Geraldine/humatheque-vlm-evaluation-results

Humathèque — VLM metadata-extraction evaluation

Structured metadata extraction from French thesis/dissertation title pages, scored against a human-validated visible-only ground truth (catalogue role/jury data not on the page was removed).

Leaderboard

model	overall	surface	stable	role	valid_json
lift	0.931	0.51	0.92	0.96	1
qwen35-9b	0.93	0.713	0.921	0.955	1
gemma4-12b	0.914	0.671	0.914	0.913	1
ministral3-8b	0.91	0.736	0.893	0.954	1
qwen3vl-8b	0.909	0.765	0.893	0.952	1
nuextract3	0.875	0.453	0.859	0.919	1
olmocr2	0.788	0.511	0.798	0.762	1

Stable vs role-attribution fields

Role fields (advisor/jury_president/reviewers/committee_members) are the hard case — the catalogue over-reports them, so this measures whether models correctly extract only what is on the page.

Per-field scores

Surface fidelity

Secondary, case- and accent-sensitive score, computed only on values whose content already matched the ground truth: does the model also render the string exactly (casing, French accents, civility stripping)? The primary scores above stay case-insensitive, so a casing error is not double-penalized as a content miss.

Thèse vs mémoire

Normal vs outlier annotations

Split by annot_type: outlier pages are the harder / atypical title pages, normal the standard layouts.

Output reliability

Confidence calibration

Per-field table

model	advisor	author	co_tutelle_institutions	committee_members	defense_year	degree_type	discipline	doctoral_school	granting_institution	jury_president	language	reviewers	subtitle	title	volume
lift	0.93	1	0.789	0.909	0.975	0.797	0.956	0.772	0.991	1	0.924	1	0.937	0.978	1
qwen35-9b	0.907	1	0.987	0.916	0.949	0.696	0.987	0.652	0.99	1	0.987	0.996	0.898	0.987	1
gemma4-12b	0.899	1	0.945	0.852	0.975	0.722	0.957	0.577	0.981	0.987	1	0.916	0.924	0.975	1
ministral3-8b	0.924	1	0.987	0.937	0.861	0.671	0.967	0.658	0.97	0.962	0.899	0.994	0.841	0.986	0.987
qwen3vl-8b	0.914	1	0.899	0.904	0.987	0.62	0.98	0.557	0.986	1	0.949	0.992	0.873	0.976	1
nuextract3	0.901	0.987	0.684	0.845	0.937	0.62	0.931	0.584	0.998	0.987	0.899	0.943	0.819	0.993	1
olmocr2	0.859	1	0.582	0.644	0.823	0.468	0.966	0.586	0.922	0.848	0.886	0.698	0.593	0.948	1

Prompts used

Prompts are selected per row by doc_type (thèse → thesis, mémoire → dissertation). The free-instruction models (olmocr2, gemma4-12b, qwen35-9b, qwen3vl-8b, ministral3-8b) share the thesis/dissertation prompt below. nuextract3 and lift are not free-instruction models and use model-specific strict-JSON schemas.

Free-instruction prompt — thesis.txt (the dissertation.txt variant differs only by the mémoire degree_type vocabulary and “thesis” → “dissertation”)

