Tuning Large Multimodal Models for Videos using Reinforcement Learning from AI Feedback,
Daechul Ahn1,3, Yura Choi1,3, Youngjae Yu1, Dongyeop Kang2, Jonghyun Choi3,†
1Yonsei University, 2University of Minnesota, 3Seoul National University
†Corresponding Author
ACL 2024 (To appear)
- [Jul 16, 2024] 🎙️ VLM-RLAIF has been selected for ✨oral presentation✨ at ACL 2024! See you in Bangkok 🇹🇭
- [Jun 16, 2024] 🔥 Our next work on aligning large video multimodal model, i-SRT🚄, is now available [arXiv, code]
- [May 31, 2024] 🥳 VLM-RLAIF is accepted to ACL 2024 !
Abstract: Recent advancements in large language models have influenced the development of video large multimodal models (VLMMs). Previous approaches for VLMMs involve Supervised Fine-Tuning (SFT) with instruction-tuned datasets, integrating LLM with visual encoders, and additional learnable parameters. Here, aligning video with text, and vice versa, remains a challenge, primarily due to the insufficient quality and quantity of multimodal instruction-tune data compared to that of text-only. This discrepancy often results in alignments that poorly ground the video content. To address this, we present a novel alignment strategy that employs a multimodal AI system equipped with Reinforcement Learning from AI Feedback (RLAIF), providing self-preference feedback to refine itself and facilitating the alignment of video and text modalities. Our approach uniquely integrates detailed video descriptions as context into a multimodal AI system during preference feedback generation to enrich the understanding of video content, a process we call context-aware reward modeling. Empirical evaluations on various video benchmarks demonstrate that our VLM-RLAIF outperforms existing approaches, including the SFT model.
Pipeline of VLM-RLAIF
Check PREPARE_DATASET.md to prepare training & validation datasets
| Model | Size | Checkpoint | corr. | detail. | context | temp. | const. |
|---|---|---|---|---|---|---|---|
| RLAIF | 7B | SNUMPR/vlm_rlaif_video_llava_7b | 3.63 | 3.25 | 4.00 | 3.23 | 3.32 |
| SFT | 7B | SNUMPR/vlm_sft_video_llava_7b | 2.79 | 2.82 | 3.37 | 2.28 | 2.49 |
| Dataset Usage | Link |
|---|---|
| SFT (short) | SNUMPR/vlm_rlaif_datasets/SFT_short.json |
| SFT (long) | SNUMPR/vlm_rlaif_datasets/SFT_long.json |
| Preference dataset (for RM) | SNUMPR/vlm_rlaif_datasets/RM_13b_v1_dataset_39k.json |
| PPO init | SNUMPR/vlm_rlaif_datasets/PPO_init.json |
| RLAIF | SNUMPR/vlm_rlaif_datasets/RL_data.json |
Check PREPARE_DATASET.md to prepare training & validation datasets
- Zero-shot QA
bash Evaluation/zeroshotqa/scripts/zeroshotqa_pipeline.sh
- Video Generative Benchmark
bash Evaluation/scripts/videochatgpt_pipeline.sh
- Refer to the RLAIF folder to train reward model, policy model, and do PPO
Available Soon
@inproceedings{ahnCYKC24,
author = {Daechul Ahn and Yura Choi and Youngjae Yu and Dongyeop Kang and Jonghyun Choi},
title = {Tuning Large Multimodal Models for Videos using Reinforcement Learning from AI Feedback},
booktitle = {ACL},
year = {2024}
}
GNU GENERAL PUBLIC LICENSE