Geraldine/mlintern-gemma4-unimarc-training

📚 Gemma-4 UNIMARC Training Dataset

Training dataset for fine-tuning google/gemma-4-E2B-it to extract bibliographic metadata from book/thesis cover images and generate valid UNIMARC records.

Overview

This dataset combines real bibliographic data from two sources with synthetic augmentation to train a Vision-Language Model (VLM) for automated UNIMARC cataloguing.

Source	Samples	Type	Description
vojtam/BiblioPage	1,708	Books (title pages)	Czech library book title pages with structured metadata
Geraldine/humatheque-vlm-sudoc-grounded	79	Theses	French doctoral thesis covers with SUDOC records
Synthetic (books)	500	Books	Generated UNIMARC records for French-language books
Synthetic (theses)	500	Theses	Generated UNIMARC records for French doctoral theses
Total	2,787		Train: 2,508 / Test: 279

⚠️ Important: UNIMARC 214 (NOT 210)

This dataset uses the UNIMARC 214 field for publication, production, distribution and manufacture data. The former field 210 is deprecated and must NOT be used in new records.

Field 214 uses the first indicator to distinguish the function:

• ind1=0: Production
• ind1=1: Publication (most common — publisher, city, date)
• ind1=2: Distribution
• ind1=3: Manufacture (printer)
• ind1=4: Copyright notice date

See ABES documentation on field 214.

UNIMARC Fields Covered

Field	Label	Subfields	Source
010	ISBN	$a	Books (synthetic)
101	Language	$a (ISO 639-2)	All
102	Country of publication	$a (ISO 3166)	All
200	Title & responsibility	$a title, $e subtitle, $f author, $g co-author, $h part no., $i part name	All
205	Edition statement	$a	Books
214	Publication/Distribution	$a place, $b publisher, $d date	All (replaces 210)
215	Physical description	$a extent, $c illustrations, $d dimensions	All
225	Series	$a name, $v number	Books
300	General note	$a	Some
328	Thesis note	$a (degree : discipline : institution : year)	Theses
700	Primary author	$a surname, $b forename, $4 role code	All
701	Co-authors	$a, $b, $4	Some
702	Secondary responsibility	$a, $b, $4 (illustrator=310, translator=730, director=651)	Some
710	Corporate author	$a, $4 (727=thesis institution)	Theses

Dataset Format

Each sample contains conversational prompt/completion pairs (JSON-encoded) designed for TRL's SFTTrainer:

{
    "prompt": [  # JSON string
        {"role": "system", "content": "Tu es un catalogueur expert en format UNIMARC..."},
        {"role": "user", "content": [
            {"type": "image"},
            {"type": "text", "text": "Analyse cette image de couverture et génère la notice UNIMARC..."}
        ]}
    ],
    "completion": [  # JSON string
        {"role": "assistant", "content": [
            {"type": "text", "text": "{\"200\": [{\"ind1\": \"1\", \"$a\": \"Title\", ...}], \"214\": [...]}"}
        ]}
    ],
    "source": "vojtam/BiblioPage",  # origin
    "doc_type": "book",              # book or these
    "image_uri": "https://..."       # image URL (null for synthetic)
}

UNIMARC Output Example (Book)

{
  "010": [{"$a": "978-2-07-036822-8"}],
  "101": [{"$a": "fre"}],
  "102": [{"$a": "FR"}],
  "200": [{"ind1": "1", "$a": "Introduction à la sociologie", "$e": "théorie et applications", "$f": "Jean Dupont"}],
  "214": [
    {"ind1": "1", "$a": "Paris", "$b": "Gallimard", "$d": "2023"},
    {"ind1": "2", "$a": "Paris", "$b": "Hachette Livre Distribution"}
  ],
  "215": [{"$a": "248 p.", "$d": "21 cm"}],
  "225": [{"$a": "Collection U", "$v": "42"}],
  "700": [{"$a": "Dupont", "$b": "Jean", "$4": "070"}]
}

UNIMARC Output Example (Thesis)

{
  "101": [{"$a": "fre"}],
  "102": [{"$a": "FR"}],
  "200": [{"ind1": "1", "$a": "Contribution à l'étude de la migration transnationale en Europe", "$e": "une analyse empirique", "$f": "Marie Laurent"}],
  "214": [{"ind1": "1", "$a": "[Paris]", "$b": "Sorbonne Université", "$d": "2024"}],
  "215": [{"$a": "420 p.", "$d": "30 cm"}],
  "328": [{"$a": "Thèse de doctorat : Sociologie : Sorbonne Université : 2024"}],
  "700": [{"$a": "Laurent", "$b": "Marie", "$4": "070"}],
  "702": [{"$a": "Martin", "$b": "Pierre", "$4": "651"}],
  "710": [{"$a": "Sorbonne Université", "$4": "727"}]
}

Training Script

The repository includes train.py for fine-tuning google/gemma-4-E2B-it:

# Basic training (LoRA, bf16, A100 80GB)
python train.py

# QLoRA for smaller GPUs (A10G 24GB)
python train.py --use_qlora

# Custom configuration
python train.py \
  --num_train_epochs 5 \
  --learning_rate 2e-4 \
  --lora_r 32 \
  --per_device_train_batch_size 2 \
  --gradient_accumulation_steps 4 \
  --push_to_hub_model_id my-org/my-unimarc-model

Training Details

Parameter	Default	Notes
Method	LoRA (SFT)	LoRA on LLM attention + MLP; vision encoder frozen
Learning rate	1e-4	Higher than base SFT (2e-5) due to LoRA
Batch size	1 × 8 grad accum	Effective batch = 8
Epochs	3
LoRA rank	16	Target: q/k/v/o_proj + gate/up/down_proj
Precision	bf16	Native for Gemma 4
Packing	Disabled	Mandatory for VLM (image tokens)
Max length	None	No truncation to preserve image tokens
Loss	Assistant-only	Loss computed on UNIMARC output only

Dependencies

transformers>=5.5.0
trl>=1.2.0
peft>=0.15.0
datasets>=4.0.0
torch>=2.4.0
accelerate>=1.0.0
bitsandbytes>=0.45.0  # only for QLoRA
trackio                # optional, for monitoring

Data Sources & Licensing

• vojtam/BiblioPage: From BiblioPage: A Dataset for Bibliographic Information Extraction. Czech library book title pages with CC license.
• Geraldine/humatheque-vlm-sudoc-grounded: French doctoral thesis covers from SUDOC/ABES.
• Synthetic data: Generated following UNIMARC standard (IFLA) and ABES/SUDOC conventions.

References

• UNIMARC Manual (IFLA)
• ABES SUDOC UNIMARC Documentation
• ABES 214 Field Documentation
• BiblioPage Paper (arXiv:2503.19658)
• TRL SFT Trainer Documentation
• Gemma 4 Model Card

Citation

@misc{mlintern-gemma4-unimarc,
  title={Gemma-4 UNIMARC Training Dataset for Bibliographic Record Generation},
  author={Geraldine},
  year={2026},
  howpublished={\\url{https://huggingface.co/datasets/Geraldine/mlintern-gemma4-unimarc-training}},
}