Despina/re_mixtune

MixTune (Domain-Balanced RE Instruction Mixture)

MixTune is an aggregated instruction-tuning mixture for relation extraction (RE) that balances general-domain and literary data. It draws equally from a general-domain pool and a literary pool, giving a single prompt/label corpus used to fine-tune small language models in the paper "Sub-Billion, Super-Frontier: Fine-Tuned Small Language Models Rival Zero-Shot Frontier LLMs on General and Literary Relation Extraction" (Christou & Tsoumakas, 2026) arXiv:2606.22606.

It is one of the three training regimes in the paper:

• GenTune — general-domain only (Despina/re_gentune)
• LitTune — literary only (Despina/re_littune)
• MixTune — a domain-balanced mix of the two (this dataset)

Component datasets

MixTune is built by pooling each domain, then sampling an equal number of examples from the general and literary pools (domain-balanced, shuffled with seed 42):

General pool

Source	Repo
TACRED	Despina/tacred
SemEval-2010 Task 8	Despina/semeval2010_task8
CoNLL04	Despina/conll04
NYT11	Despina/nyt_11
GIDS	Despina/gids
Re-DocRED	Despina/re-docred
REBEL (subsampled)	Despina/rebel-dataset

Literary pool

Source	Repo
Biographical	Despina/biographical
PG-Fiction	Despina/project-gutenberg-fiction-relations

Because the mix is domain-balanced (equal general vs. literary), the literary datasets are up-weighted relative to their raw size, and the general pool is subsampled down to match.

Dataset structure

Each source example is reshaped into a unified instruction-tuning row. The mixture is materialized once per shot setting (0, 2, 5), and each shot setting has its own train / eval / test split:

Split	Shot
train_0, eval_0, test_0	0-shot prompts
train_2, eval_2, test_2	2-shot prompts
train_5, eval_5, test_5	5-shot prompts

Columns

Column	Description
prompt	The instruction prompt for the example at this shot count. For the 0-shot splits, an "\nAnswer: " cue is appended to the prompt.
relation	The gold relation label (the target completion).
dataset	The source dataset the row came from (e.g. conll04, biographical), so the mixture can be stratified by source or domain.

Pick one shot setting per experiment (e.g. train on train_2 / evaluate on test_2) rather than mixing splits across shot counts.

Usage

from datasets import load_dataset

# a full DatasetDict with all shot splits
mixtune = load_dataset("Despina/re_mixtune")
print(mixtune)  # train_0, eval_0, test_0, train_2, ... test_5

# or a single split
train2 = load_dataset("Despina/re_mixtune", split="train_2")
print(train2[0]["prompt"], "->", train2[0]["relation"])

Licensing and redistribution

MixTune includes TACRED and NYT11 (in its general pool), which are derived from Linguistic Data Consortium (LDC) corpora (LDC newswire / the NYT Annotated Corpus, LDC2008T19). Those components are not freely redistributable: using this mixture requires the appropriate LDC licenses, and this aggregate should not be redistributed publicly as-is. If you need a shareable mixture, rebuild MixTune with src/build_mixtures.py while excluding the LDC-restricted sources.

Each component also retains its own upstream license (CC BY-SA for DocRED/REBEL, MIT for Biographical, public-domain source texts for PG-Fiction, etc.); see the individual dataset cards under the Despina/ namespace.

Citation

If you use this dataset, please cite our paper (and the underlying source datasets you rely on):

@article{christou2026subbillion,
  title        = {Sub-Billion, Super-Frontier: Small Language Models Rival
                  Zero-Shot Frontier LLMs on General and Literary Relation Extraction},
  author       = {Christou, Despina and Tsoumakas, Grigorios},
  journal      = {arXiv preprint arXiv:2606.22606},
  year         = {2026},
  url          = {https://arxiv.org/abs/2606.22606}
}