Evaluator Lm Eval

evaluator-lm-eval

Prometheus: Inducing Fine-grained Evaluation Capability in Language Models — Seungone Kim et al. (2023) (arXiv:2310.08491, 2023)

What this evaluates

This protocol evaluates an LLM's ability to act as a fine-grained text evaluator using custom score rubrics. It tests both absolute grading (assigning a 1–5 score and generating feedback based on a rubric and reference answer) and ranking grading (predicting human preference between two responses).

Datasets

• Feedback Bench — total 2000; splits: test (2000); repo https://github.com/kaistAI/Prometheus
• Vicuna Bench — total 80; splits: test (80)
• MT Bench — total 80; splits: test (80)
• FLASK Eval — total 200; splits: test (200)
• MT Bench Human Judgments — total ?; splits: test (-1)
• HHH Alignment — total 221; splits: test (221)

Metrics

• Pearson correlation (primary) — range: [−1, 1]
  • Measures the linear correlation between the evaluator model's scores and human/GPT-4 reference scores across all instances.
• Kendall-Tau correlation — range: [−1, 1]
  • Measures the ordinal association between predicted and reference scores by counting concordant and discordant pairs.
• Spearman correlation — range: [−1, 1]
  • A rank-based measure of statistical dependence between the evaluator's scores and reference scores.
• accuracy — range: [0, 1]
  • The fraction of response pairs where the evaluator correctly predicts the human-preferred response.

Input / output format

Input: Instruction, response(s) to evaluate, custom score rubric, and reference answer(s). For ranking tasks, two candidate responses are provided alongside the instruction and rubric.

Output: A textual feedback critique and a numerical score on a 1–5 scale. For ranking evaluation, the model scores each candidate independently.

Scoring recipe

def score_absolute(model_scores, ref_scores):
    return pearsonr(model_scores, ref_scores)

def score_ranking(model, pairs):
    correct = 0
    for pair in pairs:
        s1 = model.score(pair.r1, temp=1.0)
        s2 = model.score(pair.r2, temp=1.0)
        while s1 == s2:
            s1 = model.score(pair.r1, temp=1.0)
            s2 = model.score(pair.r2, temp=1.0)
        if (s1 > s2) == (pair.human_pref == 1):
            correct += 1
    return correct / len(pairs)

Common pitfalls

• Ranking evaluation uses temperature 1.0 and iterative independent scoring to break ties, which the authors explicitly note is not a fair comparison to standard paired ranking models.
• Reference answers for several benchmarks (Vicuna, MT Bench) are synthetically generated by GPT-4 rather than human-written, which may bias the evaluation.
• Absolute grading requires the model to score without an opponent, making it inherently harder than ranking but more practical for real-world use.

Evidence (verbatim from paper)

For the experiments that measure the correlation, we use 3 different correlation metrics: Pearson, Kdendall-Tau, and Spearman. For measuring the quality of the generated feedback, we conduct a pairwise comparison between the feedback generated by Prometheus, GPT-3.5-Turbo, and GPT-4, asking human evaluators to choose which has better quality and why they thought so.

Citation

@misc{kim2023prometheus,
  title={Prometheus: Inducing Fine-grained Evaluation Capability in Language Models},
  author={Seungone Kim et al. (2023)},
  year={2023},
  note={arXiv:2310.08491}
}

• arXiv: 2310.08491