Hinglish-IndicVisionBench-Benchmark

We contribute a Roman Hinglish extension of 98 matched IndicVisionBench image questions, aligned with their English and Hindi counterparts, plus the recorded Qwen3-VL-4B-Instruct evaluation across all three languages.

What is included

• 98 shared images and aligned item IDs, so each question can be compared directly across English, Hindi and Roman Hinglish.
• Three language splits: hinglish, hindi, and english, with 98 items each.
• Exact questions, answer choices and reference answers, together with all 294 recorded model responses.
• Explicit model identity and result columns: model_name, model_revision, model_output, predicted_answer, and correct.

The English/Hindi image-question material and reference keys come from IndicVisionBench by krutrim-ai-labs. This study adds the aligned Roman Hinglish wording and the three-language Qwen3-VL evaluation. Source attribution and upstream terms are retained.

Browse the benchmark

Select trilingual_evaluation in Subset, then hinglish, hindi, or english in Split. The viewer starts with image → question → expected answer → model output → predicted answer → correctness → model name. Exact full prompts, checkpoint revision and source identifiers appear afterwards. The three splits share the same item_id values and images; there are 98 matched items, not 294 independent images. Each source image is stored once.

Evaluated model and results

Model: Qwen/Qwen3-VL-4B-Instruct. Revision: ebb281ec70b05090aa6165b016eac8ec08e71b17.

Language split	Correct	Items	Accuracy
hinglish	88	98	89.8%
hindi	88	98	89.8%
english	92	98	93.9%

model_output preserves the recorded response. predicted_answer decodes its option letter using that language's ordered choices. expected_answer is the text associated with the published reference key. correct compares the parsed prediction with that key. These are the saved evaluation results from the study.

Alignment and source record

The evaluation uses 98 selected English/Hindi pairs from IndicVisionBench's parallel VQA shard at revision 45269760b15938c6a345f74373f56c7a5ab7ecc8. The original shard contains 1,166 rows. The selected English/Hindi pairs share the image and aligned choices; Roman Hinglish counterparts preserve that question/choice structure. source_english_id and source_hindi_id retain the original IDs. Four reference-key mismatches excluded during selection were vqa_par_0102, vqa_par_0120, vqa_par_0261, and vqa_par_1087.

This repository contains only the IndicVisionBench-related evaluation. The separate StarVLA rollout repository contains robot recordings and LIBERO-specific VLM item sets. The study notebooks explain the broader multilingual VLA investigation.

Use

from datasets import load_dataset
repo = "SolusOps/Hinglish-IndicVisionBench-Benchmark"
hinglish = load_dataset(repo, "trilingual_evaluation", split="hinglish")
hindi = load_dataset(repo, "trilingual_evaluation", split="hindi")
english = load_dataset(repo, "trilingual_evaluation", split="english")
# Join the three splits by item_id for matched comparisons.