cometadata/funding-extraction-artifact-data-mix-grpo-mixed-reward

Funding Extraction Training Data

Training, evaluation, and test data for fine-tuning LLMs to extract structured funding metadata (funder names, award IDs, funding schemes, award titles) from academic paper funding statements.

Dataset Structure

data/
├── full/                     # Complete unsplit dataset
│   ├── train.jsonl           # 5,264 real Crossref funding statements
│   └── synthetic.jsonl       # 10,124 LLM-generated funding statements
├── sft/                      # Supervised fine-tuning split
│   ├── train.jsonl           # 3,528 real examples
│   └── synthetic.jsonl       # 7,240 synthetic examples
├── rl/                       # Reinforcement learning split
│   ├── train.jsonl           # 1,160 real examples (RL training)
│   ├── synthetic.jsonl       # 1,916 synthetic examples (RL training)
│   ├── eval_train.jsonl      # 576 real examples (RL evaluation)
│   └── eval_synthetic.jsonl  # 968 synthetic examples (RL evaluation)
└── test/
    └── arxiv_test.jsonl      # 300 held-out arXiv examples (never seen during training)

Full dataset (data/full/)

The complete corpus before splitting. Use this if you want to create your own splits.

File	Examples	Source
train.jsonl	5,264	Real Crossref funding statements with gold-standard labels
synthetic.jsonl	10,124	LLM-generated funding statements with controlled labels

Pre-split for SFT + RL pipeline

The full dataset is pre-split into non-overlapping subsets for a two-stage training pipeline (SFT then GRPO RL):

Directory	File	Examples	Pipeline stage
data/sft/	train.jsonl	3,528	SFT training (real)
data/sft/	synthetic.jsonl	7,240	SFT training (synthetic)
data/rl/	train.jsonl	1,160	RL training (real)
data/rl/	synthetic.jsonl	1,916	RL training (synthetic)
data/rl/	eval_train.jsonl	576	RL evaluation (real)
data/rl/	eval_synthetic.jsonl	968	RL evaluation (synthetic)

Held-out test (data/test/)

File	Examples	Description
arxiv_test.jsonl	300	Held-out arXiv funding statements, never seen during training

Data Augmentation

50% of training examples include synthetic noise augmentation for robustness to real-world document formats:

Augmentation	Frequency	Example
OCR-like case noise	97%	researCh for research
Digit/letter swaps	74%	Character-level OCR errors
Unicode artifacts	64%	Zero-width joiners, non-breaking spaces
XML/HTML tags	36%	<ext-link>, <bold>, <xref>
LaTeX markup	32%	\textbf{}, \href{}

The augmentation field in each record indicates whether and how it was augmented. The original field contains the clean text before augmentation.

Schema

Each JSONL record contains:

{
  "id": "unique-record-id",
  "doi": "10.xxxx/...",
  "funding_statement": "The augmented (or original) funding text to extract from",
  "original": "The clean funding text before augmentation",
  "funders": [
    {
      "funder_name": "National Science Foundation",
      "awards": [
        {
          "award_ids": ["DMS-1613002"],
          "funding_scheme": ["Standard Grant"],
          "award_title": ["Some Project Title"]
        }
      ]
    }
  ],
  "base_type": "real|synthetic",
  "category": "category label",
  "augmentation": "description of augmentation applied or null"
}

arxiv_test.jsonl uses a slightly different schema (doi, funding_statement, funders, markdown) without the augmentation fields.

Intended Use

This dataset is designed for training and evaluating models on the task of structured funding metadata extraction from academic papers. The hierarchical label structure (funder -> awards -> award_ids/scheme/title) supports both flat and hierarchical evaluation metrics.

Associated Model

• adambuttrick/funding-extraction-llama-3.1-8b-instruct-grpo-stepfinal — LoRA adapter trained on this data using SFT + GRPO