score.py

#!/usr/bin/env python
"""Scorer for the long-context agent benchmark: does a model still use information that lies far back in a long

conversation? Five sets (facts, tool_calls, rules_en, rules_cs, rules_de), each at 8k and 32k tokens.



    python score.py --model Qwen/Qwen3.5-2B --data ./longctx-agent-bench --out result.json

    python score.py --model my/model --data ./longctx-agent-bench --reference Qwen/Qwen3.5-2B --out result.json

    python score.py --model my/model --data ./longctx-agent-bench --reference_scores ref_scores.json.gz \

        --save_scores my_scores.json.gz --out result.json [--sets facts,rules_en] [--lengths 8k] [--limit 2]



Everything is teacher-forced: one forward pass over each item's text, the model's next-token distribution read at

the target positions, nothing generated. Needs torch and transformers (pyarrow only for .parquet files).



Metrics (per set and length):

  facts       query_acc: a query counts as recalled iff every token of its answer (the number and the final ".") is

              the model's top-1 prediction; token_acc over the answer tokens.

  tool_calls  acc_far: next-token top-1 accuracy at the target tokens (a tool's name, its parameter names, the token

              that closes them, the next parameter / closing tag, a documented command's name) whose definition lies

              far back in the context; acc_fresh: the far targets without an earlier call of the same tool in the

              last 512 tokens (the model cannot copy a recent call); acc_near: the rest.

  rules_<lg>  each item holds one context twice, without and with a planted policy rule; at every target (a later

              action the rule governs) score = log p(the rule-implied first token); effect = score(with rule) -

              score(without rule). The REFERENCE model (default: the scored model itself) defines which targets

              qualify: |effect_ref| >= 1.0 nat. Over the qualifying targets: retained = sum effect / sum effect_ref,

              follow = share of targets with effect / effect_ref >= 0.5, per placement (rule in the system prompt /

              in a later user message) and family (substitution / redirect / required) and over all families.

Headline rows (result["rows"]):

  "Facts recalled at 8k / 32k tokens"                        facts query_acc

  "Distant tool calls at 8k / 32k tokens"                    tool_calls acc_far

  "Rules followed at 8k / 32k tokens"                        rules_en follow, rule in the system prompt, all families

  "Rules kept far back, worst case (1.00 = as the original)" rules_en: the lowest retained of the 12 cells

                                                             (3 families x 2 placements x 2 lengths)

  "Czech rules kept far back, worst case"                    rules_cs, the same

  "German rules kept far back, worst case"                   rules_de, the same

  "Rules in Czech / German followed at 32k tokens"           rules_cs / rules_de follow at 32k (system prompt, all)

Tokenization: every item stores the text and, for the Qwen3.5 tokenizer, its token ids. --mode ids feeds those ids

(models with the same vocabulary: the exact evaluated sequences); --mode text tokenizes the text with the model's own

tokenizer and maps the targets by character offsets (any tokenizer; a target whose characters do not start at a

token boundary of that tokenizer is skipped and counted under "unaligned"); --mode auto (default): ids when the

tokenizer's vocabulary matches the dataset's tokenizer_fingerprint, else text.

Backends (--backend): "hf" (default) any transformers causal LM; "thinctx" a ThinCtx model through the thinctx

runtime's public scoring call Engine.score (thinctx 1.1, upcoming; thinctx 1.0.x cannot score ThinCtx models

teacher-forced, and the backend then stops with that message). Other models / runtimes: score_sets(backend, tokenizer, ...) with any object

that has read(ids, positions, tokens) -> (top-1 id per position, log p of the given token ids per position), or

hidden(ids) -> [n, D] (the final hidden states of the token ids) and logits(h) -> [m, V] (HFBackend).

