alignmenter 0.3.2__tar.gz → 0.3.4__tar.gz

{alignmenter-0.3.2/src/alignmenter.egg-info → alignmenter-0.3.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: alignmenter
-Version: 0.3.2
+Version: 0.3.4
 Summary: Durable application alignment evaluations, evidence review, saved comparisons, and CI gates.
 Author: Alignmenter
 License-Expression: Apache-2.0

{alignmenter-0.3.2 → alignmenter-0.3.4}/src/alignmenter/_version.py

@@ -1,3 +1,3 @@
 """Single source for distribution and runtime version metadata."""
 
-__version__ = "0.3.2"
+__version__ = "0.3.4"

{alignmenter-0.3.2 → alignmenter-0.3.4}/src/alignmenter/cli.py

@@ -32,6 +32,7 @@ from alignmenter.providers.classifiers import load_safety_classifier
 from alignmenter.providers.judges import load_judge_provider
 from alignmenter.providers.openai import OpenAICustomGPTProvider
 from alignmenter.release_cli import register_release_commands
+from alignmenter.rubric_grade_cli import register_rubric_grade_command
 from alignmenter.run_config import load_run_options
 from alignmenter.runner import RunConfig, Runner
 from alignmenter.schemas.evaluation import EvaluationSpec, JudgeBudget, JudgeContract
@@ -48,6 +49,7 @@ from alignmenter.storage.runs import RunStore
 
 app = typer.Typer(help="Alignmenter — audit your model's alignment signals.")
 register_release_commands(app)
+register_rubric_grade_command(app)
 
 persona_app = typer.Typer(help="Persona helper commands.")
 dataset_app = typer.Typer(help="Dataset helper commands.")

