RCaz/eu-funding-proposals-sft

EU Funding Proposals — SFT Dataset

A proposal-writing supervised fine-tuning (SFT) dataset derived from the CORDIS database of EU-funded research projects.

Unlike a factual Q&A dataset, this dataset trains a model to write proposal sections in the style of actual funded EU projects.

Format

Conversational ChatML (messages column) compatible with TRL SFTTrainer:

{
  "messages": [
    {"role": "system", "content": "You are an expert grant writer..."},
    {"role": "user", "content": "Draft the 'Objectives and Scope' section of a Horizon Europe proposal for the project '...'."},
    {"role": "assistant", "content": "**Objectives**\n\n..."}
  ]
}

Splits

Split	Examples
train	106,189
validation	5,618

How it was built

1. Started from RCaz/eu-funding-cordis-qa (413K factual Q&A pairs about CORDIS projects).
2. Filtered to examples containing a substantial **Objective:** block (≥250 chars) — these contain genuine, funded, proposal-quality text. ~25.6% of source examples qualified.
3. Each kept example was reframed as a proposal-writing task: a prompt asking the model to draft one of five proposal sections (abstract, objectives, concept, innovation, impact), with the real funded objective text as the gold-standard completion.
4. A grant-writer system prompt was added to set the writing persona.

Source

Built from the European Commission's CORDIS open data. See RCaz/eu-funding-cordis-qa.

Companion dataset

RCaz/eu-funding-proposals-dpo — the preference (DPO) version of the same data, for refining proposal-writing quality after SFT.