Datasets:

ML-Intern-lab
/

pre2post-chess-replication

#!/usr/bin/env python
"""Multi-turn chess puzzle evaluation for pre2post-chess checkpoints (T0/T2).
Two protocols, matched to the authors' released code
(repo pavelslab-nyu/pre2post-chess):
--protocol sft   (for SFT checkpoints; sft/evaluation/multiturn_puzzle_evaluator.py):
   CoT phase: generate up to 1024 tokens from {prompt}<T>, stopping at </T> or
   <eos>. If </T> never appears -> failure (cot_format).
   Move loop: per solver move, generate up to 32 tokens stopping at <call_env>
   or <eos>; the FIRST LAN move in the emission (excluding the stop token) must
   be legal and match the ground truth; env reply injected after <call_env>;
   max 5 solver moves. temp 1.0.
--protocol verl  (for RL checkpoints; rl/verl workers + reward_function_multiturn.py):
   Phase 1: fixed thinking budget (1024 tokens); a </T> emitted before the
   budget is exhausted is replaced with <sep> and thinking continues; at
   exhaustion </T> is force-appended. Phase 2: up to max_env_calls=6 turns
   within answer_budget tokens; stop at <call_env>, env reply appended after
   it. Extraction per turn: first LAN move after </T>, else last LAN move in
   the segment (authors' fallback). temp 1.0, response 2560 total.
Success (both) = every extracted solver move matches the ground-truth UCI line
(= the paper's pass@1; pass@k = solved within k samples).
"""
import argparse
import json
import os
import re
import sys
import time

import torch
import chess
import pandas as pd
from huggingface_hub import hf_hub_download, snapshot_download
from transformers import AutoModelForCausalLM, AutoTokenizer

BENCH_REPO = "pavelslab-nyu/chess_puzzle_benchmark"

LAN_RE = re.compile(r"^([KQRBNP])([a-h][1-8])(x?)([a-h][1-8])(=([QRBN]))?$")
CASTLE_RE = re.compile(r"^(O-O|O-O-O)([+#])?$")

BOS, EOS, T_START, T_END, SEP, CALL_ENV = 0, 1, 81, 82, 83, 84


def load_benchmark(bins, think=True, limit_per_bin=None, seed=0):
    cfg = "think" if think else "non-think"
    frames = []
    for b in bins:
        f = hf_hub_download(BENCH_REPO, f"{cfg}/test_{b}.parquet", repo_type="dataset")
        df = pd.read_parquet(f)
        if limit_per_bin and limit_per_bin > 0:
            df = df.sample(n=min(limit_per_bin, len(df)), random_state=seed)
        frames.append(df)
    df = pd.concat(frames, ignore_index=True)

    def _gt(v):
        if isinstance(v, list):
            return list(v)
        return list(eval(v))
    df["gt"] = df["reward_model"].apply(lambda r: _gt(r["ground_truth"]))
    df["env_replies"] = df["extra_info"].apply(lambda r: list(r["env_replies"]))
    df["fen"] = df["extra_info"].apply(lambda r: r["FEN"])
    df["rating"] = df["difficulty"]
    df["bin"] = df["data_source"]
    return df


def lan_to_uci(lan, board):
    if lan is None:
        return None
    s = lan.rstrip("+#")
    if s in ("O-O", "O-O-O"):
        for mv in board.legal_moves:
            if not board.is_castling(mv):
                continue
            to_file = chess.square_file(mv.to_square)
            if (s == "O-O" and to_file == 6) or (s == "O-O-O" and to_file == 2):
                return mv.uci()
        return None
    m = LAN_RE.match(s)
    if not m:
        return None
    piece, frm, _cap, to, _eq, promo = m.groups()
    sq = chess.parse_square(frm)
    p = board.piece_at(sq)
    if p is None or p.symbol().upper() != piece:
        return None
    uci = frm + to + (promo.lower() if promo else "")
    return uci if uci in {mv.uci() for mv in board.legal_moves} else None


