README.md

metadata

dataset_info:
  features:
    - name: messages
      list:
        - name: content
          dtype: string
        - name: role
          dtype: string
    - name: source
      dtype: string
    - name: split
      dtype: string
    - name: mode
      dtype: string
    - name: category
      dtype: string
    - name: file_id
      dtype: string
    - name: ref_count
      dtype: int64
    - name: language
      dtype: string
  splits:
    - name: train
      num_bytes: 3012211
      num_examples: 1708
    - name: validation
      num_bytes: 199719
      num_examples: 115
  download_size: 887241
  dataset_size: 3211930
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*

Reference Parsing Finetuning Dataset

A fine-tuning dataset for bibliographic reference extraction and parsing, combining LinkedBooks, CEX, and EXCITE datasets into conversation-style examples for LLM SFT.

Dataset Description

This dataset teaches models to extract and parse bibliographic references from text into structured JSON format. Examples follow a conversational format with system/user/assistant messages, using various prompt variants for diversity.

Data Sources:

• LinkedBooks: Multi-language reference strings with structured metadata
• CEX: English academic papers with TEI XML parsed references
• EXCITE: Multi-language academic papers with parsed references

Data Fields

Field	Type	Description
messages	list	Conversation messages with role (system/user/assistant) and content
source	string	Data source: linkedbook, cex, or excite
split	string	Dataset split: train or valid
mode	string	Example type: single (1 reference) or group (multiple references)
language	string	Language code (e.g., en, de, fr)
ref_count	int	Number of references in the example
file_id	string|null	Source document ID (for CEX/EXCITE)
category	string|null	Document category (for CEX)

Splits

Split	Examples	Description
train	~1,708	Main training data
valid	~115	Validation set

Distribution:

• ~70% single-reference examples, ~30% multi-reference groups
• ~10-15% LinkedBook, ~30-35% CEX, ~50-55% EXCITE

Data Creation and Processing

1. Data Loading: Loads references from LinkedBooks (Training and Validation JSONL), CEX (JSON + TEI XML), and EXCITE (JSON + XML)
2. Validation: Filters invalid references (missing titles/authors, unparsed authors, mismatched counts)
3. Sampling: Stratified sampling by category/class (30% train rate for CEX/EXCITE)
4. Grouping: Groups references into batches (3-20 refs per group with weighted probabilities)
5. Prompt Variants: Applies 5 prompt variants with weighted distribution (40% detailed, 25% minimal, 25% task-based, 5% ultra-minimal, 5% no prompt)
6. Format Conversion: Converts to conversation-style format with structured JSON output

Credits

The dataset is being developed by Yurui Zhu (Odoma). This work is carried out in the context of the EU-funded GRAPHIA project (grant ID: 101188018).