"""
from __future__ import annotations

import argparse
import gzip
import hashlib
import json
import time
from pathlib import Path

import numpy as np
import torch

SETS = ("facts", "tool_calls", "rules_en", "rules_cs", "rules_de")
SPLITS = ("8k", "32k")
FAMILIES = ("substitution", "redirect", "required")
PLACEMENTS = ("system", "user")
KINDS = ("name", "param", "next", "command")
MIN_EFFECT = 1.0             # nats: a rules target qualifies iff the reference reacts at least this much
FOLLOW = 0.5                 # follow: effect >= FOLLOW x the reference's effect
CHUNK = 256                  # positions per [positions, vocab] logits block (memory only)
ROWS = {
    "Facts recalled at 8k / 32k tokens": ("facts", "query_acc"),
    "Distant tool calls at 8k / 32k tokens": ("tool_calls", "acc_far"),
    "Rules followed at 8k / 32k tokens": ("rules_en", "follow"),
    "Rules kept far back, worst case (1.00 = as the original)": ("rules_en", "worst"),
    "Czech rules kept far back, worst case": ("rules_cs", "worst"),
    "German rules kept far back, worst case": ("rules_de", "worst"),
    "Rules in Czech / German followed at 32k tokens": ("rules_csde", "follow"),
}


# ---------------------------------------------------------------- data
def resolve_data(data: str) -> Path:
    """A local dataset folder, or a Hugging Face dataset repo id (downloaded with huggingface_hub)."""
    p = Path(data)
    if p.is_dir():
        return p
    from huggingface_hub import snapshot_download
    return Path(snapshot_download(repo_id=data, repo_type="dataset"))


def load_items(root: Path, name: str, split: str) -> list:
    """The items of one set and length: <root>[/data]/<set>/<split>.jsonl[.gz] or .parquet."""
    for base in (root / "data" / name, root / name):
        for ext in (".jsonl.gz", ".jsonl", ".parquet"):
            f = base / f"{split}{ext}"
            if not f.is_file():
                continue
            if ext == ".parquet":
                import pyarrow.parquet as pq
                return pq.read_table(f).to_pylist()
            op = gzip.open if ext == ".jsonl.gz" else open
            with op(f, "rt", encoding="utf-8") as fh:
                return [json.loads(line) for line in fh if line.strip()]
    raise FileNotFoundError(f"no {name}/{split}.jsonl[.gz] / .parquet under {root}")


def load_meta(root: Path) -> dict:
    for f in (root / "meta.json", root / "data" / "meta.json"):
        if f.is_file():
            return json.loads(f.read_text(encoding="utf-8"))
    return {}


def fingerprint(tok) -> str:
    """sha256 of the vocabulary (token string <TAB> id per line, by id): the dataset's tokenizer_fingerprint."""
    v = sorted(tok.get_vocab().items(), key=lambda kv: kv[1])
    return hashlib.sha256("\n".join(f"{t}\t{i}" for t, i in v).encode("utf-8")).hexdigest()


# ---------------------------------------------------------------- model backends
class HFBackend:
    """A transformers causal LM: hidden(ids) = the decoder's final hidden states (one forward, no cache),

    logits(h) = the output head as the model's own forward applies it."""

    def __init__(self, model):
        self.model = model.eval()
        self.dev = next(model.parameters()).device
        self.dec = model.get_decoder() if hasattr(model, "get_decoder") else None
        self.head = model.get_output_embeddings()
        cfg = model.config
        cfg = getattr(cfg, "text_config", None) or cfg
        self.scale = float(getattr(cfg, "logits_scaling", 1.0) or 1.0)       # Granite: logits / logits_scaling
        self.mult = float(getattr(cfg, "logit_scale", 1.0) or 1.0)           # Cohere: logits * logit_scale
        self.cap = getattr(cfg, "final_logit_softcapping", None)             # Gemma 2 / 3

    @torch.no_grad()
    def hidden(self, ids: list) -> torch.Tensor:
        x = torch.tensor([ids], dtype=torch.long, device=self.dev)
        if self.dec is not None:
            return self.dec(input_ids=x, use_cache=False).last_hidden_state[0]
        out = self.model(input_ids=x, use_cache=False, output_hidden_states=True)
        return out.hidden_states[-1][0]

    @torch.no_grad()
    def logits(self, h: torch.Tensor) -> torch.Tensor:
        z = self.head(h)
        if self.mult != 1.0:
            z = z * self.mult
        if self.scale != 1.0:
            z = z / self.scale
        if self.cap:
            z = torch.tanh(z / self.cap) * self.cap
        return z


class ThinCtxBackend:
    """A ThinCtx model through the thinctx runtime, public API only: thinctx.load(name) and