def is_lan_word(w):
    return bool(LAN_RE.match(w.rstrip("+#"))) or bool(CASTLE_RE.match(w))


class Adapter:
    def __init__(self, tok):
        self.custom = tok.custom_tokenizer
        self.vocab = self.custom._tok2id

    def encode_prompt(self, prompt):
        return list(self.custom.encode(prompt))  # [bos] + pgn + <T>

    def lan_ids(self, lan):
        toks = self.custom._lan_move_to_tokens(lan)
        return [self.vocab[t] for t in toks]

    def decode(self, ids):
        return self.custom.decode(list(ids))


def first_lan_word(text):
    for w in text.split():
        if w in ("<T>", "</T>", "<call_env>", "<sep>"):
            continue
        if is_lan_word(w):
            return w
    return None


def extract_move_verl(adapter, gen_ids):
    """verl reward extraction: after </T>, else last LAN move in segment."""
    text = adapter.decode(gen_ids)
    idx = text.find("</T>")
    if idx >= 0:
        ans = text[idx + 4:].replace("<T>", " ").replace("</T>", " ").replace("<call_env>", " ")
        w = first_lan_word(ans)
        if w:
            return w, "after_t_end"
    seg = text.split("<call_env>")[0]
    for w in reversed(seg.split()):
        if w in ("<T>", "</T>", "<sep>"):
            continue
        if is_lan_word(w):
            return w, "fallback"
    return None, "none"


def parse_args():
    ap = argparse.ArgumentParser()
    ap.add_argument("--hf-repo", default=None)
    ap.add_argument("--model-subdir", default=None)
    ap.add_argument("--protocol", choices=["sft", "verl"], default="sft")
    ap.add_argument("--bins", nargs="+", default=["B3", "B4"])
    ap.add_argument("--k", type=int, default=16)
    ap.add_argument("--temp", type=float, default=1.0)
    ap.add_argument("--top-k", type=int, default=None)
    ap.add_argument("--thinking-budget", type=int, default=1024)   # verl protocol
    ap.add_argument("--answer-budget", type=int, default=1536)     # verl protocol
    ap.add_argument("--cot-cap", type=int, default=1024)           # sft protocol
    ap.add_argument("--move-cap", type=int, default=32)            # sft protocol
    ap.add_argument("--max-moves", type=int, default=5)            # sft protocol
    ap.add_argument("--round", type=int, default=256)
    ap.add_argument("--max-env-calls", type=int, default=6)
    ap.add_argument("--limit-per-bin", type=int, default=0)
    ap.add_argument("--seed", type=int, default=0)
    ap.add_argument("--debug-gens", type=int, default=0)
    ap.add_argument("--sweep", default=None)
    ap.add_argument("--out", default=None)
    return ap.parse_args()


DEFAULTS = dict(hf_repo=None, model_subdir=None, protocol="sft", bins=["B3", "B4"],
                k=16, temp=1.0, top_k=None, thinking_budget=1024, answer_budget=1536,
                cot_cap=1024, move_cap=32, max_moves=5, round=256, max_env_calls=6,
                limit_per_bin=0, seed=0, debug_gens=0, out=None)


@torch.no_grad()
def gen_round(model, contexts, max_new, stop_ids, temp, top_k, device):
    """Batched (left-padded) generation round; returns list of new-id lists."""
    maxlen = max(len(c) for c in contexts)
    input_ids = torch.full((len(contexts), maxlen), BOS, dtype=torch.long)
    attn = torch.zeros((len(contexts), maxlen), dtype=torch.long)
    for j, c in enumerate(contexts):
        input_ids[j, maxlen - len(c):] = torch.tensor(c)
        attn[j, maxlen - len(c):] = 1
    input_ids, attn = input_ids.to(device), attn.to(device)
    out = model.generate(
        input_ids=input_ids, attention_mask=attn, max_new_tokens=max_new,
        do_sample=temp > 0, temperature=max(temp, 1e-5), top_k=top_k,
        use_cache=True, pad_token_id=BOS,
        eos_token_id=stop_ids if stop_ids else None)
    return [strip_pads(out[j, maxlen:].tolist()) for j in range(len(contexts))]


def strip_pads(ids):
    """Remove trailing pad (BOS) tokens left by batched generation of finished
    sequences; also cut everything after the first EOS if present."""
    if EOS in ids:
        ids = ids[: ids.index(EOS) + 1]
    while ids and ids[-1] == BOS:
        ids = ids[:-1]
    return ids