alignmenter-0.3.4/src/alignmenter/dataset_cli.py

@@ -0,0 +1,398 @@
+"""Dataset-management commands: stats, validate, dedupe, merge, split, sample, manifest, import.
+
+Registered onto the existing `dataset` sub-app (alongside lint/sanitize/bootstrap)
+via register_dataset_commands(dataset_app). Builds on the shared primitives
+(content_digest, read/write_jsonl) and schemas/dataset.py.
+"""
+
+from __future__ import annotations
+
+import json
+import random
+from collections import Counter, defaultdict
+from pathlib import Path
+
+import typer
+
+from alignmenter.schemas.dataset import (
+    ProvenanceEntry,
+    build_manifest,
+    dataset_digest,
+    validate_records,
+)
+from alignmenter.schemas.execution import content_digest
+from alignmenter.utils.io import read_jsonl, write_json, write_jsonl
+
+
+def _group_key(record: dict, by: str) -> str:
+    """Grouping key for split — keeps a case + its variants on the same side."""
+    if not isinstance(record, dict):
+        return ""  # non-dict rows group together (like the other commands, don't crash)
+    metadata = record.get("metadata") or {}
+    session = record.get("session_id") or ""
+    if by == "split_group":
+        return str(metadata.get("split_group") or session)
+    if by == "group":
+        for tag in record.get("tags") or []:
+            if isinstance(tag, str) and tag.startswith("group:"):
+                return tag
+        return session
+    if by == "persona":
+        return str(record.get("persona_id") or session)
+    return session
+
+
+def register_dataset_commands(dataset_app: typer.Typer) -> None:
+    @dataset_app.command("stats")
+    def stats(
+        path: Path = typer.Argument(..., exists=True, dir_okay=False),
+        as_json: bool = typer.Option(False, "--json", help="Emit the coverage report as JSON."),
+    ):
+        """Coverage report — record/session counts, role balance, tag + persona histograms."""
+        records = read_jsonl(path)
+        manifest = build_manifest(records, id=path.stem, revision="stats")
+        roles = Counter(r.get("role") for r in records if isinstance(r, dict))
+        with_context = sum(
+            1
+            for r in records
+            if isinstance(r, dict)
+            and isinstance(r.get("metadata"), dict)
+            and "context" in r["metadata"]
+        )
+        if as_json:
+            typer.echo(
+                json.dumps(
+                    {
+                        "records": manifest.record_count,
+                        "sessions": manifest.session_count,
+                        "roles": {k: v for k, v in roles.items()},
+                        "with_context": with_context,
+                        "tags": manifest.tags,
+                        "personas": manifest.personas,
+                        "content_digest": manifest.content_digest,
+                    },
+                    indent=2,
+                )
+            )
+            return
+        typer.echo(
+            f"records {manifest.record_count} · sessions {manifest.session_count} · turns-with-context {with_context}"
+        )
+        typer.echo(
+            "roles: "
+            + ", ".join(f"{k}={v}" for k, v in sorted(roles.items(), key=lambda kv: str(kv[0])))
+        )
+        if manifest.tags:
+            typer.echo("tags:  " + ", ".join(f"{k}={v}" for k, v in sorted(manifest.tags.items())))
+        if manifest.personas:
+            typer.echo(
+                "personas: " + ", ".join(f"{k}={v}" for k, v in sorted(manifest.personas.items()))
+            )
+        typer.echo(f"content_digest: {manifest.content_digest[:12]}…")
+
+    @dataset_app.command("validate")
+    def validate(
+        path: Path = typer.Argument(..., exists=True, dir_okay=False),
+        strict: bool = typer.Option(
+            False,
+            "--strict",
+            help="Require turn_index+tags+persona_id and an assistant turn per session.",
+        ),
+    ):
+        """Schema validator — lenient by default, --strict adds turn_index+tags+persona_id + an assistant turn per session.
+
+        Complements `dataset lint`, which additionally checks turn-index contiguity,
+        scenario-tag coverage, and persona-file existence.
+        """
+        records = read_jsonl(path)
+        errors = validate_records(records, strict=strict)
+        for error in errors:
+            typer.echo(error, err=True)
+        typer.echo(f"{len(records)} records, {len(errors)} error(s)")
+        raise typer.Exit(1 if errors else 0)
+
+    @dataset_app.command("dedupe")
+    def dedupe(
+        path: Path = typer.Argument(..., exists=True, dir_okay=False),
+        out: Path | None = typer.Option(
+            None, "--out", help="Output path (default: <stem>.dedup.jsonl)."
+        ),
+        in_place: bool = typer.Option(False, "--in-place"),
+    ):
+        """Drop content-identical duplicate records (order preserved)."""
+        records = read_jsonl(path)
+        seen: set[str] = set()
+        kept = []
+        for record in records:
+            digest = content_digest(record)
+            if digest in seen:
+                continue
+            seen.add(digest)
+            kept.append(record)
+        destination = path if in_place else (out or path.with_name(f"{path.stem}.dedup.jsonl"))
+        write_jsonl(destination, kept)
+        typer.echo(
+            f"removed {len(records) - len(kept)} duplicate(s); wrote {len(kept)} -> {destination}"
+        )
