surabhi1407/docvqa-hallucination-critic
收藏Hugging Face2026-03-26 更新2026-03-29 收录
下载链接：
https://hf-mirror.com/datasets/surabhi1407/docvqa-hallucination-critic
下载链接
链接失效反馈官方服务：
资源简介：
---
dataset_info:
features:
- name: image
dtype: image
- name: question
dtype: string
- name: provided_answer
dtype: string
- name: label
dtype: string
- name: original_true_answer
dtype: string
- name: question_type
dtype: string
splits:
- name: train
num_examples: 5201
configs:
- config_name: default
data_files:
- split: train
path: data/train/shard-*.parquet
---
# docvqa-hallucination-critic
Hard-negative hallucination dataset derived from DocVQA validation split.
## Purpose
Binary classification stress-test for vision-language models (e.g. LLaVA base).
Each row contains a document image, a question, and a `provided_answer` that is
either the true answer (**Valid**) or a plausible-but-wrong distractor generated
by DeepSeek (**Hallucination**).
## Schema
| Column | Description |
|---|---|
| `image` | Document image (from DocVQA) |
| `question` | Original DocVQA question |
| `provided_answer` | True answer OR hard-negative distractor |
| `label` | `Valid` or `Hallucination` |
| `original_true_answer` | Always the ground-truth answer |
| `question_type` | `yes_no` / `date` / `numerical` / `named_entity` / `free_text` |
## Stats
- Total rows: 5201
- Shards: 14
提供机构：
surabhi1407