# --------------------------------------------------------------------------- #
# sft protocol (natural CoT termination; sft/evaluation/multiturn_puzzle_evaluator.py)
# --------------------------------------------------------------------------- #

@torch.no_grad()
def eval_puzzle_sft(model, adapter, row, args, device):
    gt = row["gt"][: args.max_moves]
    env_replies = row["env_replies"]
    base = adapter.encode_prompt(row["prompt"])  # already ends with <T>
    k = args.k
    fail = {"ok": 0, "cot_format": 0, "move_format": 0, "illegal_move": 0, "wrong_move": 0}
    success = [0] * k
    first_ok = [0] * k
    gen_tokens = [0] * k
    boards = [chess.Board(row["fen"]) for _ in range(k)]
    dbg = 0

    # CoT phase: stop at </T> or <eos>, cot_cap tokens
    ctxs = [list(base) for _ in range(k)]
    finished = [False] * k
    has_t_end = [False] * k
    while not all(finished):
        act = [i for i in range(k) if not finished[i]]
        gens = gen_round(model, [ctxs[i] for i in act], args.cot_cap, [T_END, EOS],
                         args.temp, args.top_k, device)
        for j, i in enumerate(act):
            toks = gens[j]
            gen_tokens[i] += len(toks)
            ctxs[i] = ctxs[i] + toks
            has_t_end[i] = T_END in toks
            finished[i] = has_t_end[i] or toks and toks[-1] == EOS or len(toks) < args.cot_cap
    for i in range(k):
        if not has_t_end[i]:
            fail["cot_format"] += 1
    # continue only for samples with a well-formed CoT
    alive = [i for i in range(k) if has_t_end[i]]

    # Move loop
    turn = 0
    while alive and turn < len(gt):
        next_alive = []
        gens = gen_round(model, [ctxs[i] for i in alive], args.move_cap, [CALL_ENV, EOS],
                         args.temp, args.top_k, device)
        for j, i in enumerate(alive):
            toks = gens[j]
            gen_tokens[i] += len(toks)
            hit_call = CALL_ENV in toks
            if hit_call:
                cut = toks.index(CALL_ENV) + 1
                emission = toks[:cut]           # keep raw emission incl. <call_env>
                move_ids = toks[: cut - 1]
            else:
                emission = toks
                move_ids = toks
            lan = first_lan_word(adapter.decode(move_ids)) if move_ids else None
            if not hit_call:
                fail["move_format"] += 1
                if args.debug_gens and dbg < args.debug_gens:
                    print(f"[gen] turn={turn} move_format "
                          f"tail={adapter.decode(toks)[-160:]!r}", flush=True)
                    dbg += 1
                continue
            board_j = boards[j]
            uci = lan_to_uci(lan, board_j) if lan else None
            if uci is None:
                fail["illegal_move"] += 1
                if args.debug_gens and dbg < args.debug_gens:
                    print(f"[gen] turn={turn} illegal lan={lan!r} "
                          f"tail={adapter.decode(toks)[-160:]!r}", flush=True)
                    dbg += 1
                continue
            if uci != gt[turn]:
                fail["wrong_move"] += 1
                if args.debug_gens and dbg < args.debug_gens:
                    print(f"[gen] turn={turn} wrong lan={lan!r} uci={uci} gt={gt[turn]} "
                          f"tail={adapter.decode(toks)[-160:]!r}", flush=True)
                    dbg += 1
                continue
            if turn == 0:
                first_ok[i] = 1
            board_j.push(chess.Move.from_uci(uci))
            ctxs[i] = ctxs[i] + emission + adapter.lan_ids(lan)
            if turn == len(gt) - 1:
                success[i] = 1
                fail["ok"] += 1
                continue
            o = env_replies[turn]
            o_uci = lan_to_uci(o, board_j)
            if o_uci is None:
                continue
            board_j.push(chess.Move.from_uci(o_uci))
            ctxs[i] = ctxs[i] + adapter.lan_ids(o)
            next_alive.append(i)
        alive = next_alive
        turn += 1
    return success, first_ok, gen_tokens, fail