+
+    @dataset_app.command("merge")
+    def merge(
+        paths: list[Path] = typer.Argument(..., exists=True, dir_okay=False),
+        out: Path = typer.Option(..., "--out"),
+        dedupe: bool = typer.Option(
+            False, "--dedupe", help="Drop content-identical records after merging."
+        ),
+        namespace_sessions: bool = typer.Option(
+            False,
+            "--namespace-sessions",
+            help="Prefix session_id with the source file stem to avoid collisions.",
+        ),
+    ):
+        """Concatenate datasets into one (optionally dedupe / namespace session ids)."""
+        merged: list[dict] = []
+        for source in paths:
+            records = read_jsonl(source)
+            if namespace_sessions:
+                for record in records:
+                    if isinstance(record, dict) and record.get("session_id"):
+                        record["session_id"] = f"{source.stem}:{record['session_id']}"
+            merged.extend(records)
+        if dedupe:
+            seen: set[str] = set()
+            deduped = []
+            for record in merged:
+                digest = content_digest(record)
+                if digest in seen:
+                    continue
+                seen.add(digest)
+                deduped.append(record)
+            merged = deduped
+        write_jsonl(out, merged)
+        typer.echo(f"merged {len(paths)} file(s) -> {len(merged)} records -> {out}")
+
+    @dataset_app.command("split")
+    def split(
+        path: Path = typer.Argument(..., exists=True, dir_okay=False),
+        out: Path = typer.Option(
+            ..., "--out", help="Output directory for train.jsonl + holdout.jsonl."
+        ),
+        holdout: float = typer.Option(
+            0.2, "--holdout", min=0.0, max=1.0, help="Target holdout fraction of records."
+        ),
+        by: str = typer.Option(
+            "split_group",
+            "--by",
+            help="Grouping unit kept together: split_group | group | session | persona.",
+        ),
+        seed: int = typer.Option(42, "--seed"),
+    ):
+        """Group-aware train/holdout split — a case and its variants never straddle the boundary."""
+        if by not in {"split_group", "group", "session", "persona"}:
+            raise typer.BadParameter("--by must be split_group, group, session, or persona")
+        records = read_jsonl(path)
+        groups: dict[str, list[dict]] = defaultdict(list)
+        for record in records:
+            groups[_group_key(record, by)].append(record)
+        names = sorted(groups)
+        random.Random(seed).shuffle(names)
+        target = holdout * len(records)
+        held: set[str] = set()
+        accumulated = 0
+        for name in names:
+            if accumulated >= target:
+                break
+            held.add(name)
+            accumulated += len(groups[name])
+        train = [r for r in records if _group_key(r, by) not in held]
+        holdout_records = [r for r in records if _group_key(r, by) in held]
+        write_jsonl(Path(out) / "train.jsonl", train)
+        write_jsonl(Path(out) / "holdout.jsonl", holdout_records)
+        typer.echo(
+            f"train {len(train)} · holdout {len(holdout_records)} "
+            f"({len(held)}/{len(names)} '{by}' groups) -> {out}"
+        )
+
+    @dataset_app.command("sample")
+    def sample(
+        path: Path = typer.Argument(..., exists=True, dir_okay=False),
+        out: Path = typer.Option(..., "--out", help="Output subset JSONL."),
+        n: int = typer.Option(..., "--n", min=1, help="Number of units to keep."),
+        by: str = typer.Option(
+            "session", "--by", help="Sampling unit kept whole: session | group | persona."
+        ),
+        filter_tag: list[str] = typer.Option(
+            [],
+            "--filter-tag",
+            help="Keep only units with a record carrying one of these tags (repeatable, OR). "
+            "e.g. --filter-tag tool:nutrition",
+        ),
+        seed: int = typer.Option(42, "--seed"),
+    ):
+        """Down-sample a dataset to N whole units, optionally pre-filtered by tag.
+
+        The change-aware selection primitive: a random seeded subset for regular
+        runs, or `--filter-tag tool:<name>` for a change-scoped subset. Units are
+        kept whole (a case + its variants never split) and output is ordered by
+        unit key for a stable content digest.
+        """
+        if by not in {"session", "group", "persona"}:
+            raise typer.BadParameter("--by must be session, group, or persona")
+        records = read_jsonl(path)
+        wanted = set(filter_tag)
+
+        groups: dict[str, list[dict]] = defaultdict(list)
+        for record in records:
+            groups[_group_key(record, by)].append(record)
+
+        def _keep(unit: list[dict]) -> bool:
+            if not wanted:
+                return True
+            return any(
+                isinstance(t, str) and t in wanted
+                for r in unit
+                if isinstance(r, dict)
+                for t in (r.get("tags") or [])
+            )
+
+        eligible = [name for name, unit in groups.items() if _keep(unit)]
+        if not eligible:
+            typer.echo(f"no units match filter {sorted(wanted)}", err=True)
+            raise typer.Exit(1)
+
+        names = sorted(eligible)
+        random.Random(seed).shuffle(names)
+        picked = sorted(names[:n])
+        subset = [r for name in picked for r in groups[name]]
