
+            meta = precheck(m)
+            if any(x["id"] == meta["id"] for x in metas):
+                continue  # same repo typed twice with different casing
+            metas.append(meta)
+        model_ids = [x["id"] for x in metas]
+        for i, (m, meta) in enumerate(zip(model_ids, metas), 1):
+            yield _status(f"Loading `{m}` on CPU ({i}/{len(model_ids)})… first load downloads the weights."), empty_html("Loading models…"), ""
+            players.append(load_player(m, meta))
+    except ModelRejected as e:
+        yield _status(str(e), "err"), empty_html("Match cancelled."), ""
+        return
+    if RANDOM_ID in baselines:
+        players.append(RandomPlayer())
+    if ORACLE_ID in baselines:
+        players.append(OracleReaderPlayer())
+
+    seed = int(seed) if seed else random.SystemRandom().randrange(1, 10**9)
+    games = [TetrisGame(seed) for _ in players]
+    mode = "friendly"
+    yield _status(f"Seed {seed} · {protocol} · {mode}. Scoring the first moves…"), arena_html(games, players), ""
+
+    last = time.time()
+    try:
+        while True:
+            active = [(g, p) for g, p in zip(games, players) if g.alive and g.pieces < MAX_PIECES]
+            if not active:
+                break
+            for g, p in active:
+                choose(g, p, protocol, seed)
+            elapsed = time.time() - last
+            if elapsed < delay:
+                time.sleep(delay - elapsed)
+            last = time.time()
+            n = max(g.pieces for g in games)
+            alive = sum(g.alive for g in games)
+            yield _status(f"Seed {seed} · {protocol} · {mode} · piece {n}/{MAX_PIECES} · {alive} still playing"), arena_html(games, players), ""
+    except ModelRejected as e:
+        yield _status(str(e), "err"), arena_html(games, players), ""
+        return
+    except Exception as e:
+        yield _status(f"A model crashed during play: {type(e).__name__}: {str(e)[:200]}", "err"), arena_html(games, players), ""
+        return
+
+    ranks = rank_games(games)
+    order = sorted(range(len(players)), key=lambda i: ranks[i])
+    elos = None
+    note = "Friendly match: Elo not changed (only matches in the Ranked tab count)."
+    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 ranked_play(protocol):
+    """Start a ranked match (models picked at random) or watch the one already running."""
+    match, started = RANKED.start_or_join(protocol or "guided")
+    joined = "" if started else f" · you joined the ranked match already in progress ({match.protocol})"
+    seen = -1
+    while True:
+        version, status, boards, results, done = match.snapshot()
+        if version != seen:
+            seen = version
+            yield status + joined, boards, results
+        if done:
+            return
+        time.sleep(0.1)
+
+
+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"""
+<div class="ah-heading"><div><p class="ah-eyebrow">SMALL MODELS. ZERO-SHOT TETRIS.</p>
+<h1>SLM Tetris Arena<span class="ah-accent">.</span></h1>
+<p class="ah-intro">Decoder-only language models ({short_params(MIN_PARAMS)}–{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.</p></div>
+<span class="ah-version">SEASON {SEASON} <span class="ah-dot"></span></span></div>
+"""
+
+HOW = f"""
+### How a model plays
+For every new piece the game lists all legal placements (rotation × column, hard drop), simulates each one and
+describes the outcome in plain English. The model never sees the grid; it judges the descriptions:
+
+```
+{PROMPTS['guided'].format(desc='drops the piece into the lowest part of the board, clears one line, creates no new holes, keeps the stack low and leaves the surface flat')}
+```
+
+The model's value for a placement is **log P(" good move") − log P(" bad move")** after that prompt. The placement with the
+highest value is played; exact ties are broken by a seeded coin that is identical for every player.
+Because the value is a difference, a model's general bias towards "good" or "bad" cancels out.
+
+### Protocols (separate leaderboards)
+- **Guided**: the first line states the goal ("clear lines, avoid holes, keep the stack low"). Tests reading comprehension.
+- **Blind**: `{PROMPTS['blind'].splitlines()[0]}` No rules; the model must already know what is good in Tetris.
+
+### Rules of a match
+- Same 7-bag piece sequence for everyone. The game ends at top-out or after {MAX_PIECES} pieces.
+- Placement = score (100/300/500/800 for 1/2/3/4 lines), then lines, then pieces survived.
+- **Match** tab (friendly): pick any 2+ players (up to {MAX_MODELS} language models of any size, plus optional baselines)
+  and the seed. Nothing is recorded.
+- **Ranked** tab: press Play and the arena picks up to {MAX_MODELS} models at random from its pool of {len(POOL.ids)} models,
+  {SIZE_RULE}, with a random seed. Nobody chooses who plays, so Elo can't be farmed.
+  Models with fewer ranked games are more likely to be picked, so every model gets played. The match runs on the server and counts even if
+  you close the page; only one ranked match runs at a time, and pressing Play while one is running lets you watch it.
+- Elo: K=32, multiplayer (every pair of players counts as a game, scaled by 1/(N−1)).
