---
dataset_info:
  features:
  - name: image_id
    dtype: int64
  - name: image
    dtype: image
  - name: caption
    dtype: string
  - name: source
    dtype: string
  - name: caption_gemini-2.5-pro
    dtype: string
  - name: caption_gemini-3.1-flash-lite-preview
    dtype: string
  - name: caption_gemini-2.5-flash
    dtype: string
  - name: caption_vllm_qwen35_detail
    dtype: string
  - name: caption_vllm_qwen35_pretrained
    dtype: string
  - name: caption_vllm_qwen35_27b
    dtype: string
  - name: caption_aya_vision_8b
    dtype: string
  - name: caption_qwen3vl_4b
    dtype: string
  splits:
  - name: train
    num_bytes: 572815308
    num_examples: 5000
  download_size: 551933145
  dataset_size: 572815308
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
# Dataset de captions geradas por:
## Três modelos Gemini -> gemini-2.5-pro // gemini-3.1-flash-lite-preview // gemini-2.5-flash
## Três modelos Open Source -> Qwen3-VL-4B-Instruct // aya_vision_8b // Qwen3.5-27B
## Dois modelos VerboVision -> qwen3.5-4b-pretrained-merged // qwen3.5-4b-verbovision-detail-tags-merged
## -----------------------------------------------------------------------------------------------------------

# Dataset base: VerboVision/PraCegoVer-Filtrado-FSB
## -----------------------------------------------------------------------------------------------------------
# Quantidade de amostras:
## gemini-2.5-pro -> ~1000 amostras
## Outros -> ~5000 amostras
## -----------------------------------------------------------------------------------------------------------

# Dataset base para seleção por preferência para treinamento por reforço