Add grobid_custom_hybrid profile with SciELO-preprints reference segmenter

.github/workflows/benchmark.yml

@@ -53,18 +53,34 @@ jobs:
           github.ref == 'refs/heads/main' && 'medium' ||
           'smoke'
         }}
-      BENCHMARK_PROFILE: >-
-        ${{
-          contains(github.event.pull_request.labels.*.name, 'profile:grobid_crf') && 'grobid_crf' ||
-          contains(github.event.pull_request.labels.*.name, 'profile:biorxiv_elife') && 'biorxiv_elife' ||
-          inputs.profile ||
-          'grobid_crf'
-        }}
       BENCHMARK_RUN: benchmarks/runs/${{ github.sha }}/validation
 
     steps:
       - uses: actions/checkout@v5
 
+      - name: Write GCS model reader key (temporary access to private model bucket)
+        id: gcp_auth
+        env:
+          GCS_MODEL_READER_SA_KEY_B64: ${{ secrets.GCS_MODEL_READER_SA_KEY_B64 }}
+        run: |
+          KEY_PATH="${RUNNER_TEMP}/gcs-model-reader-key.json"
+          printf '%s' "$GCS_MODEL_READER_SA_KEY_B64" | base64 -d > "$KEY_PATH"
+          chmod 600 "$KEY_PATH"
+          echo "key_path=$KEY_PATH" >> "$GITHUB_OUTPUT"
+
+      - name: Resolve benchmark profile
+        env:
+          PR_LABELS_JSON: ${{ toJson(github.event.pull_request.labels.*.name) }}
+          DISPATCH_PROFILE: ${{ inputs.profile }}
+        run: |
+          LABEL_PROFILE=$(echo "$PR_LABELS_JSON" | jq -r '.[] | select(startswith("profile:")) | sub("^profile:"; "")' | head -n1)
+          PROFILE="${LABEL_PROFILE:-${DISPATCH_PROFILE:-grobid_crf}}"
+          if ! [[ "$PROFILE" =~ ^[A-Za-z0-9_]+$ ]]; then
+            echo "::error::Invalid profile '$PROFILE' (from label or input) - only letters, digits and underscores are allowed"
+            exit 1
+          fi
+          echo "BENCHMARK_PROFILE=$PROFILE" >> "$GITHUB_ENV"
+
       - name: Set image tag
         id: image_tag
         run: |
@@ -139,6 +155,8 @@ jobs:
           fi
           docker run -d --name sciencebeam-parser -p 8080:8070 \
             -e SCIENCEBEAM_PARSER__PRELOAD_ON_STARTUP=true \
+            -v "${{ steps.gcp_auth.outputs.key_path }}:/creds/gcp-key.json:ro" \
+            -e GOOGLE_APPLICATION_CREDENTIALS=/creds/gcp-key.json \
             "${PROFILE_ENV[@]}" ${{ steps.image_tag.outputs.value }}
 
       - name: Wait for parser

sciencebeam_parser/config/config.py

@@ -2,7 +2,7 @@
 import os
 import copy
 from pathlib import Path
-from typing import Any, Optional, Union
+from typing import Any, FrozenSet, Optional, Union
 
 import yaml
 
@@ -27,6 +27,29 @@ def _deep_merge(base: dict, overlay: dict) -> dict:
     return result
 
 
+def _resolve_sequence_model_profile(
+    seq_profiles: dict,
+    name: str,
+    _seen: FrozenSet[str] = frozenset()
+) -> dict:
+    if name not in seq_profiles:
+        raise ValueError(
+            f'Unknown sequence_model_profile {name!r}. Available: {sorted(seq_profiles)}'
+        )
+    if name in _seen:
+        raise ValueError(
+            f'Circular extends detected for sequence_model_profile {name!r} '
+            f'(chain: {" -> ".join([*_seen, name])})'
+        )
+    profile = seq_profiles[name]
+    base_name = profile.get('extends')
+    overlay = {key: value for key, value in profile.items() if key != 'extends'}
+    if not base_name:
+        return overlay
+    base = _resolve_sequence_model_profile(seq_profiles, base_name, _seen | {name})
+    return _deep_merge(base, overlay)
+
+
 class AppConfig:
     def __init__(self, props: dict):
         self.props = props
@@ -87,7 +110,7 @@ def resolve_profile(self, profile_name: Optional[str] = None) -> 'AppConfig':
                     f'Profile {resolved!r} references unknown sequence_model_profile '
                     f'{seq_name!r}. Available: {sorted(seq_profiles)}'
                 )
-            overlay['models'] = seq_profiles[seq_name]
+            overlay['models'] = _resolve_sequence_model_profile(seq_profiles, seq_name)
 
         for key, value in profile.items():
             if key != 'sequence_models':

sciencebeam_parser/resources/default_config/config.yml

@@ -179,6 +179,14 @@ sequence_model_profiles:
     citation:
       path: 'https://github.com/kermitt2/grobid/raw/0.9.0/grobid-home/models/citation'
       engine: 'wapiti'
+  grobid_custom_hybrid:
+    extends: grobid_crf_0_9_0
+    reference_segmenter:
+      # TODO: switch back once the release is public
+      # path: 'https://github.com/eLifePathways/sciencebeam-models/releases/download/dev/2026-07-01-delft-grobid-reference-segmenter-CustomBidLSTM_CRF-grobid_090-scielo_preprints-v1.tar.gz'
+      # path: 'gs://sciencebeam-v2-models/delft/reference-segmenter/models/2026-07-01_CustomBidLSTM_CRF_grobid_090_scielo_preprints_ore/reference-segmenter/'
+      path: 'gs://sciencebeam-v2-models/wapiti/reference-segmenter/models/2026-07-01_wapiti_grobid_090_scielo_preprints_ore/reference-segmenter'
+      engine: 'wapiti'
 
