Graphgen Eval

graphgen-eval

GraphGen: Enhancing Supervised Fine-Tuning for LLMs with Knowledge-Driven Synthetic Data Generation — Chen et al. (2025) (arXiv:2505.20416, 2025)

What this evaluates

Evaluates the ability of LLMs to answer knowledge-intensive questions across atomic, aggregated, and multi-hop reasoning scenarios in agricultural, medical, and general domains. It measures how well supervised fine-tuning with synthetic knowledge-graph data improves closed-book QA performance.

Datasets

• SeedEval — total 582; splits: test (582)
• PQArefEval — total 5815; splits: test (5815)
• HotpotEval — total 7405; splits: test (7405)

Metrics

• ROUGE-F (primary) — range: [0, 1]
  • F1 score measuring the overlap of n-grams between the generated answer and the ground truth reference.

Input / output format

Input: A natural language question for closed-book QA.

Output: A natural language text answer.

Scoring recipe

def evaluate(predictions, references):
    scores = []
    for pred, ref in zip(predictions, references):
        score = compute_rouge_f1(pred, ref)  # ROUGE-F as specified in the paper
        scores.append(score)
    return sum(scores) / len(scores)

Common pitfalls

• The evaluation is strictly closed-book; allowing external retrieval during inference invalidates the benchmark.
• ROUGE-F measures lexical overlap, not factual correctness or reasoning depth, so high scores do not guarantee accurate multi-hop answers.
• Baseline methods can cause performance degradation on aggregated QA tasks due to single-segment generation limits, which should be attributed to data synthesis flaws rather than the evaluation metric itself.

Evidence (verbatim from paper)

We use ROUGE-F as the metric. The baseline methods exhibit varying performance across the three datasets, while GraphGen consistently achieves optimal results.

Citation

@misc{chen2025graphgen,
  title={GraphGen: Enhancing Supervised Fine-Tuning for LLMs with Knowledge-Driven Synthetic Data Generation},
  author={Chen et al. (2025)},
  year={2025},
  note={arXiv:2505.20416}
}

• arXiv: 2505.20416