Despina/frontier-re-generations

Frontier LLM Relation-Extraction Generations

This dataset contains the raw relation-extraction predictions of frontier LLMs (GPT-5.4, Claude Sonnet 4.6, and Gemini 2.5 Pro) on the nine evaluation datasets used 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.

The frontier comparison is routed through a commercial API (OpenRouter) and is therefore not bit-for-bit reproducible, so we release the exact generations behind the paper's frontier baseline tables rather than only the code. Each row is one model's answer for one evaluation example.

Dataset structure

A single merged CSV (all_frontier_generations.csv), one row per (model × evaluation example × generation setting).

Column	Description
eval_dataset_name	Which evaluation dataset the example comes from (e.g. conll04, biographical).
prompt_hash	Truncated MD5 of the input prompt; stable per-example id (also used for dedup/resume).
relation	Gold relation label for the example.
gen_type	gen_open (generic RE system prompt) or gen_constrained (allowed label set injected into the system prompt).
model_id	Frontier model display name: GPT-5.4, Claude Sonnet 4.6, or Gemini 2.5 Pro.
tuned_dataset_name	Always frontier (these models are evaluated zero-/few-shot, not fine-tuned).
model_shot	Always n/a (no fine-tuning shot for frontier models).
prompt_shot	In-context demonstrations in the prompt (0 or 2).
generated_relation	The model's raw output string (the predicted relation; may need normalization before scoring). Rows beginning with ERROR: denote failed API calls.
input_tokens	Prompt tokens reported by the API for this call.
output_tokens	Completion tokens reported by the API for this call.

Scoring in the paper normalizes generated_relation against relation with the metric library in the accompanying code (positive-class micro-F1, no-relation class excluded); see the repo for the exact evaluation.

Usage

import pandas as pd
from datasets import load_dataset

# via pandas
df = pd.read_csv("hf://datasets/Despina/frontier-re-generations/all_frontier_generations.csv")
print(df.groupby(["model_id", "eval_dataset_name"]).size())

# ...or via the datasets library
ds = load_dataset("Despina/frontier-re-generations", split="train")
print(ds[0])

Licensing and terms

These are outputs generated by third-party commercial models (OpenAI GPT-5.4, Anthropic Claude Sonnet 4.6, Google Gemini 2.5 Pro) accessed via OpenRouter. Use of the generations is subject to the respective model providers' terms of service. The gold relation labels and the input examples derive from the underlying evaluation datasets, each of which retains its own license (see the individual dataset cards under the Despina/ namespace; note in particular that TACRED / LDC-derived sources carry redistribution restrictions).

Citation

If you use these generations, please cite our paper:

@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}
}