You are extracting metadata from a French graduate thesis title page.
General rules:
- Return ONLY valid JSON.
- Do not add markdown.
- Do not add explanations.
- Do not invent missing data.
- Multivalued fields are always JSON arrays; use [] when empty. Scalar fields use null when absent.
- If source text is written in all-caps, transform it into sentence case with French accents.
- For person names, apply sentence case (e.g. Jean-Daniel DUBOIS → Jean-Daniel Dubois, MARIE-FRANCE Dupont → Marie-France Dupont) and remove all civility and title prefixes (e.g. Monsieur, Madame, M., Mme, Professeur, Pr, Dr, etc.) before returning the name.
Role attribution rules:
- A role is explicit only when a textual role label is directly associated with the person.
- Assign a person to a specific role ONLY if that role is explicitly written on the title page.
- Do NOT infer roles from position, ordering, typography, academic conventions, or prior knowledge.
- Do not infer role equivalence from document structure or visual grouping.
- If people are listed without explicit role labels, place them ONLY in committee_members.
- If no jury or committee members are explicitly listed, use [].
- Do NOT guess who is advisor, reviewer, or jury president.
{
  "title": "Main title as it appears on the title page",
  "subtitle": "Subtitle or remainder of the title, usually following a colon.",
  "author": "Full name of the author (student) who wrote the graduate thesis",
  "degree_type": "Academic degree sought by the author. Possible values are Thèse d'État | Thèse de doctorat | Thèse de 3e cycle | Thèse d'université | Thèse de docteur-ingénieur | Thèse d'exercice",
  "discipline": "Academic field or discipline of the graduate thesis as explicitly mentioned. If not explicitly mentioned, infer it from the title and any other visible contextual elements (this is the only field where inference is allowed). Always return the discipline label in French (e.g. Histoire, Droit public, Sciences de l'éducation).",
  "volume": "Volume number if the title page explicitly mentions a volume, returned as the Arabic numeral only (e.g. \"1\" when the page prints Tome 1, TOME PREMIER or vol. I). Null if not explicitly mentioned.",
  "granting_institution": "Institution where the graduate thesis was submitted and the degree is granted",
  "co_tutelle_institutions": "List of institutions involved in a joint supervision or co-tutelle agreement.",
  "doctoral_school": "Doctoral school or graduate program, if explicitly mentioned",
  "defense_year": "Year the graduate thesis was defended. Format yyyy",
  "advisor": "List of ONLY persons explicitly identified as graduate thesis advisor/supervisor/director.",
  "jury_president": "ONLY the person explicitly identified as president/chair of the jury. Null if not explicitly stated.",
  "reviewers": "List of ONLY official reviewers/rapporteurs explicitly identified as such.",
  "committee_members": "List of persons explicitly listed as jury or committee members whose role is not explicitly identified as president/chair or reviewer/rapporteur. Include unlabeled persons here.",
  "language": "Language in ISO 639-3 codes. Example: fre, eng, ita...",
  "confidence": "Confidence score between 0.0 and 1.0 reflecting certainty based only on explicitly visible evidence on the page"
}

NuExtract3 — JSON template + free-text instructions (thesis variant)

JSON template (via chat_template_kwargs.template):

{
  "title": "string",
  "subtitle": "string",
  "author": "string",
  "degree_type": [
    "Thèse d'État",
    "Thèse de doctorat",
    "Thèse de 3e cycle",
    "Thèse d'université",
    "Thèse de docteur-ingénieur",
    "Thèse d'exercice"
  ],
  "discipline": "string",
  "volume": "string",
  "granting_institution": "string",
  "co_tutelle_institutions": [
    "string"
  ],
  "doctoral_school": "string",
  "defense_year": "integer",
  "advisor": [
    "string"
  ],
  "jury_president": "string",
  "reviewers": [
    "string"
  ],
  "committee_members": [
    "string"
  ],
  "language": "string"
}

Instructions (via chat_template_kwargs.instructions):

Extract only what is explicitly visible on the title page; do not invent missing data and do not infer roles from position, ordering, typography, or academic convention. Multivalued fields are always arrays: use [] when empty; use null for absent scalar fields. If source text is all-caps, transform it to sentence case with French accents. For person names, apply sentence case and remove civility/title prefixes (Monsieur, Madame, M., Mme, Professeur, Pr, Dr, etc.). A role is explicit only when a textual role label is directly associated with the person. If people are listed without an explicit role label, place them ONLY in committee_members. Exception for discipline: if not explicitly mentioned, infer it from the title and visible context, and always return the label in French (e.g. Histoire, Droit public, Sciences de l'éducation). volume is the volume number when the page explicitly mentions a volume, returned as the Arabic numeral only (e.g. "1" for Tome 1 or vol. I); null otherwise.

lift — JSON-Schema-in-prompt (thesis variant)

Extract structured data from this document according to the provided JSON schema.  The document is provided as images, in page order.

