sora 13274243a0 Bump vendored EvalScope and add K3-ready DPV4 configs.
Keep K3 suite selection and report-schema scoring in bash, merge K3/vision dataset_args into dpv4 yamls, and pin EvalScope at 735d920ee911 with local patches.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-02 07:30:48 +00:00

4.8 KiB

VTCBench

Overview

VTCBench (Vision-Text Compression Benchmark) evaluates long-context understanding when text is represented as rendered images, and compares it with a pure-text baseline.

Task Description

  • Task Type: Long-context question answering with image-based and text-based evaluation modes
  • Input: Rendered context images plus a question (VTC mode), or the source text plus a question (Text mode)
  • Output: Short free-form answer
  • Domain: Retrieval, associative reasoning, and long-term dialogue memory

Key Features

  • Provides matched VTC and Text modes for measuring the effect of vision-text compression
  • Includes Retrieval, Reasoning, and Memory subsets derived from RULER, NoLiMa, and LoCoMo
  • Uses pre-rendered multi-image documents to preserve the benchmark's visual layouts
  • Supports contexts spanning multiple document images

Evaluation Notes

  • Default configuration uses 0-shot evaluation in VTC mode
  • Use --dataset-args '{"vtcbench": {"extra_params":{"eval_mode":"text"}}}' to enable the Text baseline
  • Retrieval and Reasoning use the official fractional contains_all score
  • Memory uses the official maximum ROUGE-L F1 across reference answers
  • The unified score metric dispatches to the official metric for each subset; its report macro_score is the unweighted mean across the three tasks
  • Text mode strips HTML tags and normalizes whitespace in the same way as the official static evaluator
  • Content inside <think>...</think> is excluded before scoring, matching the official evaluator
  • Long-context requests may require a larger model timeout
  • If dataset casting reports an offset overflow, set DATASET_TF_BATCH_SIZE=1
  • Paper | Code

Properties

Property Value
Benchmark Name vtcbench
Dataset ID MLLM-CL/VTCBench
Paper Paper
Tags LongContext, MultiModal, QA, Reasoning, Retrieval
Metrics score, contains_all, rouge_l
Default Shots 0-shot
Evaluation Split test

Data Statistics

Metric Value
Total Samples 2,200
Prompt Length (Mean) 236.71 chars
Prompt Length (Min/Max) 89 / 384 chars

Per-Subset Statistics:

Subset Samples Prompt Mean Prompt Min Prompt Max
Retrieval 800 110.38 89 141
Reasoning 800 368.69 363 384
Memory 600 229.15 186 283

Image Statistics:

Metric Value
Total Images 26,554
Images per Sample min: 1, max: 62, mean: 12.07
Resolution Range 896x896 - 896x896
Formats jpeg

Sample Example

Subset: Retrieval

{
  "input": [
    {
      "id": "c51f44e8",
      "content": [
        {
          "image": "[BASE64_IMAGE: jpeg, ~367.1KB]"
        },
        {
          "image": "[BASE64_IMAGE: jpeg, ~366.1KB]"
        },
        {
          "image": "[BASE64_IMAGE: jpeg, ~385.2KB]"
        },
        {
          "image": "[BASE64_IMAGE: jpeg, ~377.7KB]"
        },
        {
          "image": "[BASE64_IMAGE: jpeg, ~333.5KB]"
        },
        {
          "text": "\n\nQuestion:What are all the special magic numbers for foolish-rawhide mentioned in the provided text?"
        }
      ]
    }
  ],
  "target": "4075987, 5943250",
  "id": 0,
  "group_id": 0,
  "metadata": {
    "problem": "What are all the special magic numbers for foolish-rawhide mentioned in the provided text?",
    "answers": [
      "4075987",
      "5943250"
    ],
    "subset": "Retrieval",
    "eval_mode": "vtc"
  }
}

Prompt Template

No prompt template defined.

Extra Parameters

Parameter Type Default Description
eval_mode str vtc Evaluation mode: vtc (images+problem) or text (text+problem). Choices: ['vtc', 'text']

Usage

Using CLI

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets vtcbench \
    --limit 10  # Remove this line for formal evaluation

Using Python

from evalscope import run_task
from evalscope.config import TaskConfig

task_cfg = TaskConfig(
    model='YOUR_MODEL',
    api_url='OPENAI_API_COMPAT_URL',
    api_key='EMPTY_TOKEN',
    datasets=['vtcbench'],
    dataset_args={
        'vtcbench': {
            # subset_list: ['Retrieval', 'Reasoning', 'Memory']  # optional, evaluate specific subsets
            # extra_params: {}  # uses default extra parameters
        }
    },
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)