    Engine.score(ids, positions, tokens) -> {"argmax": [int per position], "logprobs": [[float per requested token]

    per position]} (the next-token distribution after ids[:p + 1], the sequence processed as the runtime processes a

    prompt). Engine.score comes with thinctx 1.1; thinctx 1.0.x does not have it."""

    def __init__(self, name: str):
        import thinctx
        self.eng = thinctx.load(name)
        if not callable(getattr(self.eng, "score", None)):
            raise SystemExit(f"thinctx {getattr(thinctx, '__version__', '?')} has no Engine.score (teacher-forced "
                             "log-probabilities at given positions): scoring a ThinCtx model needs thinctx >= 1.1")

    def read(self, ids: list, positions: list, tokens: list) -> tuple[list, list]:
        r = self.eng.score(list(ids), list(positions), [list(t) for t in tokens])
        return [int(v) for v in r["argmax"]], [[float(v) for v in row] for row in r["logprobs"]]


def load_hf(name: str, dtype: str = "bfloat16", device: str = "auto", attn: str | None = "sdpa"):
    """(HFBackend, tokenizer) of a hub id or local folder (attn: the attention implementation, None: the model's

    default)."""
    from transformers import AutoModelForCausalLM, AutoTokenizer
    dev = ("cuda" if torch.cuda.is_available() else "cpu") if device == "auto" else device
    tok = AutoTokenizer.from_pretrained(name)
    kw = {"dtype": getattr(torch, dtype)} | ({"attn_implementation": attn} if attn else {})
    model = AutoModelForCausalLM.from_pretrained(name, **kw).to(dev)
    return HFBackend(model), tok


# ---------------------------------------------------------------- tokenization
class TextTok:
    """The model tokenizer's view of a text: ids (with the tokenizer's BOS when it adds one), char spans."""

    def __init__(self, tok, text: str):
        enc = tok(text, add_special_tokens=False, return_offsets_mapping=True)
        ids, off = list(enc["input_ids"]), [tuple(o) for o in enc["offset_mapping"]]
        bos = getattr(tok, "bos_token_id", None)
        if bos is not None and tok("a", add_special_tokens=True)["input_ids"][:1] == [bos]:
            ids, off = [bos] + ids, [(0, 0)] + off
        self.tok, self.text, self.ids = tok, text, ids
        self.s = np.array([o[0] for o in off], dtype=np.int64)
        self.e = np.array([o[1] for o in off], dtype=np.int64)

    def overlapping(self, c0: int, c1: int) -> list:
        """Indices of the (non-empty) tokens overlapping characters [c0, c1)."""
        return np.flatnonzero((self.s < c1) & (self.e > c0) & (self.e > self.s)).tolist()

    def starting(self, c: int) -> int | None:
        """The index of the token starting at character c, None without a token boundary there."""
        j = np.flatnonzero((self.s == c) & (self.e > self.s))
        if not len(j):
            return None
        j = int(j[0])
        return j if j == 0 or self.e[:j].max(initial=0) <= c else None

    def first_token(self, c: int, cont: str, ctx: int = 48) -> int | None:
        """The model's first token of the continuation `cont` written at character c (re-tokenized with up to ctx

        preceding characters, which must keep their own tokenization)."""
        pre = self.text[max(0, c - ctx):c]
        a = self.tok(pre, add_special_tokens=False)["input_ids"]
        b = self.tok(pre + cont, add_special_tokens=False)["input_ids"]
        return int(b[len(a)]) if len(b) > len(a) and list(b[:len(a)]) == list(a) else None


# ---------------------------------------------------------------- requests: positions to read per sequence
class Reads:
    """Every sequence to run once (keyed by its ids), with the positions to read: argmax and log p of given ids."""