# --------------------------------------------------------------------------- #
# verl protocol (forced thinking budget + answer phase; for RL checkpoints)
# --------------------------------------------------------------------------- #

@torch.no_grad()
def eval_puzzle_verl(model, adapter, row, args, device):
    gt, env_replies = row["gt"], row["env_replies"]
    base = adapter.encode_prompt(row["prompt"])
    k = args.k
    total_budget = args.thinking_budget + args.answer_budget
    ctx = [list(base) for _ in range(k)]
    used = [0] * k
    done_think = [False] * k
    success = [0] * k
    gen_tokens = [0] * k
    src_counts = {"after_t_end": 0, "fallback": 0, "none": 0}
    boards = [chess.Board(row["fen"]) for _ in range(k)]
    dbg = 0

    while not all(done_think):
        act = [i for i in range(k) if not done_think[i]]
        gens = gen_round(model, [ctx[i] for i in act], args.round, [T_END, EOS],
                         args.temp, args.top_k, device)
        for j, i in enumerate(act):
            toks = gens[j][: args.thinking_budget - used[i]]
            if not toks:
                ctx[i] += [T_END]; used[i] += 1; done_think[i] = True; continue
            gen_tokens[i] += len(toks)
            used[i] += len(toks)
            if toks[-1] == T_END:
                if used[i] < args.thinking_budget:
                    ctx[i] += toks[:-1] + [SEP]
                else:
                    ctx[i] += toks
                    done_think[i] = True
            elif toks[-1] == EOS:
                ctx[i] += toks + [T_END]
                done_think[i] = True
            else:
                ctx[i] += toks
                if used[i] >= args.thinking_budget:
                    ctx[i] += [T_END]
                    done_think[i] = True
    for i in range(k):
        if ctx[i][-1] != T_END:
            ctx[i] += [T_END]

    turn = 0
    alive = list(range(k))
    n_calls = [0] * k
    while alive and turn < len(gt) and turn < args.max_env_calls:
        next_alive = []
        remaining = [max(total_budget - used[i], 1) for i in alive]
        max_round = max(1, min(args.round, max(remaining)))
        gens = gen_round(model, [ctx[i] for i in alive], max_round, [CALL_ENV, EOS],
                         args.temp, args.top_k, device)
        for j, i in enumerate(alive):
            toks = gens[j][: remaining[j]]
            if not toks:
                continue
            gen_tokens[i] += len(toks)
            used[i] += len(toks)
            hit_call = CALL_ENV in toks
            if hit_call:
                cut = toks.index(CALL_ENV) + 1
                ctx[i] += toks[:cut]
                emission = toks[:cut]
                n_calls[i] += 1
            else:
                ctx[i] += toks
                emission = toks
            lan, src = extract_move_verl(adapter, emission)
            src_counts[src] += 1
            if args.debug_gens and dbg < args.debug_gens:
                print(f"[gen] turn={turn} src={src} parsed={lan!r} gt={gt[turn]!r} "
                      f"tail={adapter.decode(emission)[-200:]!r}", flush=True)
                dbg += 1
            board_j = boards[j]
            uci = lan_to_uci(lan, board_j) if lan else None
            wrong = (uci is None) or (uci != gt[turn])
            if hit_call:
                if wrong:
                    continue
                board_j.push(chess.Move.from_uci(uci))
                ctx[i] += adapter.lan_ids(lan)
                if turn == len(gt) - 1:
                    success[i] = 1
                    continue
                o = env_replies[turn]
                o_uci = lan_to_uci(o, board_j)
                if o_uci is None:
                    continue
                board_j.push(chess.Move.from_uci(o_uci))
                ctx[i] += adapter.lan_ids(o)
                next_alive.append(i)
            else:
                if (not wrong) and turn == len(gt) - 1:
                    success[i] = 1
        alive = next_alive
        turn += 1
    return success, [0] * k, gen_tokens, src_counts


