MM-RLHF
The Next Step Forward in Multimodal LLM Alignment
Stars: 116
MM-RLHF is a comprehensive project for aligning Multimodal Large Language Models (MLLMs) with human preferences. It includes a high-quality MLLM alignment dataset, a Critique-Based MLLM reward model, a novel alignment algorithm MM-DPO, and benchmarks for reward models and multimodal safety. The dataset covers image understanding, video understanding, and safety-related tasks with model-generated responses and human-annotated scores. The reward model generates critiques of candidate texts before assigning scores for enhanced interpretability. MM-DPO is an alignment algorithm that achieves performance gains with simple adjustments to the DPO framework. The project enables consistent performance improvements across 10 dimensions and 27 benchmarks for open-source MLLMs.
README:
[📖 arXiv Paper] [📊 MM-RLHF Data] [📝 Homepage]
[🏆 Reward Model] [🔮 MM-RewardBench] [🔮 MM-SafetyBench] [📈 Evaluation Suite]
[2025/03/04] 🔥 MM-RLHF is now available on model scope.
[2025/02/27] 🔥 MM-RLHF is now fully supported by swift. Simply process your data using scripts/convert_to_swift.py and execute the scripts/swift.sh command to get started.
[2025/02/10] 🔥 We are proud to open-source MM-RLHF, a comprehensive project for aligning Multimodal Large Language Models (MLLMs) with human preferences. This release includes:
- A high-quality MLLM alignment dataset (120K samples, created by over 50 experts over two months, including ratings and manual annotations across eight dimensions.).
- A strong Critique-Based MLLM reward model which is trained on human annotations, achieving state-of-the-art (SOTA) performance on public benchmarks.
- A novel alignment algorithm MM-DPO, effectively integrates reward signals to improve the data efficiency of DPO training..
- Two new benchmarks designed for the reward model and multimodal safety, addressing gaps in existing benchmarks in these areas..
Our dataset and algorithms enable consistent performance improvements across 10 dimensions and 27 benchmarks for open-source MLLMs.
1. MM-RLHF Dataset (data.jsonl in MM-RLHF Data)
- 20k instructions covering image understanding, video understanding, and safety-related tasks.
- Each instruction includes 3-5 model-generated responses, along with human-annotated scores, rankings, and fine-grained textual feedback.
-
80k comparison pairs derived from ranked samples for each instruction, suitable for RLHF training.
- We release the MM-RLHF-Reward-7B, a Critique-Based Reward Model that generates critiques of candidate texts before assigning scores, offering enhanced interpretability and more informative feedback.
- Includes the training algorithm for the reward model, enabling researchers to reproduce and extend our work.
- Complete training code for MM-DPO, a novel alignment algorithm that achieves significant performance gains with simple adjustments to the DPO framework.
- MM-RLHF-RewardBench: Evaluates the quality of reward models.
- MM-RLHF-SafetyBench: Focuses on MLLM safety, including tasks like adversarial attacks, red teaming, jailbreaking, and harmful content detection.
git clone https://github.com/yfzhang114/MM-RLHF
cd MM-RLHFInstall the inference package:
conda create -n llava python=3.10 -y
conda activate llava
pip install --upgrade pip # Enable PEP 660 support.
pip install -e .
pip install -e ".[train]"
pip install flash-attn --no-build-isolationDownload 📊 MM-RLHF Data, unzip the image and video datasets, and the final structure should look like this:
MM-RLHF
-- | long
-- | mcq
-- | safety
-- | short
-- | data.jsonl
-- | dpo_pairs.jsonl
Here, data.jsonl contains all labeled information, and dpo_pairs.jsonl contains pairs of different ranks used for subsequent DPO and reward model training. The other folders contain the image and video frames.