    def __init__(self):
        self.seqs = {}           # key -> {"ids": int64 array, "pos": {p: set(token ids for log p)}}
        self.out = {}            # (key, p) -> (argmax, {token: log p})

    def seq(self, ids) -> str:
        arr = np.asarray(ids, dtype=np.int64)
        key = hashlib.sha1(arr.tobytes()).hexdigest()
        self.seqs.setdefault(key, {"ids": arr, "pos": {}})
        return key

    def read(self, key: str, p: int, want=()) -> tuple:
        s = self.seqs[key]
        if not 0 <= int(p) < len(s["ids"]):
            raise IndexError(f"position {p} outside a sequence of {len(s['ids'])} tokens")
        s["pos"].setdefault(int(p), set()).update(int(t) for t in want)
        return key, int(p)

    def run(self, backend, log=None):
        n = len(self.seqs)
        t0 = time.time()
        for i, (key, s) in enumerate(sorted(self.seqs.items(), key=lambda kv: len(kv[1]["ids"]))):
            ps = sorted(s["pos"])
            if hasattr(backend, "read"):                 # a runtime that reads the positions itself
                need = [sorted(s["pos"][p]) for p in ps]
                am, lps = backend.read(s["ids"].tolist(), ps, need)
                for p, a_, ts, lp in zip(ps, am, need, lps):
                    self.out[(key, p)] = (int(a_), dict(zip(ts, (float(v) for v in lp))))
                if log is not None and ((i + 1) % 20 == 0 or i + 1 == n):
                    log(f"  {i + 1} / {n} sequences ({time.time() - t0:.0f} s)")
                continue
            h = backend.hidden(s["ids"].tolist())
            for c0 in range(0, len(ps), CHUNK):
                pc = ps[c0:c0 + CHUNK]
                z = backend.logits(h[torch.tensor(pc, device=h.device)]).float()
                am = z.argmax(-1).tolist()
                need = [sorted(s["pos"][p]) for p in pc]
                lp = z.log_softmax(-1) if any(need) else None
                for k, p in enumerate(pc):
                    self.out[(key, p)] = (int(am[k]), {t: float(lp[k, t]) for t in need[k]})
            del h
            if log is not None and ((i + 1) % 20 == 0 or i + 1 == n):
                log(f"  {i + 1} / {n} sequences ({time.time() - t0:.0f} s)")


# ---------------------------------------------------------------- per-set requests and results
def _plan_facts(items, mode, tok, R: Reads):
    plan = []
    for it in items:
        qs = []
        if mode == "ids":
            k = R.seq(it["qwen_ids"][:-1])
            for q in it["targets"]:
                qs.append([R.read(k, p) + (int(t),) for p, t in zip(q["qwen_pos"], q["qwen_answer_ids"])])
        else:
            tt = TextTok(tok, it["text"])
            k = R.seq(tt.ids)
            for q in it["targets"]:
                js = [j for j in tt.overlapping(*q["answer_char"]) if j >= 1]
                qs.append([R.read(k, j - 1) + (tt.ids[j],) for j in js] or None)
        plan.append((it["id"], qs))
    return plan


def _plan_tools(items, mode, tok, R: Reads):
    plan = []
    for it in items:
        tt = None if mode == "ids" else TextTok(tok, it["text"])
        k = R.seq(it["qwen_ids"][:-1] if mode == "ids" else tt.ids)
        ts = []
        for t in it["targets"]:
            if mode == "ids":
                y = int(t["qwen_target_id"])
                rd = [R.read(k, t["qwen_pos"], [y]) + (y,)]
            else:
                js = [j for j in tt.overlapping(*t["char"]) if j >= 1]
                rd = [R.read(k, j - 1, [tt.ids[j]]) + (tt.ids[j],) for j in js] or None
            ts.append((t, rd))
        plan.append((it["id"], ts))
    return plan


def _rule_read(R, mode, k, tt, t, which: str):
    """The read of a rules target in one of the item's two texts (k: its sequence; tt: its TextTok in text mode)."""
    if mode == "ids":
        y = int(t["expected_qwen_id"])
        return R.read(k, t[f"qwen_pos_{which}"], [y]) + (y,)
    c = int(t[f"char_{which}"])
    exp, orig = t["expected_text"], t["original_text"]
    ws = min(len(orig) - len(orig.lstrip()), len(exp) - len(exp.lstrip()))
    for d in range(ws, -1, -1):           # the latest token boundary within the shared leading whitespace
        j = tt.starting(c + d)
        if j is None or j < 1:
            continue
        y = tt.first_token(c + d, exp[d:])
        return None if y is None else R.read(k, j - 1, [y]) + (y,)
    return None