+
+### Baselines
+- **🎲 Random**: every placement ties, so it plays uniformly at random. This is the floor a model should beat.
+- **📏 Oracle reader**: reads the same descriptions and ranks them with fixed common sense (holes > lines > height > surface > landing).
+  This is roughly the ceiling for a perfect reader of the text.
+
+Baselines can join friendly matches for comparison. They never play ranked and never change anyone's Elo.
+
+### Model requirements
+Public, not gated, loads with `AutoModelForCausalLM` + `AutoTokenizer` (PyTorch or safetensors weights), between {fmt_params(MIN_PARAMS)} and {fmt_params(MAX_PARAMS)} parameters.
+Models with custom code (`auto_map`) load with `trust_remote_code=True`. That code runs on this Space's CPU, so only
+submit repos you trust. Prompts are in English (the language most pre-training corpora such as FineWeb-edu use).
+
+Results and every match (seed, commit SHA of each model, scores) are published in
+[`{RESULTS_REPO}`](https://huggingface.co/datasets/{RESULTS_REPO}).
+"""
+
+RANKED_INTRO = f"""<p class="ah-note">Press <b>Play</b> and the arena picks up to {MAX_MODELS} language models <b>at random</b>
+from its pool of {len(POOL.ids)} models, {SIZE_RULE}, with a random seed.
+The result updates the public Elo leaderboard. The match runs on the server: it finishes and counts even if you close the page.
+Only one ranked match runs at a time; if one is already running, you watch it.</p>"""
+
+with gr.Blocks(title="SLM Tetris Arena") as demo:
+    gr.HTML(INTRO, padding=False)
+    if not STORE.persistent:
+        gr.HTML('<div class="ah-warn">⚠️ Results are not being saved: add an <code>HF_TOKEN</code> secret with write access '
+                'to the results dataset.</div>', padding=False)
+    with gr.Tabs():
+        with gr.Tab("⚔️ Match"):
+            with gr.Row():
+                with gr.Column(scale=3):
+                    models = gr.Dropdown(
+                        choices=SUGGESTED, value=DEFAULT_MODELS, multiselect=True, allow_custom_value=True,
+                        max_choices=MAX_MODELS, label=f"Language models (1–{MAX_MODELS})",
+                        info="Pick from the list or type any Hub repo id (owner/name) and press Enter.",
+                    )
+                    baselines = gr.CheckboxGroup(BASELINE_CHOICES, value=[RANDOM_ID], label="Baselines (optional, never rated)")
+                with gr.Column(scale=2):
+                    protocol = gr.Radio(PROTOCOL_CHOICES, value="guided", label="Protocol")
+                    seed = gr.Number(value=42, precision=0, label="Seed (0 = random)",
+                                     info="Friendly match: any models, nothing is recorded. Elo only changes in the Ranked tab.")
+                    delay = gr.Slider(0, 0.5, value=0.12, step=0.02, label="Seconds per piece (viewing speed)")
+            with gr.Row():
+                start = gr.Button("▶ Start match", variant="primary")
+                stop = gr.Button("■ Stop", variant="secondary")
+            status = gr.Markdown(_status("Ready.", "ok"))
+            boards = gr.HTML(empty_html())
+            results = gr.HTML()
+        with gr.Tab("🏅 Ranked"):
+            gr.HTML(RANKED_INTRO, padding=False)
+            with gr.Row(equal_height=True):
+                r_protocol = gr.Radio(PROTOCOL_CHOICES, value="guided", label="Protocol", scale=3)
+                r_play = gr.Button("▶ Play ranked match", variant="primary", scale=1)
+            r_status = gr.Markdown(_status("Ready. Press Play: the arena picks the models.", "ok"))
+            r_boards = gr.HTML(empty_html("Press <b>Play ranked match</b>. The arena picks the models at random."))
+            r_results = gr.HTML()
+        with gr.Tab("🏆 Leaderboard"):
+            lb_protocol = gr.Radio(PROTOCOL_CHOICES, value="guided", label="Protocol")
+            lb = gr.HTML(leaderboard_view("guided"), padding=False)
+            lb_refresh = gr.Button("↻ Refresh", variant="secondary")
+        with gr.Tab("📖 How it works"):
+            gr.Markdown(HOW)
+
+    match_event = start.click(
+        run_match, [models, baselines, protocol, seed, delay], [status, boards, results], concurrency_limit=1,
+    )
+    stop.click(stop_status, status, status, cancels=[match_event])
+    # viewers only watch; the ranked match itself runs in one background thread
+    r_play.click(ranked_play, r_protocol, [r_status, r_boards, r_results], concurrency_limit=16,
+                 concurrency_id="ranked").then(leaderboard_view, lb_protocol, lb)
+    lb_protocol.change(leaderboard_view, lb_protocol, lb)
+    lb_refresh.click(refresh_leaderboard, lb_protocol, lb)
+    demo.load(leaderboard_view, lb_protocol, lb)
+
+demo.queue(max_size=32)
+
+if __name__ == "__main__":
+    demo.launch(css=APP_CSS + CSS + LB_CSS, theme=THEME, ssr_mode=False)