crystal-metrics 0.2.0

Transparent multimodal reasoning metrics from the CRYSTAL benchmark (Match F1, Ordered Match F1, accuracy).

crystal-metrics

Transparent multimodal reasoning metrics from the CRYSTAL benchmark.

Your model gets the right answer. But does it actually reason? Standard benchmarks only check the final answer, so a lucky guess scores the same as sound reasoning. crystal-metrics scores the reasoning chain itself — step-level precision/recall, ordering, and answer accuracy.

What it measures

Metric	Measures
Match F1	Step-level F1 of predicted vs. reference reasoning steps via semantic-similarity matching
Precision	Fraction of predicted steps that match a reference step (few wrong things)
Recall	Fraction of reference steps that were covered (completeness — the hard part)
Ordered Match F1	Match F1 penalized for out-of-order reasoning (Kendall's τ or LIS ratio)
Accuracy	Multi-format final-answer correctness (yes/no, numeric, multiple choice, free text)

It also ships the RL reward functions used to train models on CRYSTAL — Causal Process Reward (CPR) and Semantic Process Reward (SPR) — in crystal_metrics.rewards (pure Python, model-agnostic). See the rewards docs.

Install

pip install crystal-metrics          # core metrics (no LLM required)
pip install crystal-metrics[judge]   # + optional LLM judge for free-form answers

Requires Python 3.8+. The default embedding model all-distilroberta-v1 is downloaded and cached on first use.

Quickstart

from crystal_metrics import MLLMReasoningEvaluator

evaluator = MLLMReasoningEvaluator()  # all-distilroberta-v1, threshold τ=0.35 (paper defaults)

m = evaluator.evaluate_single(
    predicted_steps=[
        "Three objects sit on the table",
        "The middle console is the smallest",
        "Therefore the answer is C",
    ],
    reference_steps=[
        "There are three objects in the image",
        "Compare the sizes of the three objects",
        "The middle object is smallest",
        "Select option C",
    ],
    alpha=0.3,  # enable Ordered Match F1 (0 = order-insensitive)
)

print(f"Match F1:         {m.match_f1:.3f}")
print(f"Precision:        {m.precision:.3f}")
print(f"Recall:           {m.recall:.3f}")
print(f"Ordered Match F1: {m.ordered_match_f1:.3f}")

Answer accuracy

from crystal_metrics import AccuracyCalculator

calc = AccuracyCalculator(use_llm_grader=False)          # rule-based, no LLM
acc = calc.evaluate_dataset(predictions, references)
print(acc["overall_accuracy"], acc["type_statistics"])

The optional LLM judge (for free-form text) needs the [judge] extra and any OpenAI-compatible endpoint (e.g. a local Ollama server):

calc = AccuracyCalculator(use_llm_grader=True, llm_model="gpt-oss:120b",
                          base_url="http://localhost:11434/v1")

Command line

crystal-metrics evaluate predictions.json references.json --alpha 0.3

=== CRYSTAL metrics ===
  samples           : 3
  match_f1          : 0.5524
  precision         : 0.6667
  recall            : 0.4722
  ordered_match_f1  : 0.4952
  accuracy          : 0.6667

Data format

// predictions
{"<id>": {"question": "...", "reasoning_steps": ["..."], "answer": "..."}}
// references
{"<id>": {"reference_steps": ["..."], "answer": "..."}}

Paper defaults

Setting	Value	Source
Embedding model	all-distilroberta-v1	Paper §4.3
Similarity threshold τ	0.35	Ablation-validated
Recommended alpha	0.3	Paper
Numeric tolerance	ε_abs = 0.05, ε_rel = 0.10	Paper Eq. (2)

Documentation

• Installation
• Quickstart
• Metric definitions
• CLI reference
• RL reward functions (CPR / SPR)

Benchmark: 🤗 waybarrios/CRYSTAL · Project: github.com/waybarrios/crystal

Citation

@misc{barrios2026crystal,
  title   = {Beyond Final Answers: CRYSTAL Benchmark for Transparent
             Multimodal Reasoning Evaluation},
  author  = {Wayner Barrios and SouYoung Jin},
  year    = {2026},
  eprint  = {2603.13099},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url     = {https://arxiv.org/abs/2603.13099}
}

License

MIT — see the project repository.