def _plan_rules(items, mode, tok, R: Reads):
    plan = []
    for it in items:
        if mode == "ids":
            a = b = None
            ka, kb = R.seq(it["qwen_ids_without_rule"][:-1]), R.seq(it["qwen_ids_with_rule"][:-1])
        else:
            a, b = TextTok(tok, it["text_without_rule"]), TextTok(tok, it["text_with_rule"])
            ka, kb = R.seq(a.ids), R.seq(b.ids)
        ts = [(t, _rule_read(R, mode, ka, a, t, "without_rule"), _rule_read(R, mode, kb, b, t, "with_rule"))
              for t in it["targets"]]
        plan.append((it, ts))
    return plan


def _val(R, rd):
    key, p, y = rd
    am, lp = R.out[(key, p)]
    return am == y, lp.get(y)


def _facts_scores(plan, R) -> dict:
    out = {}
    for iid, qs in plan:
        out[iid] = [None if q is None else [bool(_val(R, rd)[0]) for rd in q] for q in qs]
    return out


def _tools_scores(plan, R) -> dict:
    out = {}
    for iid, ts in plan:
        row = []
        for t, rd in ts:
            if rd is None:
                row.append(None)
                continue
            v = [_val(R, x) for x in rd]
            row.append([bool(all(ok for ok, _ in v)), float(sum(lp for _, lp in v))])
        out[iid] = row
    return out


def _rules_scores(plan, R) -> dict:
    out = {}
    for it, ts in plan:
        out[it["id"]] = [None if a is None or b is None else [_val(R, a)[1], _val(R, b)[1]] for _, a, b in ts]
    return out


def facts_metrics(items, sc) -> dict:
    q_ok = q_n = t_ok = t_n = skipped = 0
    for it in items:
        for q in sc[it["id"]]:
            if q is None:
                skipped += 1
                continue
            q_ok, q_n = q_ok + int(all(q)), q_n + 1
            t_ok, t_n = t_ok + sum(q), t_n + len(q)
    return {"query_acc": q_ok / max(1, q_n), "token_acc": t_ok / max(1, t_n), "n_queries": q_n,
            "unaligned": skipped}


def tools_metrics(items, sc) -> dict:
    acc, skipped = {}, 0
    for it in items:
        for t, v in zip(it["targets"], sc[it["id"]]):
            if v is None:
                skipped += 1
                continue
            groups = ["far" if t["far"] else "near"] + (["fresh"] if t["fresh"] else []) + (
                [f"far_{t['kind']}"] if t["far"] else [])
            for g in groups:
                a = acc.setdefault(g, [0, 0, 0.0])
                a[0], a[1], a[2] = a[0] + int(v[0]), a[1] + 1, a[2] + v[1]
    out = {"unaligned": skipped}
    for g, (ok, n, lp) in acc.items():
        out[f"acc_{g}"] = ok / n
        out[f"n_{g}"] = n
        if not g.startswith("far_"):
            out[f"logp_{g}"] = lp / n
    return out


def rules_summary(targets: list, eff_ref: np.ndarray, eff_sys: np.ndarray, min_effect: float = MIN_EFFECT) -> dict:
    """Per family (+ "all") metrics over the far targets [(family, ...)] given the reference's and the model's

    effects (module docstring)."""
    out = {}
    for fam in list(FAMILIES) + ["all"]:
        ids = [i for i, t in enumerate(targets) if fam == "all" or t["family"] == fam]
        if not ids:
            continue
        ef, es = eff_ref[ids], eff_sys[ids]
        q = np.abs(ef) >= float(min_effect)
        o = {"n": len(ids), "n_qual": int(q.sum())}
        if q.any():
            efq, esq = ef[q], es[q]
            o |= {"eff_ref": float(efq.mean()), "eff": float(esq.mean()), "frac_pos_ref": float((efq > 0).mean()),
                  "retained": float(esq.sum() / efq.sum()) if abs(efq.sum()) > 1e-9 else None,
                  "follow": float((esq / efq >= FOLLOW).mean()), "sign": float((esq * efq > 0).mean())}
        out[fam] = o
    return out