+        write_jsonl(out, subset)
+        typer.echo(
+            f"sampled {len(picked)}/{len(eligible)} eligible '{by}' units "
+            f"({len(subset)} records"
+            + (f", filter {sorted(wanted)}" if wanted else "")
+            + f") -> {out}"
+        )
+
+    @dataset_app.command("manifest")
+    def manifest(
+        path: Path = typer.Argument(..., exists=True, dir_okay=False),
+        out: Path | None = typer.Option(
+            None, "--out", help="Write the manifest JSON here (default: stdout)."
+        ),
+        identifier: str | None = typer.Option(
+            None, "--id", help="Dataset id (default: file stem)."
+        ),
+        revision: str = typer.Option("v1", "--revision"),
+        verify: Path | None = typer.Option(
+            None, "--verify", help="Verify the data still matches this manifest's content_digest."
+        ),
+    ):
+        """Build (or --verify) a content-addressed dataset manifest with provenance."""
+        records = read_jsonl(path)
+        if verify is not None:
+            try:
+                expected = json.loads(Path(verify).read_text()).get("content_digest")
+            except (json.JSONDecodeError, OSError) as exc:
+                raise typer.BadParameter(f"Could not read manifest {verify}: {exc}") from exc
+            actual = dataset_digest(records)
+            match = expected == actual
+            typer.echo(
+                f"{'OK' if match else 'MISMATCH'}: manifest {str(expected)[:12]} vs data {actual[:12]}"
+            )
+            raise typer.Exit(0 if match else 2)
+        built = build_manifest(
+            records,
+            id=identifier or path.stem,
+            revision=revision,
+            provenance=(
+                ProvenanceEntry(
+                    kind="authored",
+                    ref=str(path),
+                    count=len(records),
+                    digest=dataset_digest(records),
+                ),
+            ),
+        )
+        payload = built.model_dump(mode="json")
+        if out is not None:
+            write_json(out, payload)
+            typer.echo(f"wrote manifest -> {out}")
+        else:
+            typer.echo(json.dumps(payload, indent=2))
+
+    @dataset_app.command("import")
+    def import_corpus_cmd(
+        source: str = typer.Argument(..., help="Corpus adapter (e.g. healthbench)."),
+        input_path: Path = typer.Argument(
+            ...,
+            metavar="INPUT",
+            exists=True,
+            dir_okay=False,
+            help="Local corpus JSONL you fetched/licensed.",
+        ),
+        out: Path = typer.Option(..., "--out", help="Output dataset JSONL."),
+        sample: int | None = typer.Option(
+            None,
+            "--sample",
+            min=1,
+            help="Down-sample to N sessions (stratified by the source's axis).",
+        ),
+        seed: int = typer.Option(42, "--seed"),
+        no_stratify: bool = typer.Option(
+            False, "--no-stratify", help="Plain random sample instead of stratified."
+        ),
+        manifest: bool = typer.Option(
+            False, "--manifest", help="Also write <out>.manifest.json with import provenance."
+        ),
+    ):
+        """Adapt a public eval corpus into Alignmenter turn records.
+
+        Corpora are fetched and licensed by you (importers never download); point
+        this at a local JSONL. Example:
+            alignmenter dataset import healthbench oss_eval.jsonl --out hb.jsonl --sample 60
+        """
+        from alignmenter.importers import available, get_importer, import_corpus
+
+        try:
+            mapper, default_prefix = get_importer(source)
+        except KeyError as exc:
+            raise typer.BadParameter(
+                f"unknown source {source!r}; available: {', '.join(available())}"
+            ) from exc
+
+        rows = read_jsonl(input_path)
+        records, report = import_corpus(
+            rows,
+            mapper,
+            sample=sample,
+            seed=seed,
+            stratify_prefix=None if no_stratify else default_prefix,
+        )
+        if not records:
+            typer.echo(
+                f"no records imported from {input_path} "
+                f"({report['input_rows']} rows, {report['skipped']} skipped)",
+                err=True,
+            )
+            raise typer.Exit(1)
+
+        write_jsonl(out, records)
+        deduped = report.get("deduped", 0)
+        typer.echo(
+            f"imported {report['records_out']} records / {report['sessions_out']} sessions "
+            f"from {report['input_rows']} rows ({report['skipped']} skipped"
+            + (f", {deduped} duplicate(s) dropped" if deduped else "")
+            + f") -> {out}"
+        )
+        if report.get("strata"):
+            typer.echo("strata: " + ", ".join(f"{k}={v}" for k, v in report["strata"].items()))
+
+        if manifest:
+            provenance = ProvenanceEntry(
+                kind="import",
+                ref=f"{source}:{input_path.name}",
+                count=len(records),
+                digest=dataset_digest(records),
+            )
+            built = build_manifest(records, id=out.stem, revision="v1", provenance=(provenance,))
+            manifest_path = out.with_suffix(out.suffix + ".manifest.json")
+            write_json(manifest_path, built.model_dump(mode="json"))
+            typer.echo(f"wrote manifest -> {manifest_path}")