Specify the model to learn the critic and provide the reward's loss weight, such as critic_rewards_weight and float_rewards_weight.
sh scripts/train/critic_reward_7b.shStep 1: Precompute Logits with Reference Model
To save GPU memory during DPO training, precompute logits using the reference model. Specify DATA_PATH and OUTPUT_DATA_PATH in the script.
sh scripts/train/generate_ref_logits.shThe output file will add elements like "reference_chosen_logp" and "reference_rejected_logp" to the data.
Step 2: Run DPO Algorithm with Precomputed Logits
sh scripts/train/dpo_ov7b.shFor alignment models, the evaluation code is available in the mmrlhf-eval repository. This code provides various evaluation tasks that test the alignment capabilities of your model across multiple benchmarks. These benchmarks include standard tests for model robustness, safety, and hallucination handling in multimodal contexts.
For reward models, we offer the MM-RLHF-RewardBench (available on Hugging Face at MM-RLHF-RewardBench) for detailed evaluation. To perform the evaluation, download the required images and the mm_reward_bench.jsonl file from the repository and place them in the appropriate directories. Then, follow these steps:
-
Download the dataset and necessary files:
- Ensure the repository contains the image data and the
mm_reward_bench.jsonlfile in the designated folder (path_to_data/).
- Ensure the repository contains the image data and the
-
Run the reward model evaluation: This step involves evaluating the reward model against the dataset by running the following command:
python llava/eval/eval_mm_reward_bench.py --model-path your_reward_model --question-file path_to_data/mm_reward_bench.jsonl --answers-file your_answer_file
-
Calculate performance metrics: After generating the answer file, you can calculate the performance of your reward model by running the following command:
python llava/eval/cal_performance_mmreward_bench.py --input_file your_answer_file
If you find it useful for your research and applications, please cite related papers/blogs using this BibTeX:
@article{zhang2025mm,
title={MM-RLHF: The Next Step Forward in Multimodal LLM Alignment},
author={Zhang, Yi-Fan and Yu, Tao and Tian, Haochen and Fu, Chaoyou and Li, Peiyan and Zeng, Jianshu and Xie, Wulin and Shi, Yang and Zhang, Huanyu and Wu, Junkang and others},
journal={arXiv preprint arXiv:2502.10391},
year={2025}
}- MME-RealWorld: Could Your Multimodal LLM Challenge High-Resolution Real-World Scenarios that are Difficult for Humans?
- MME-Survey: A Comprehensive Survey on Evaluation of Multimodal LLMs
- Beyond LLaVA-HD: Diving into High-Resolution Large Multimodal Models
- VITA-1.5: Towards GPT-4o Level Real-Time Vision and Speech Interaction
For Tasks:
Click tags to check more tools for each tasksFor Jobs:
Alternative AI tools for MM-RLHF
Similar Open Source Tools
MM-RLHF
MM-RLHF is a comprehensive project for aligning Multimodal Large Language Models (MLLMs) with human preferences. It includes a high-quality MLLM alignment dataset, a Critique-Based MLLM reward model, a novel alignment algorithm MM-DPO, and benchmarks for reward models and multimodal safety. The dataset covers image understanding, video understanding, and safety-related tasks with model-generated responses and human-annotated scores. The reward model generates critiques of candidate texts before assigning scores for enhanced interpretability. MM-DPO is an alignment algorithm that achieves performance gains with simple adjustments to the DPO framework. The project enables consistent performance improvements across 10 dimensions and 27 benchmarks for open-source MLLMs.
APOLLO
APOLLO is a memory-efficient optimizer designed for large language model (LLM) pre-training and full-parameter fine-tuning. It offers SGD-like memory cost with AdamW-level performance. The optimizer integrates low-rank approximation and optimizer state redundancy reduction to achieve significant memory savings while maintaining or surpassing the performance of Adam(W). Key contributions include structured learning rate updates for LLM training, approximated channel-wise gradient scaling in a low-rank auxiliary space, and minimal-rank tensor-wise gradient scaling. APOLLO aims to optimize memory efficiency during training large language models.
