MiniMax-H3-Reason: Evaluation Data

This directory contains the local preparation scaffold for the benchmark release. Actual evaluation inputs and prompts have not yet been populated here. Files under templates/ are organization templates, not benchmark samples.

Input modalities

Every instance includes a text prompt.

Scenario	Media inputs
MSR	One image
ADR	One image and one audio clip
VDR	One video
AVIR	One audio clip and one video

Each instance lives in cases/<SCENARIO>/<ID>/, with a case.json containing its ID, subcategory, operation, prompt, and media paths. Media paths are relative to this dataset root. Missing modalities are null.

File structure

dataset/
├── README.md                     # Dataset description and usage
├── test.jsonl                    # Generated index: one case per line
├── cases/
│   ├── MSR/MSR-001/
│   │   ├── case.json             # Metadata, original prompt, and input paths
│   │   └── image.png
│   ├── ADR/ADR-001/
│   │   ├── case.json
│   │   ├── image.png
│   │   └── audio.wav
│   ├── VDR/VDR-001/
│   │   ├── case.json
│   │   └── video.mp4
│   └── AVIR/AVIR-001/
│       ├── case.json
│       ├── audio.wav
│       └── video.mp4
├── templates/
│   ├── MSR.case.json
│   ├── ADR.case.json
│   ├── VDR.case.json
│   └── AVIR.case.json
└── build_index.py                # Validate cases and generate the index

The case directories above illustrate the naming convention; they are not additional released samples. test.jsonl is generated only after real cases have been populated. Preserve the media formats used in the original evaluation; the extensions above are examples.

Each case.json contains schema_version, id, scenario, subcategory, operation, prompt, and inputs. Store the exact original prompt in prompt. The inputs object always has image, audio, and video keys, using dataset-root-relative paths for required modalities and null for the others. Keep case IDs stable across releases.

Build the index

After populating real cases, run:

python3 build_index.py

The script validates required fields, scenario-specific modalities, unique IDs, and local media paths before writing test.jsonl. It refuses to generate an empty index. The per-case files are the source of truth; regenerate the index after editing them.

Read the index

After downloading the complete dataset, including the cases/ directory:

import json
from pathlib import Path

root = Path("/path/to/MiniMax-H3-Reason")
with (root / "test.jsonl").open(encoding="utf-8") as stream:
    for line in stream:
        case = json.loads(line)
        prompt = case["prompt"]
        media = {
            modality: root / relative_path
            for modality, relative_path in case["inputs"].items()
            if relative_path is not None
        }
        # Pass prompt and media to your model's own generation interface.

operation distinguishes generation, continuation, and editing. This dataset structure contains test inputs and prompts. Model-generated outputs and evaluation results belong in a separate results release, linked by case ID.

The expected full benchmark contains 517 instances: MSR 200, ADR 146, VDR 100, and AVIR 71. This preparation scaffold does not yet contain those instances. Final license, provenance, release scope, and media preprocessing details must be documented with the actual release.