def rules_metrics(items, sc, ref) -> dict:
    tg, er, es, skipped = [], [], [], 0
    for it in items:
        for k, t in enumerate(it["targets"]):
            v, r = sc[it["id"]][k], ref[it["id"]][k]
            if v is None or r is None or None in v or None in r:
                skipped += 1
                continue
            tg.append({"family": it["family"], "placement": it["placement"]})
            er.append(r[1] - r[0])
            es.append(v[1] - v[0])
    er, es = np.asarray(er), np.asarray(es)
    out = {"by_placement": {}, "unaligned": skipped}
    for pc in PLACEMENTS:
        m = [i for i, t in enumerate(tg) if t["placement"] == pc]
        if m:
            out["by_placement"][pc] = rules_summary([tg[i] for i in m], er[m], es[m])
    return out


# ---------------------------------------------------------------- driver
def plan_sets(root: Path, sets, splits, mode: str, tok, limit: int | None = None):
    """(Reads, {(set, split): (kind, items, plan)})."""
    R, plans = Reads(), {}
    for name in sets:
        for sp in splits:
            items = load_items(root, name, sp)
            items = items[:limit] if limit else items
            if name == "facts":
                plans[(name, sp)] = ("facts", items, _plan_facts(items, mode, tok, R))
            elif name == "tool_calls":
                plans[(name, sp)] = ("tools", items, _plan_tools(items, mode, tok, R))
            else:
                plans[(name, sp)] = ("rules", items, _plan_rules(items, mode, tok, R))
    return R, plans


def raw_scores(R: Reads, plans: dict) -> dict:
    out = {}
    for (name, sp), (kind, items, plan) in plans.items():
        f = {"facts": _facts_scores, "tools": _tools_scores, "rules": _rules_scores}[kind]
        out.setdefault(name, {})[sp] = f(plan, R)
    return out


def choose_mode(mode: str, tok, meta: dict) -> str:
    if mode != "auto":
        return mode
    fp = meta.get("tokenizer_fingerprint")
    return "ids" if fp is not None and fingerprint(tok) == fp else "text"


def score_sets(backend, tok, root, sets=SETS, splits=SPLITS, mode: str = "auto", limit: int | None = None,

               log=print) -> tuple[str, dict]:
    """(mode, raw per-target scores {set: {split: {item id: [...]}}}) of a model backend."""
    root = Path(root)
    mode = choose_mode(mode, tok, load_meta(root))
    log(f"mode {mode}")
    R, plans = plan_sets(root, sets, splits, mode, tok, limit)
    log(f"{len(R.seqs)} sequences, {sum(len(s['pos']) for s in R.seqs.values())} positions")
    R.run(backend, log)
    return mode, raw_scores(R, plans)


def metrics(root, scores: dict, ref: dict | None = None, limit: int | None = None) -> dict:
    """{set: {split: metrics}} from raw scores (ref: the reference model's raw rules scores; None: the model's own)."""
    root = Path(root)
    out = {}
    for name, by in scores.items():
        for sp, sc in by.items():
            items = load_items(root, name, sp)
            items = items[:limit] if limit else items
            if name == "facts":
                m = facts_metrics(items, sc)
            elif name == "tool_calls":
                m = tools_metrics(items, sc)
            else:
                r = sc if ref is None else ref[name][sp]
                m = rules_metrics(items, sc, r)
            out.setdefault(name, {})[sp] = m
    return out


def worst(by_split: dict) -> tuple[float | None, str | None, int]:
    """The lowest retained over every (family, placement, length) cell, where it is, the number of cells."""
    cells = []
    for sp, m in by_split.items():
        for pc, sm in m.get("by_placement", {}).items():
            for fam in FAMILIES:
                v = sm.get(fam, {}).get("retained")
                if v is not None:
                    cells.append((v, f"{fam} / {pc} / {sp}"))
    if not cells:
        return None, None, 0
    lo = min(cells)
    return lo[0], lo[1], len(cells)


def rows(m: dict) -> dict:
    """The headline rows (module docstring) as the results page prints them (2 decimals)."""
    def f(v):
        return "–" if v is None else f"{v:.2f}"

    def follow(name, sp):
        return m.get(name, {}).get(sp, {}).get("by_placement", {}).get("system", {}).get("all", {}).get("follow")