## JSON Schema
```json
{
  "type": "object",
  "properties": {
    "title": {
      "type": "string",
      "description": "Main title as it appears on the title page"
    },
    "subtitle": {
      "type": "string",
      "description": "Subtitle or remainder of the title, usually following a colon. Null if not present."
    },
    "author": {
      "type": "string",
      "description": "Full name of the author (student) who wrote the graduate thesis. Apply sentence case and remove all civility/title prefixes (Monsieur, Madame, M., Mme, Professeur, Pr, Dr, etc.)."
    },
    "degree_type": {
      "type": "string",
      "enum": [
        "Thèse d'État",
        "Thèse de doctorat",
        "Thèse de 3e cycle",
        "Thèse d'université",
        "Thèse de docteur-ingénieur",
        "Thèse d'exercice"
      ],
      "description": "Academic degree sought by the author. Possible values are Thèse d'État | Thèse de doctorat | Thèse de 3e cycle | Thèse d'université | Thèse de docteur-ingénieur | Thèse d'exercice"
    },
    "discipline": {
      "type": "string",
      "description": "Academic field or discipline of the graduate thesis as explicitly mentioned. If not explicitly mentioned, infer it from the title and any other visible contextual elements (this is the only field where inference is allowed). Always return the discipline label in French (e.g. Histoire, Droit public, Sciences de l'éducation)."
    },
    "volume": {
      "type": "string",
      "description": "Volume number if the title page explicitly mentions a volume, returned as the Arabic numeral only (e.g. \"1\" when the page prints Tome 1, TOME PREMIER or vol. I). Null if not explicitly mentioned."
    },
    "granting_institution": {
      "type": "string",
      "description": "Institution where the graduate thesis was submitted and the degree is granted"
    },
    "co_tutelle_institutions": {
      "type": "array",
      "items": {
        "type": "string"
      },
      "description": "List of institutions involved in a joint supervision or co-tutelle agreement. Empty array if none."
    },
    "doctoral_school": {
      "type": "string",
      "description": "Doctoral school or graduate program, if explicitly mentioned"
    },
    "defense_year": {
      "type": "integer",
      "description": "Year the graduate thesis was defended. Format yyyy"
    },
    "advisor": {
      "type": "array",
      "items": {
        "type": "string"
      },
      "description": "List of ONLY persons explicitly identified as graduate thesis advisor/supervisor/director. Empty array if the role is not explicitly stated."
    },
    "jury_president": {
      "type": "string",
      "description": "ONLY the person explicitly identified as president/chair of the jury. Null if not explicitly stated."
    },
    "reviewers": {
      "type": "array",
      "items": {
        "type": "string"
      },
      "description": "List of ONLY official reviewers/rapporteurs explicitly identified as such. Empty array if not explicitly stated."
    },
    "committee_members": {
      "type": "array",
      "items": {
        "type": "string"
      },
      "description": "List of persons explicitly listed as jury or committee members whose role is not explicitly identified as president/chair or reviewer/rapporteur. Include unlabeled persons here. Empty array if none."
    },
    "language": {
      "type": "string",
      "description": "Language in ISO 639-3 codes. Example: fre, eng, ita..."
    }
  }
}

Instructions

• Return a JSON object matching the schema
• Use the correct type for each field (string, number, array)
• Do not invent missing data. Multivalued fields are always JSON arrays; use [] when empty. Scalar fields use null when absent.
• If source text is written in all-caps, transform it into sentence case with French accents.
• For person names, apply sentence case (e.g. Jean-Daniel DUBOIS → Jean-Daniel Dubois) and remove all civility and title prefixes (Monsieur, Madame, M., Mme, Professeur, Pr, Dr, etc.).
• A role is explicit only when a textual role label is directly associated with the person. Do NOT infer roles from position, ordering, typography, academic conventions, or prior knowledge. If people are listed without explicit role labels, place them ONLY in committee_members.

</details>