gneubig/materials-figure-qa

Materials Figure QA

This repository contains 250 figure-grounded multimodal question-answer pairs from recent materials-related arXiv papers. Each item requires interpreting a rendered figure and applying materials-science reasoning. Figures with aspect ratio greater than 4:1 were removed; retained figures were downsampled so their longest dimension is at most 2048 pixels.

The filtered split preserves the original deterministic split:

• validation: 128 examples
• test: 122 examples

Each split contains a typed Parquet file (validation.parquet or test.parquet) for the Hugging Face viewer, plus an images/ directory. The Parquet schema exposes image, question, answer, caption, reasoning, difficulty, and provenance columns. The JSONL image_path values are relative to the split directory.

To load a split with the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("gneubig/materials-figure-qa", data_files={"validation": "validation.parquet", "test": "test.parquet"})
print(dataset["validation"].features)

Run inference with an arbitrary LLM

infer_materials_qa.py uses an OpenAI-compatible chat-completions endpoint. It sends each question with its figure as a base64 image data URL and writes JSONL predictions. It can be used with OpenAI-compatible hosted services or a local server such as vLLM, LM Studio, or Ollama's OpenAI-compatible endpoint.

export OPENAI_API_KEY=...
python infer_materials_qa.py \
  --data /path/to/materials-figure-qa/validation \
  --output predictions/validation.jsonl \
  --model YOUR_VISION_MODEL \
  --base-url https://api.openai.com/v1

For a local endpoint:

export LOCAL_API_KEY=none
python infer_materials_qa.py \
  --data ./validation \
  --output predictions/validation.jsonl \
  --model Qwen/Qwen2.5-VL-7B-Instruct \
  --base-url http://localhost:8000/v1 \
  --api-key-env LOCAL_API_KEY

Semantic evaluation results

An independent multimodal judge scored retained predictions on a 0–3 scale: 3 fully correct, 2 mostly correct, 1 partially correct, and 0 incorrect. These scores reuse cached answers and verdicts after removing the five filtered test items.

Model	Dev strict	Dev 2+	Test strict	Test 2+
GPT-5.5	68.8%	79.7%	59.3%	73.1%
Gemini 3.5 Flash	33.9%	48.0%	42.0%	51.3%
Claude Sonnet 4.6	28.3%	51.2%	26.8%	54.5%
Claude Opus 4.6	24.8%	52.0%	16.2%	48.6%
Gemini 2.5 Pro	5.5%	11.7%	5.1%	14.5%

strict counts only score 3; 2+ counts scores 2 or 3. The verifier had six errors overall; denominators above use scored retained predictions.

Run both splits with the same model:

python infer_materials_qa.py --data ./validation  --output predictions/validation.jsonl  --model YOUR_MODEL --base-url YOUR_BASE_URL
python infer_materials_qa.py --data ./test --output predictions/test.jsonl --model YOUR_MODEL --base-url YOUR_BASE_URL

The inference script is resumable: existing candidate_id values in the output are skipped.

Calculate scores

score_materials_qa.py provides a deterministic lexical baseline. It reports:

• exact_match: normalized string equality
• token_f1: token-overlap F1 between prediction and reference answer

The deterministic scorer requires predictions whose candidate_id values match the split records. It reports only available predictions; missing calls reduce availability and are not counted as correct. Exact match is intentionally strict and is usually inappropriate for free-form scientific answers, so token F1 is included as a lexical baseline. The included historical evaluation files are not part of the uploaded dataset and should not be interpreted as complete dev/test benchmark results.

• availability: fraction of split items with a successful prediction

python score_materials_qa.py \
  --data-root /path/to/materials-figure-qa \
  --predictions predictions/all.jsonl \
  --output scores.json

Predictions can be combined into one JSONL file, or the command can be run separately for each split. Prediction records must contain candidate_id, model, and answer.

Interpreting the scores

The deterministic metrics are useful for regression testing and exact-answer tasks, but they are not a complete measure of scientific correctness. Many answers have valid paraphrases and require visual and domain reasoning. For a stronger benchmark, supplement these metrics with expert review or an independent multimodal judge. Report API availability separately from answer quality.

Data provenance and qualification

The examples were extracted from LaTeX figure-label references. The current selection retained examples whose automated verification marked the question and reference answer valid, and where at least two of five evaluated multimodal models received a score of 0 or 1. This automated filtering is intended to identify difficult examples for expert review; it should not be treated as a substitute for expert gold-label validation.