Variation of https://huggingface.co/tarekziade/distilvit
Trained on 270k images from Flickr10k and COCO. Training source code: https://github.com/tarekziade/distilvit
Results:
- eval_loss: 0.2305169701576233
- eval_rouge1: 39.511
- eval_rouge2: 14.7798
- eval_rougeL: 35.9476
- eval_rougeLsum: 35.9497
- eval_gen_len: 11.695219762592236