alignmenter-0.3.4/src/alignmenter/importers/__init__.py

@@ -0,0 +1,163 @@
+"""Dataset importers: adapt public eval corpora into Alignmenter turn records.
+
+Each importer is a pure ``row -> list[record]`` mapper registered by name.
+``import_corpus`` owns the shared concerns — session-level sampling (optionally
+stratified so a small sample keeps topic coverage), deterministic ordering, and a
+counts report — so an adapter only has to describe the mapping.
+
+Importers never download: the corpora are fetched and licensed by the user, then
+this maps rows already read from a local JSONL. New corpora (K-QA, MedSafetyBench)
+add a module + one ``IMPORTERS`` entry.
+"""
+
+from __future__ import annotations
+
+import random
+from collections import defaultdict
+from collections.abc import Callable, Iterable
+
+from alignmenter.importers.healthbench import healthbench_to_records
+from alignmenter.schemas.execution import content_digest
+
+Mapper = Callable[[dict], list[dict]]
+
+# name -> (row mapper, default stratification tag-prefix)
+IMPORTERS: dict[str, tuple[Mapper, str]] = {
+    "healthbench": (healthbench_to_records, "theme:"),
+}
+
+
+def available() -> list[str]:
+    return sorted(IMPORTERS)
+
+
+def get_importer(name: str) -> tuple[Mapper, str]:
+    """Return (mapper, default_stratify_prefix); raises KeyError if unknown."""
+    return IMPORTERS[name]
+
+
+def _session_of(records: list[dict]) -> str:
+    return records[0].get("session_id", "") if records else ""
+
+
+def _stratum_of(records: list[dict], prefix: str) -> str:
+    for r in records:
+        for tag in r.get("tags") or []:
+            if isinstance(tag, str) and tag.startswith(prefix):
+                return tag
+    return "(none)"
+
+
+def _dedupe_and_disambiguate(sessions: list[list[dict]]) -> tuple[list[list[dict]], int]:
+    """Guarantee unique session ids across mapped sessions.
+
+    Two source rows can map to the same session_id — a content-addressed id
+    collides when rows share prompt content and carry no source id. Left alone
+    that emits duplicate turn_index within one session, which fails
+    ``validate_records``. So: drop byte-identical duplicates, and suffix an id
+    that collides with *different* content (`<id>#2`, `#3`, …). Returns the
+    cleaned sessions plus the count of exact duplicates dropped.
+    """
+    seen: dict[str, str] = {}  # session_id -> session content digest
+    kept: list[list[dict]] = []
+    dropped = 0
+    for session in sessions:
+        sid = _session_of(session)
+        digest = content_digest(session)
+        if sid not in seen:
+            seen[sid] = digest
+            kept.append(session)
+            continue
+        if seen[sid] == digest:
+            dropped += 1  # exact duplicate row
+            continue
+        suffix = 2
+        new_sid = f"{sid}#{suffix}"
+        while new_sid in seen:
+            suffix += 1
+            new_sid = f"{sid}#{suffix}"
+        for record in session:
+            record["session_id"] = new_sid
+        seen[new_sid] = digest
+        kept.append(session)
+    return kept, dropped
+
+
+def _stratified_sample(
+    sessions: list[list[dict]], sample: int, prefix: str, rng: random.Random
+) -> list[list[dict]]:
+    """Round-robin across strata (shuffled within each) for even coverage."""
+    buckets: dict[str, list[list[dict]]] = defaultdict(list)
+    for s in sessions:
+        buckets[_stratum_of(s, prefix)].append(s)
+    for b in buckets.values():
+        rng.shuffle(b)
+    keys = sorted(buckets)
+    picked: list[list[dict]] = []
+    i = 0
+    while len(picked) < sample and any(buckets[k] for k in keys):
+        key = keys[i % len(keys)]
+        if buckets[key]:
+            picked.append(buckets[key].pop())
+        i += 1
+    return picked
+
+
+def import_corpus(
+    rows: Iterable[dict],
+    mapper: Mapper,
+    *,
+    sample: int | None = None,
+    seed: int = 42,
+    stratify_prefix: str | None = None,
+) -> tuple[list[dict], dict]:
+    """Map ``rows`` to turn records, optionally down-sampling whole sessions.
+
+    Returns ``(records, report)``. Deterministic in input order: sessions are
+    deduped/disambiguated to unique ids, then sorted by id before any sampling,
+    so the same ``seed`` picks the same subset (and emits the same order — hence a
+    stable content digest) regardless of how the rows were ordered on input.
+    ``report`` carries input/skip/dedupe/session counts and the per-stratum
+    breakdown of what was kept.
+    """
+    sessions: list[list[dict]] = []
+    total = 0
+    skipped = 0
+    for row in rows:
+        total += 1
+        recs = mapper(row) if isinstance(row, dict) else []
+        if recs:
+            sessions.append(recs)
+        else:
+            skipped += 1
+
+    sessions, deduped = _dedupe_and_disambiguate(sessions)
+    # Sort before sampling so selection is input-order-invariant for a fixed seed.
+    sessions.sort(key=_session_of)
+
+    report: dict = {
+        "input_rows": total,
+        "skipped": skipped,
+        "deduped": deduped,
+        "sessions_in": len(sessions),
+    }
+
+    if sample is not None and sample < len(sessions):
+        rng = random.Random(seed)
+        if stratify_prefix:
+            sessions = _stratified_sample(sessions, sample, stratify_prefix, rng)
+        else:
+            rng.shuffle(sessions)
+            sessions = sessions[:sample]
+
+    sessions.sort(key=_session_of)
+    records = [r for s in sessions for r in s]
+
+    strata: dict[str, int] = defaultdict(int)
+    axis = stratify_prefix or "theme:"
+    for s in sessions:
+        strata[_stratum_of(s, axis)] += 1
+    report["sessions_out"] = len(sessions)
+    report["records_out"] = len(records)
+    report["strata"] = dict(sorted(strata.items()))
+    return records, report

