RCaz/eu-funding-proposals-dpo

EU Funding Proposals — DPO Preference Dataset

A proposal-writing preference dataset (DPO) derived from the CORDIS database of EU-funded research projects.

It trains a model to prefer high-quality, funded-style proposal text over degraded alternatives, refining writing quality after supervised fine-tuning.

Format

Conversational preference format compatible with TRL DPOTrainer:

{
  "prompt":   [{"role": "user", "content": "Draft the 'Objectives and Scope' section ..."}],
  "chosen":   [{"role": "assistant", "content": "**Objectives**\n\n<real funded project text>"}],
  "rejected": [{"role": "assistant", "content": "<degraded version>"}]
}

Splits

How it was built

1. 1.Started from RCaz/eu-funding-cordis-qa and filtered to examples with a substantial **Objective:** block (≥250 chars) — genuine, funded, proposal-quality text.
2. 2.Chosen = the real funded project objective, lightly formatted as a proposal section.
3. 3.Rejected = a degraded version of the same text, produced by one of four degraders:
4. 4.structure removal — strips headers/bullets, flattens to a plain paragraph
5. 5.vague terminology — replaces precise EU terms with generic words (e.g. dissemination → sharing results, innovation → new idea)
6. 6.incomplete — truncates to ~1/3, simulating an unfinished draft
7. 7.generic LLM filler — replaces specifics with bland boilerplate

This teaches the model to favor structured, EU-terminology-rich, complete, specific proposal text over generic or poorly structured output.

Intended use

Run after SFT on RCaz/eu-funding-proposals-sft. Use the SFT model as the DPO reference policy (e.g. RCaz/Qwen2.5-7B-EU-Funding-Expert).

Recommended starting hyperparameters: beta=0.1, lr=1e-6, 1–2 epochs, LoRA r=64.

Source

Built from the European Commission's CORDIS open data.

Companion dataset

RCaz/eu-funding-proposals-sft — the SFT version of the same data.