Vodalus-Expert-LLM-Forge
Vodalus Expert LLM Forge is a tool designed for crafting datasets and efficiently fine-tuning models using free open-source tools. It includes components for data generation, LLM interaction, RAG engine integration, model training, fine-tuning, and quantization. The tool is suitable for users at all levels and is accompanied by comprehensive documentation. Users can generate synthetic data, interact with LLMs, train models, and optimize performance for local execution. The tool provides detailed guides and instructions for setup, usage, and customization.
OpenManus-RL
OpenManus-RL is an open-source initiative focused on enhancing reasoning and decision-making capabilities of large language models (LLMs) through advanced reinforcement learning (RL)-based agent tuning. The project explores novel algorithmic structures, diverse reasoning paradigms, sophisticated reward strategies, and extensive benchmark environments. It aims to push the boundaries of agent reasoning and tool integration by integrating insights from leading RL tuning frameworks and continuously updating progress in a dynamic, live-streaming fashion.
MathVerse
MathVerse is an all-around visual math benchmark designed to evaluate the capabilities of Multi-modal Large Language Models (MLLMs) in visual math problem-solving. It collects high-quality math problems with diagrams to assess how well MLLMs can understand visual diagrams for mathematical reasoning. The benchmark includes 2,612 problems transformed into six versions each, contributing to 15K test samples. It also introduces a Chain-of-Thought (CoT) Evaluation strategy for fine-grained assessment of output answers.
WatermarkRemover-AI
WatermarkRemover-AI is an advanced application that utilizes AI models for precise watermark detection and seamless removal. It leverages Florence-2 for watermark identification and LaMA for inpainting. The tool offers both a command-line interface (CLI) and a PyQt6-based graphical user interface (GUI), making it accessible to users of all levels. It supports dual modes for processing images, advanced watermark detection, seamless inpainting, customizable output settings, real-time progress tracking, dark mode support, and efficient GPU acceleration using CUDA.
llms-interview-questions
This repository contains a comprehensive collection of 63 must-know Large Language Models (LLMs) interview questions. It covers topics such as the architecture of LLMs, transformer models, attention mechanisms, training processes, encoder-decoder frameworks, differences between LLMs and traditional statistical language models, handling context and long-term dependencies, transformers for parallelization, applications of LLMs, sentiment analysis, language translation, conversation AI, chatbots, and more. The readme provides detailed explanations, code examples, and insights into utilizing LLMs for various tasks.
Simplifine
Simplifine is an open-source library designed for easy LLM finetuning, enabling users to perform tasks such as supervised fine tuning, question-answer finetuning, contrastive loss for embedding tasks, multi-label classification finetuning, and more. It provides features like WandB logging, in-built evaluation tools, automated finetuning parameters, and state-of-the-art optimization techniques. The library offers bug fixes, new features, and documentation updates in its latest version. Users can install Simplifine via pip or directly from GitHub. The project welcomes contributors and provides comprehensive documentation and support for users.
aibrix
AIBrix is an open-source initiative providing essential building blocks for scalable GenAI inference infrastructure. It delivers a cloud-native solution optimized for deploying, managing, and scaling large language model (LLM) inference, tailored to enterprise needs. Key features include High-Density LoRA Management, LLM Gateway and Routing, LLM App-Tailored Autoscaler, Unified AI Runtime, Distributed Inference, Distributed KV Cache, Cost-efficient Heterogeneous Serving, and GPU Hardware Failure Detection.