alignmenter-0.3.4/src/alignmenter/importers/healthbench.py

@@ -0,0 +1,91 @@
+"""HealthBench (OpenAI, MIT) → Alignmenter turn records.
+
+HealthBench rows are physician-authored health conversations paired with rubric
+criteria but NO reference answer — grading is rubric-based, not exact-match. So
+each row becomes a *prompt session* (the conversation up to the turn the model
+must answer); capture fills the assistant turn, and the physician rubrics ride in
+the final user turn's ``metadata`` for a later rubric evaluator. Deterministic
+safety checks (refusal / escalation / non-diagnostic scope) still score the
+captured response for free.
+
+Row shape (``oss_eval.jsonl``)::
+
+    {"prompt": [{"role": "user", "content": "..."}, ...],
+     "rubrics": [{"criterion": "...", "points": 5, "tags": [...]}, ...],
+     "example_tags": ["theme:emergency_referrals", ...],
+     "prompt_id": "abc123"}
+
+The corpus is MIT-licensed but the user fetches it themselves — this adapter only
+maps rows already read from a local file.
+"""
+
+from __future__ import annotations
+
+import hashlib
+
+_ROLES = {"user", "assistant"}
+
+
+def _stable_id(prompt: list) -> str:
+    """Deterministic id for rows without a prompt_id (content-addressed)."""
+    raw = "\n".join(f"{m.get('role')}:{m.get('content')}" for m in prompt if isinstance(m, dict))
+    return hashlib.sha256(raw.encode("utf-8")).hexdigest()[:16]
+
+
+def _theme_tags(example_tags) -> list[str]:
+    """HealthBench axis tags (``theme:...``, ``physician_agreed_category:...``)
+    pass through unchanged; a bare tag gets a ``theme:`` prefix so stratified
+    sampling always has an axis to balance on."""
+    tags: list[str] = []
+    for t in example_tags or []:
+        if isinstance(t, str) and t.strip():
+            tags.append(t if ":" in t else f"theme:{t}")
+    return tags
+
+
+def healthbench_to_records(row: dict) -> list[dict]:
+    """Map one HealthBench row to a list of turn records (empty list = skip)."""
+    prompt = row.get("prompt")
+    if not isinstance(prompt, list) or not prompt:
+        return []
+
+    msgs: list[tuple[str, str]] = []
+    for m in prompt:
+        if not isinstance(m, dict):
+            return []
+        role = m.get("role")
+        content = m.get("content")
+        if role not in _ROLES or not isinstance(content, str) or not content.strip():
+            return []
+        msgs.append((role, content))
+
+    # The model answers the last turn; if it isn't a user turn there's nothing
+    # to capture against, so skip the row rather than emit an ungradeable session.
+    if msgs[-1][0] != "user":
+        return []
+
+    pid = row.get("prompt_id") or _stable_id(prompt)
+    session_id = f"healthbench:{pid}"
+    base_tags = ["source:healthbench", "group:healthbench", *_theme_tags(row.get("example_tags"))]
+
+    records: list[dict] = []
+    for i, (role, content) in enumerate(msgs):
+        records.append(
+            {
+                "session_id": session_id,
+                "turn_index": i + 1,
+                "role": role,
+                "text": content,
+                "tags": list(base_tags),
+                "persona_id": "reference",
+            }
+        )
+
+    rubrics = [r for r in (row.get("rubrics") or []) if isinstance(r, dict)]
+    records[-1]["metadata"] = {
+        "benchmark": "healthbench",
+        "prompt_id": pid,
+        "example_tags": [t for t in (row.get("example_tags") or []) if isinstance(t, str)],
+        "rubrics": rubrics,
+    }
+    return records