def run_eval(cfg):
    args = argparse.Namespace(**{**DEFAULTS, **cfg})
    torch.manual_seed(args.seed)

    local = snapshot_download(args.hf_repo, allow_patterns=f"{args.model_subdir}/*")
    path = os.path.join(local, args.model_subdir)
    device = "cuda" if torch.cuda.is_available() else "cpu"
    model = AutoModelForCausalLM.from_pretrained(path, trust_remote_code=True).to(device).eval()
    tok = AutoTokenizer.from_pretrained(path, trust_remote_code=True)
    adapter = Adapter(tok)

    df = load_benchmark(args.bins, think=True, limit_per_bin=args.limit_per_bin, seed=args.seed)
    print(f"evaluating {args.hf_repo}/{args.model_subdir} [{args.protocol}] on {len(df)} "
          f"puzzles ({args.bins}), k={args.k}, temp={args.temp}", flush=True)

    eval_fn = eval_puzzle_sft if args.protocol == "sft" else eval_puzzle_verl
    results = []
    diag_total = {"ok": 0, "cot_format": 0, "move_format": 0, "illegal_move": 0,
                  "wrong_move": 0, "after_t_end": 0, "fallback": 0, "none": 0}
    t0 = time.time()
    for _, row in df.iterrows():
        success, first_ok, gen_tokens, diag = eval_fn(model, adapter, row, args, device)
        for key, val in diag.items():
            diag_total[key] = diag_total.get(key, 0) + val
        results.append({
            "puzzle_id": row["extra_info"]["PuzzleId"],
            "bin": row["bin"],
            "rating": int(row["rating"]),
            "n_moves": len(row["gt"]),
            "successes": success,
            "pass1_mean": sum(success) / len(success),
            "solved_within_k": float(any(success)),
            "first_move_acc": sum(first_ok) / len(first_ok),
            "gen_tokens": sum(gen_tokens),
        })
        if len(results) % 10 == 0:
            print(f"  {len(results)}/{len(df)} puzzles, "
                  f"pass1={sum(r['pass1_mean'] for r in results)/len(results):.3f}, "
                  f"{time.time()-t0:.0f}s", flush=True)

    by_bin = {}
    for b in sorted(set(r["bin"] for r in results)):
        rs = [r for r in results if r["bin"] == b]
        by_bin[b] = {"n": len(rs),
                     "pass1_mean": sum(r["pass1_mean"] for r in rs) / len(rs),
                     "passk_any": sum(r["solved_within_k"] for r in rs) / len(rs),
                     "first_move_acc": sum(r["first_move_acc"] for r in rs) / len(rs),
                     "mean_gen_tokens": sum(r["gen_tokens"] for r in rs) / len(rs)}
    n = len(results)
    summary = {
        "config": vars(args),
        "summary": {
            "n_puzzles": n,
            "by_bin": by_bin,
            "pass1_mean_overall": sum(r["pass1_mean"] for r in results) / n,
            "passk_any_overall": sum(r["solved_within_k"] for r in results) / n,
            "first_move_acc_overall": sum(r["first_move_acc"] for r in results) / n,
            "diag": diag_total,
            "wall_seconds": time.time() - t0,
        },
    }
    os.makedirs(os.path.dirname(args.out) or ".", exist_ok=True)
    with open(args.out, "w") as f:
        json.dump(summary, f)
    print(json.dumps(summary["summary"], indent=2), flush=True)
    return summary["summary"]