RealtimeSTT_LLM_TTS
RealtimeSTT is an easy-to-use, low-latency speech-to-text library for realtime applications. It listens to the microphone and transcribes voice into text, making it ideal for voice assistants and applications requiring fast and precise speech-to-text conversion. The library utilizes Voice Activity Detection, Realtime Transcription, and Wake Word Activation features. It supports GPU-accelerated transcription using PyTorch with CUDA support. RealtimeSTT offers various customization options for different parameters to enhance user experience and performance. The library is designed to provide a seamless experience for developers integrating speech-to-text functionality into their applications.
vearch
Vearch is a cloud-native distributed vector database designed for efficient similarity search of embedding vectors in AI applications. It supports hybrid search with vector search and scalar filtering, offers fast vector retrieval from millions of objects in milliseconds, and ensures scalability and reliability through replication and elastic scaling out. Users can deploy Vearch cluster on Kubernetes, add charts from the repository or locally, start with Docker-compose, or compile from source code. The tool includes components like Master for schema management, Router for RESTful API, and PartitionServer for hosting document partitions with raft-based replication. Vearch can be used for building visual search systems for indexing images and offers a Python SDK for easy installation and usage. The tool is suitable for AI developers and researchers looking for efficient vector search capabilities in their applications.
qapyq
qapyq is an image viewer and AI-assisted editing tool designed to help curate datasets for generative AI models. It offers features such as image viewing, editing, captioning, batch processing, and AI assistance. Users can perform tasks like cropping, scaling, editing masks, tagging, and applying sorting and filtering rules. The tool supports state-of-the-art captioning and masking models, with options for model settings, GPU acceleration, and quantization. qapyq aims to streamline the process of preparing images for training AI models by providing a user-friendly interface and advanced functionalities.
CortexON
CortexON is an open-source, multi-agent AI system designed to automate and simplify everyday tasks. It integrates specialized agents like Web Agent, File Agent, Coder Agent, Executor Agent, and API Agent to accomplish user-defined objectives. CortexON excels at executing complex workflows, research tasks, technical operations, and business process automations by dynamically coordinating the agents' unique capabilities. It offers advanced research automation, multi-agent orchestration, integration with third-party APIs, code generation and execution, efficient file and data management, and personalized task execution for travel planning, market analysis, educational content creation, and business intelligence.
griptape
Griptape is a modular Python framework for building AI-powered applications that securely connect to your enterprise data and APIs. It offers developers the ability to maintain control and flexibility at every step. Griptape's core components include Structures (Agents, Pipelines, and Workflows), Tasks, Tools, Memory (Conversation Memory, Task Memory, and Meta Memory), Drivers (Prompt and Embedding Drivers, Vector Store Drivers, Image Generation Drivers, Image Query Drivers, SQL Drivers, Web Scraper Drivers, and Conversation Memory Drivers), Engines (Query Engines, Extraction Engines, Summary Engines, Image Generation Engines, and Image Query Engines), and additional components (Rulesets, Loaders, Artifacts, Chunkers, and Tokenizers). Griptape enables developers to create AI-powered applications with ease and efficiency.
replexica
Replexica is an i18n toolkit for React, to ship multi-language apps fast. It doesn't require extracting text into JSON files, and uses AI-powered API for content processing. It comes in two parts: 1. Replexica Compiler - an open-source compiler plugin for React; 2. Replexica API - an i18n API in the cloud that performs translations using LLMs. (Usage based, has a free tier.) Replexica supports several i18n formats: 1. JSON-free Replexica compiler format; 2. .md files for Markdown content; 3. Legacy JSON and YAML-based formats.
For similar tasks
MM-RLHF
MM-RLHF is a comprehensive project for aligning Multimodal Large Language Models (MLLMs) with human preferences. It includes a high-quality MLLM alignment dataset, a Critique-Based MLLM reward model, a novel alignment algorithm MM-DPO, and benchmarks for reward models and multimodal safety. The dataset covers image understanding, video understanding, and safety-related tasks with model-generated responses and human-annotated scores. The reward model generates critiques of candidate texts before assigning scores for enhanced interpretability. MM-DPO is an alignment algorithm that achieves performance gains with simple adjustments to the DPO framework. The project enables consistent performance improvements across 10 dimensions and 27 benchmarks for open-source MLLMs.