    out = {}
    for label, (name, key) in ROWS.items():
        if key == "worst":
            if name in m:
                out[label] = f(worst(m[name])[0])
        elif name == "rules_csde":
            if "rules_cs" in m or "rules_de" in m:
                out[label] = f"{f(follow('rules_cs', '32k'))} / {f(follow('rules_de', '32k'))}"
        elif name in m:
            vals = [follow(name, sp) if key == "follow" else m[name].get(sp, {}).get(key) for sp in SPLITS]
            out[label] = " / ".join(f(v) for v in vals)
    return out


def _dump(p: str, obj):
    op = gzip.open if p.endswith(".gz") else open
    with op(p, "wt", encoding="utf-8") as f:
        json.dump(obj, f)


def _load(p: str):
    op = gzip.open if p.endswith(".gz") else open
    with op(p, "rt", encoding="utf-8") as f:
        return json.load(f)


def main(argv=None):
    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("--model", required=True, help="a Hugging Face causal LM (hub id or local folder)")
    ap.add_argument("--backend", default="hf", choices=("hf", "thinctx"),
                    help="hf: transformers; thinctx: a ThinCtx model through thinctx >= 1.1 (Engine.score)")
    ap.add_argument("--data", required=True, help="the dataset folder (or its hub id)")
    ap.add_argument("--sets", default=",".join(SETS))
    ap.add_argument("--lengths", default=",".join(SPLITS))
    ap.add_argument("--reference", default=None, help="the reference model of the rules sets (default: --model)")
    ap.add_argument("--reference_scores", default=None, help="the reference's raw scores (--save_scores of a run)")
    ap.add_argument("--save_scores", default=None, help="write the raw per-target scores here (.json or .json.gz)")
    ap.add_argument("--out", default=None, help="the result JSON (default: stdout only)")
    ap.add_argument("--mode", default="auto", choices=("auto", "ids", "text"))
    ap.add_argument("--dtype", default="bfloat16")
    ap.add_argument("--device", default="auto")
    ap.add_argument("--attn", default="sdpa", help="attention implementation (\"\" = the model's default)")
    ap.add_argument("--limit", type=int, default=None, help="only the first N items per set and length (a smoke test)")
    a = ap.parse_args(argv)
    root = resolve_data(a.data)
    sets = [s for s in a.sets.split(",") if s]
    splits = [s for s in a.lengths.split(",") if s]
    bad = set(sets) - set(SETS) or set(splits) - set(SPLITS)
    if bad:
        raise SystemExit(f"unknown sets / lengths {sorted(bad)}")
    t0 = time.time()
    log = lambda s: print(s, flush=True)          # noqa: E731
    rsets = [s for s in sets if s.startswith("rules_")]
    ref, ref_name = None, None
    if a.reference_scores:
        rr = _load(a.reference_scores)
        ref, ref_name = rr["scores"], rr.get("model")
    elif a.reference and a.reference != a.model and rsets:
        log(f"reference {a.reference}")
        be, tok = load_hf(a.reference, a.dtype, a.device, a.attn or None)
        _, rs = score_sets(be, tok, root, rsets, splits, a.mode, a.limit, log)
        ref, ref_name = rs, a.reference
        del be
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
    log(f"model {a.model}")
    if a.backend == "thinctx":
        be = ThinCtxBackend(a.model)
        from transformers import AutoTokenizer
        tok = AutoTokenizer.from_pretrained(a.model)
    else:
        be, tok = load_hf(a.model, a.dtype, a.device, a.attn or None)
    mode, sc = score_sets(be, tok, root, sets, splits, a.mode, a.limit, log)
    if a.save_scores:
        _dump(a.save_scores, {"model": a.model, "mode": mode, "scores": sc})
    if ref is None:
        ref_name = a.model
    m = metrics(root, sc, ref, a.limit)
    res = {"model": a.model, "reference": ref_name, "mode": mode, "limit": a.limit, "rows": rows(m), "sets": m,
           "time_s": round(time.time() - t0, 1)}
    for label, v in res["rows"].items():
        log(f"{label}: {v}")
    if a.out:
        Path(a.out).parent.mkdir(parents=True, exist_ok=True)
        Path(a.out).write_text(json.dumps(res, indent=1, ensure_ascii=False), encoding="utf-8")


if __name__ == "__main__":
    main()