CALIB_SWEEP = [
    {"hf_repo": "pavelslab-nyu/Chess-SFT-Models", "model_subdir": "model_20m_5.3B",
     "protocol": "sft", "bins": ["B1", "B2"], "k": 8, "limit_per_bin": 25,
     "debug_gens": 4, "out": "/work/results/calib_sft20_nat.json"},
    {"hf_repo": "pavelslab-nyu/Chess-SFT-Models", "model_subdir": "model_680m_16B",
     "protocol": "sft", "bins": ["B1", "B2"], "k": 8, "limit_per_bin": 25,
     "out": "/work/results/calib_sft680_nat.json"},
    {"hf_repo": "Pre2Post-Chess-RL/Chess-RL-Models", "model_subdir": "model_20m_5.3B/global_step_5000",
     "protocol": "verl", "bins": ["B1", "B2"], "k": 8, "limit_per_bin": 25,
     "thinking_budget": 1024, "answer_budget": 1536,
     "debug_gens": 3, "out": "results/calib_rl5000_verl.json"},
]


T2_RL_RUNS = ["model_20m_0.53B", "model_20m_1.6B", "model_20m_2.6B", "model_20m_5.3B",
              "model_20m_11B", "model_20m_16B", "model_20m_21B", "model_20m_32B",
              "model_20m_40B", "model_20m_53B"]
T2_STEP_TARGETS = [300, 750, 1500, 3000, 5000]
RESULTS_REPO = "ML-Intern-lab/pre2post-chess-replication"


def build_t2_sweep():
    """RL sweep over the 20M family: verl protocol, B3+B4, steps nearest targets."""
    from huggingface_hub import HfApi
    files = HfApi().list_repo_files("Pre2Post-Chess-RL/Chess-RL-Models", repo_type="model")
    avail = {}
    for f in files:
        parts = f.split("/")
        if len(parts) == 3 and parts[1].startswith("global_step_") and f.endswith("config.json"):
            run, step = parts[0], int(parts[1].replace("global_step_", ""))
            avail.setdefault(run, []).append(step)
    specs = []
    for run in T2_RL_RUNS:
        steps = sorted(avail.get(run, []))
        if not steps:
            print(f"[sweep] no checkpoints for {run}, skipping")
            continue
        chosen = sorted(set(min(steps, key=lambda s: abs(s - t)) for t in T2_STEP_TARGETS))
        for step in chosen:
            tag = f"{run.replace('model_', '').replace('.', 'p')}_s{step}"
            specs.append({
                "hf_repo": "Pre2Post-Chess-RL/Chess-RL-Models",
                "model_subdir": f"{run}/global_step_{step}",
                "protocol": "verl", "bins": ["B3", "B4"], "k": 8, "limit_per_bin": 25,
                "thinking_budget": 1024, "answer_budget": 1536,
                "out": f"/work/results/t2_rl_{tag}.json",
                "tag": f"rl_{tag}",
            })
    return specs


def upload_result(out_path, tag):
    try:
        from huggingface_hub import HfApi
        HfApi().upload_file(path_or_fileobj=out_path,
                            path_in_repo=f"results/t2/{tag}.json",
                            repo_id=RESULTS_REPO, repo_type="dataset")
        print(f"[upload] {tag} -> {RESULTS_REPO}", flush=True)
    except Exception as e:
        print(f"[upload] FAILED {tag}: {e}", flush=True)


def run_t2_sweep():
    specs = build_t2_sweep()
    print(f"T2 sweep: {len(specs)} evals", flush=True)
    for i, sp in enumerate(specs):
        if result_exists(sp["tag"], "t2"):
            print(f"[skip] {sp['tag']} already uploaded", flush=True)
            continue
        print(f"\n===== [{i+1}/{len(specs)}] {sp['model_subdir']} =====", flush=True)
        try:
            s = run_eval(sp)
            upload_result(sp["out"], sp["tag"])
            print(f"RESULT {sp['tag']} pass1={s['pass1_mean_overall']:.4f} "
                  f"passk={s['passk_any_overall']:.4f} wall={s['wall_seconds']:.0f}s", flush=True)
        except Exception as e:
            print(f"[sweep] eval {sp['tag']} FAILED: {e}", flush=True)