 profiles:
   biorxiv_elife:
@@ -188,6 +196,11 @@ profiles:
     processors:
       fulltext:
         noise_filter_enabled: false
+  grobid_custom_hybrid:
+    sequence_models: grobid_custom_hybrid
+    processors:
+      fulltext:
+        noise_filter_enabled: false
 
 profile_aliases:
   grobid_crf: grobid_crf_0_9_0

tests/config/config_test.py

@@ -5,7 +5,11 @@
 import pytest
 import yaml
 
-from sciencebeam_parser.config.config import AppConfig, _deep_merge
+from sciencebeam_parser.config.config import (
+    AppConfig,
+    _deep_merge,
+    _resolve_sequence_model_profile
+)
 
 
 MINIMAL_PROFILE_CONFIG = {
@@ -18,10 +22,15 @@
             'segmentation': {'path': 'path_b/segmentation', 'engine': 'wapiti'},
             'header': {'path': 'path_b/header', 'engine': 'wapiti'},
         },
+        'profile_b_extended': {
+            'extends': 'profile_b',
+            'header': {'path': 'path_b_extended/header', 'engine': 'wapiti'},
+        },
     },
     'profiles': {
         'profile_a': {'sequence_models': 'profile_a'},
         'profile_b': {'sequence_models': 'profile_b'},
+        'profile_b_extended': {'sequence_models': 'profile_b_extended'},
         'profile_with_extra': {
             'sequence_models': 'profile_a',
             'processors': {'fulltext': {'use_cv_model': True}},
@@ -67,6 +76,59 @@ def test_does_not_mutate_base(self):
         assert base['a']['b'] == 1
 
 
+class TestResolveSequenceModelProfile:
+    def test_returns_profile_without_extends_unchanged(self):
+        seq_profiles = {
+            'base': {'segmentation': {'path': 'base/segmentation'}},
+        }
+        result = _resolve_sequence_model_profile(seq_profiles, 'base')
+        assert result == {'segmentation': {'path': 'base/segmentation'}}
+
+    def test_merges_extended_profile(self):
+        seq_profiles = {
+            'base': {
+                'segmentation': {'path': 'base/segmentation', 'engine': 'wapiti'},
+                'header': {'path': 'base/header', 'engine': 'wapiti'},
+            },
+            'child': {
+                'extends': 'base',
+                'header': {'path': 'child/header'},
+            },
+        }
+        result = _resolve_sequence_model_profile(seq_profiles, 'child')
+        assert result['segmentation'] == {'path': 'base/segmentation', 'engine': 'wapiti'}
+        assert result['header'] == {'path': 'child/header', 'engine': 'wapiti'}
+        assert 'extends' not in result
+
+    def test_supports_chained_extends(self):
+        seq_profiles = {
+            'grandparent': {'segmentation': {'path': 'gp/segmentation'}},
+            'parent': {'extends': 'grandparent', 'header': {'path': 'p/header'}},
+            'child': {'extends': 'parent', 'table': {'path': 'c/table'}},
+        }
+        result = _resolve_sequence_model_profile(seq_profiles, 'child')
+        assert result['segmentation'] == {'path': 'gp/segmentation'}
+        assert result['header'] == {'path': 'p/header'}
+        assert result['table'] == {'path': 'c/table'}
+
+    def test_raises_on_unknown_profile(self):
+        with pytest.raises(ValueError, match='Unknown sequence_model_profile'):
+            _resolve_sequence_model_profile({}, 'missing')
+
+    def test_raises_on_unknown_extends_target(self):
+        seq_profiles = {'child': {'extends': 'missing'}}
+        with pytest.raises(ValueError, match='Unknown sequence_model_profile'):
+            _resolve_sequence_model_profile(seq_profiles, 'child')
+
+    def test_raises_on_circular_extends(self):
+        seq_profiles = {
+            'a': {'extends': 'b'},
+            'b': {'extends': 'a'},
+        }
+        with pytest.raises(ValueError, match='Circular extends'):
+            _resolve_sequence_model_profile(seq_profiles, 'a')
+
+
 class TestAppConfigResolveProfile:
     def _make_config(self, extra: Optional[dict] = None) -> AppConfig:
         props = dict(MINIMAL_PROFILE_CONFIG)
@@ -80,6 +142,11 @@ def test_applies_sequence_model_profile(self):
         assert config['models']['segmentation']['engine'] == 'wapiti'
         assert config['models']['header']['path'] == 'path_b/header'
 
+    def test_applies_extended_sequence_model_profile(self):
+        config = self._make_config().resolve_profile('profile_b_extended')
+        assert config['models']['segmentation']['path'] == 'path_b/segmentation'
+        assert config['models']['header']['path'] == 'path_b_extended/header'
+
     def test_inherits_base_model_keys_not_in_profile(self):
         config = self._make_config().resolve_profile('profile_a')
         assert config['models']['segmentation']['path'] == 'path_a/segmentation'