CritiqueLLM
CritiqueLLM is an official implementation of a model designed for generating informative critiques to evaluate large language model generation. It includes functionalities for data collection, referenced pointwise grading, referenced pairwise comparison, reference-free pairwise comparison, reference-free pointwise grading, inference for pointwise grading and pairwise comparison, and evaluation of the generated results. The model aims to provide a comprehensive framework for evaluating the performance of large language models based on human ratings and comparisons.
Vision-LLM-Alignment
Vision-LLM-Alignment is a repository focused on implementing alignment training for visual large language models (LLMs), including SFT training, reward model training, and PPO/DPO training. It supports various model architectures and provides datasets for training. The repository also offers benchmark results and installation instructions for users.
foundations-of-gen-ai
This repository contains code for the O'Reilly Live Online Training for 'Transformer Architectures for Generative AI'. The course provides a deep understanding of transformer architectures and their impact on natural language processing (NLP) and vision tasks. Participants learn to harness transformers to tackle problems in text, image, and multimodal AI through theory and practical exercises.
For similar jobs
sweep
Sweep is an AI junior developer that turns bugs and feature requests into code changes. It automatically handles developer experience improvements like adding type hints and improving test coverage.
teams-ai
The Teams AI Library is a software development kit (SDK) that helps developers create bots that can interact with Teams and Microsoft 365 applications. It is built on top of the Bot Framework SDK and simplifies the process of developing bots that interact with Teams' artificial intelligence capabilities. The SDK is available for JavaScript/TypeScript, .NET, and Python.
ai-guide
This guide is dedicated to Large Language Models (LLMs) that you can run on your home computer. It assumes your PC is a lower-end, non-gaming setup.
classifai
Supercharge WordPress Content Workflows and Engagement with Artificial Intelligence. Tap into leading cloud-based services like OpenAI, Microsoft Azure AI, Google Gemini and IBM Watson to augment your WordPress-powered websites. Publish content faster while improving SEO performance and increasing audience engagement. ClassifAI integrates Artificial Intelligence and Machine Learning technologies to lighten your workload and eliminate tedious tasks, giving you more time to create original content that matters.
chatbot-ui
Chatbot UI is an open-source AI chat app that allows users to create and deploy their own AI chatbots. It is easy to use and can be customized to fit any need. Chatbot UI is perfect for businesses, developers, and anyone who wants to create a chatbot.
BricksLLM
BricksLLM is a cloud native AI gateway written in Go. Currently, it provides native support for OpenAI, Anthropic, Azure OpenAI and vLLM. BricksLLM aims to provide enterprise level infrastructure that can power any LLM production use cases. Here are some use cases for BricksLLM: * Set LLM usage limits for users on different pricing tiers * Track LLM usage on a per user and per organization basis * Block or redact requests containing PIIs * Improve LLM reliability with failovers, retries and caching * Distribute API keys with rate limits and cost limits for internal development/production use cases * Distribute API keys with rate limits and cost limits for students
uAgents
uAgents is a Python library developed by Fetch.ai that allows for the creation of autonomous AI agents. These agents can perform various tasks on a schedule or take action on various events. uAgents are easy to create and manage, and they are connected to a fast-growing network of other uAgents. They are also secure, with cryptographically secured messages and wallets.
griptape
Griptape is a modular Python framework for building AI-powered applications that securely connect to your enterprise data and APIs. It offers developers the ability to maintain control and flexibility at every step. Griptape's core components include Structures (Agents, Pipelines, and Workflows), Tasks, Tools, Memory (Conversation Memory, Task Memory, and Meta Memory), Drivers (Prompt and Embedding Drivers, Vector Store Drivers, Image Generation Drivers, Image Query Drivers, SQL Drivers, Web Scraper Drivers, and Conversation Memory Drivers), Engines (Query Engines, Extraction Engines, Summary Engines, Image Generation Engines, and Image Query Engines), and additional components (Rulesets, Loaders, Artifacts, Chunkers, and Tokenizers). Griptape enables developers to create AI-powered applications with ease and efficiency.