T2B_RUNS = ["model_20m_0.53B", "model_20m_1.6B", "model_20m_5.3B",
            "model_20m_11B", "model_20m_32B", "model_20m_53B"]
T2B_STEPS = [300, 1500, 5000]


def build_t2b_sweep():
    """Focused high-power sweep on B1+B2 (above the resolution floor at our
    reduced scale): 6 pretraining levels x 3 RL steps, 400 draws per point.
    Tests R_ref vs pretraining loss and RL slope vs log T (claims 4a/4b)."""
    from huggingface_hub import HfApi
    files = HfApi().list_repo_files("Pre2Post-Chess-RL/Chess-RL-Models", repo_type="model")
    avail = {}
    for f in files:
        parts = f.split("/")
        if len(parts) == 3 and parts[1].startswith("global_step_") and f.endswith("config.json"):
            avail.setdefault(parts[0], []).append(int(parts[1].replace("global_step_", "")))
    specs = []
    for run in T2B_RUNS:
        steps = sorted(avail.get(run, []))
        if not steps:
            continue
        for target in T2B_STEPS:
            step = min(steps, key=lambda s: abs(s - target))
            tag = f"{run.replace('model_', '').replace('.', 'p')}_s{step}"
            specs.append({
                "hf_repo": "Pre2Post-Chess-RL/Chess-RL-Models",
                "model_subdir": f"{run}/global_step_{step}",
                "protocol": "verl", "bins": ["B1", "B2"], "k": 8, "limit_per_bin": 40,
                "thinking_budget": 1024, "answer_budget": 1536,
                "out": f"/work/results/t2b_{tag}.json",
                "tag": f"t2b_{tag}",
            })
    # dedupe (same step chosen for two targets)
    seen, out = set(), []
    for s in specs:
        if s["tag"] in seen:
            continue
        seen.add(s["tag"])
        out.append(s)
    return out


def result_exists(tag, subdir):
    try:
        from huggingface_hub import HfApi
        path = f"results/{subdir}/{tag}.json"
        return path in set(HfApi().list_repo_files(RESULTS_REPO, repo_type="dataset"))
    except Exception:
        return False


def run_t2b_sweep():
    specs = build_t2b_sweep()
    print(f"T2b sweep: {len(specs)} evals", flush=True)
    for i, sp in enumerate(specs):
        if result_exists(sp["tag"], "t2b"):
            print(f"[skip] {sp['tag']} already uploaded", flush=True)
            continue
        print(f"\n===== [{i+1}/{len(specs)}] {sp['model_subdir']} =====", flush=True)
        try:
            s = run_eval(sp)
            upload_result(sp["out"], sp["tag"])
            print(f"RESULT {sp['tag']} pass1={s['pass1_mean_overall']:.4f} "
                  f"passk={s['passk_any_overall']:.4f} wall={s['wall_seconds']:.0f}s", flush=True)
        except Exception as e:
            print(f"[sweep] eval {sp['tag']} FAILED: {e}", flush=True)


def main():
    args = parse_args()
    if args.sweep is None and len(sys.argv) == 1:
        run_t2_sweep()   # completes any missing B3+B4 evals (skips uploaded)
        run_t2b_sweep()  # then the focused B1+B2 high-power sweep
        return
    if args.sweep:
        outs = []
        for sp in json.loads(args.sweep):
            print(f"\n===== run {sp.get('model_subdir')} [{sp.get('protocol')}] =====", flush=True)
            outs.append({"spec": sp, "summary": run_eval(sp)})
        print("\nSWEEP_SUMMARY\n" + json.dumps(outs, indent=2), flush=True)
    else:
        run_eval(vars(args))


if __name__ == "__main__":
    main()