ReST-MCTS
ReST-MCTS*: LLM Self-Training via Process Reward Guided Tree Search (NeurIPS 2024)
Stars: 196
ReST-MCTS is a reinforced self-training approach that integrates process reward guidance with tree search MCTS to collect higher-quality reasoning traces and per-step value for training policy and reward models. It eliminates the need for manual per-step annotation by estimating the probability of steps leading to correct answers. The inferred rewards refine the process reward model and aid in selecting high-quality traces for policy model self-training.
README:
📃 [ReST-MCTS*]
[GitHub]
[Website]
We develop a reinforced self-training approach, called ReST-MCTS*, based on integrating process reward guidance with tree search MCTS* for collecting higher-quality reasoning traces as well as per-step value to train policy and reward models. ReST-MCTS* circumvents the per-step manual annotation typically used to train process rewards by tree-search-based reinforcement learning: Given oracle final correct answers, ReST-MCTS* is able to infer the correct process rewards by estimating the probability this step can help lead to the correct answer. These inferred rewards serve dual purposes: they act as value targets for further refining the process reward model and also facilitate the selection of high-quality traces for policy model self-training.
Considering the different dependency versions of transformers for Mistral (or Llama) and SciGLM, you should install different environments through miniconda and install corresponding required packages by:
running Mistral (or Llama)
pip install -r requirements_mistral.txtor running SciGLM
pip install -r requirements_sciglm.txtNote that for some models on huggingface like the GLM series, you may need to install specific versions of transformers.
To run MCTS* search, you should implement a policy as well as a process reward model (value model).
You can directly set these models by providing the model paths in the file models/model.py, substituting INFERENCE_MODEL_DIR, VALUE_BASE_MODEL_DIR and VALUE_MODEL_STATE_DICT.
INFERENCE_MODEL_DIR is the local path to the policy model, model could be Llama3-8b-Instruct, Mistral-7B: MetaMATH, and SciGLM-6B.
VALUE_BASE_MODEL_DIR is the local path to the value model. Considering the different dependency versions of transformers, Mistral-7B is adopted as the backbone of the value model when the policy model is Llama3-8B-Instruct or MetaMATH: Mistral-7B. When the policy model is SciGLM, we use ChatGLM3-6B as the backbone of the value model.
Aiming to gather value train data for science, we integrate questions of a lean science dataset $D_{sci}$ within [SciInstruct] to construct $D_{V_0}$. This dataset consists of 11,554 questions, where each question is paired with a correct step-by-step solution. (See Section 4.1 of the paper for more details.)
You can download [$D_{V_0}$] and put them in PRM/data to train Mistral-7B as the initial process reward model and obtain VALUE_MODEL_STATE_DICT.
We also provide PRM/train_VM_chatglm.py and PRM/train_VM_mistral.py.
We now only provide the implementation of the llama, glm and mistral as policy, with glm and mistral as value model in models/model.py.
If you are trying with other models, you can refer to our implementation and modify relevant codes to implement the corresponding models.
Once you've implemented the policy and value model, you should modify the LOCAL_INFERENCE_IDX and LOCAL_VALUE_IDX in models/model.py to the corresponding model index.
Before running search for evaluation or generation, you have to make sure your target question dataset is in the correct format. The data file should be a json file with items in the following format:
{
"content": "Calculate the sum of the first 10 prime numbers.",
"answer": "129"
}The content entry is required, serving as the question. While the answer entry is optional, it is used for evaluation.
The implementation of MCTS* search can be found in MCTS. We provide a search interface in MCTS/task.py. To run MCTS* search for a single question, you can refer to the following script:
from MCTS.task import *
question = "Calculate the sum of the first 10 prime numbers."
task = MCTS_Task(question, 'llama', 'local', lang='en')
output = task.run()
print(output['solution'])For evaluation of MCTS* on benchmarks, you can refer to evalaute.py, setting the parameter --mode to "mcts". You should specify the benchmark name and the exact file (subset) you want to evaluate. A simple demonstration is provided below:
python evaluate.py \
--task_name "scibench" \
--file "thermo" \
--propose_method "gpt" \
--value_method "local" \
--mode "mcts" \
--evaluate "scibench" \
--iteration_limit 50 \
--use_reflection "simple" \
--branch 3You can also refer to the MCTS/args.md for more details on the search parameters.
Given question set $D_G$, we use Llama3-8B-Instruct guided by MCTS* to generate synthetic data for policy model and value model. (See Algorithm 1 of the paper for more details.)
Download policy data (positive samples) for training 1st policy model (Llama3-8b-Instruct): [Hugging Face]
Download PRM data (positive and negative samples) for training 1st reward model (Mistral-7B: MetaMATH): [Hugging Face]
Download the trained policy model: [Hugging Face]
For our methods:
Regarding Llama3-8B-Instruct and Mistral-7B: MetaMATH, we use the default repo of [MAmmoTH] to train the policy model and evaluate.
Regarding SciGLM-6B, we use the default repo of [SciGLM] to train the policy model and evaluate.
We also implement self-rewarding as our baseline in ./self_train/self_train_dpo.py.
Self-training Results:
Accuracy of Different Verifiers:
Accuracy of Different Searches (we also provide the plot code in figures/plot_math_self_training.py):
If you find our work helpful, please kindly cite our paper:
@article{zhang2024rest,
title={ReST-MCTS*: LLM Self-Training via Process Reward Guided Tree Search},
author={Zhang, Dan and Zhoubian, Sining and Hu, Ziniu and Yue, Yisong and Dong, Yuxiao and Tang, Jie},
journal={arXiv preprint arXiv:2406.03816},
year={2024}
}
For Tasks:
Click tags to check more tools for each tasksFor Jobs:
Alternative AI tools for ReST-MCTS
Similar Open Source Tools
ReST-MCTS
ReST-MCTS is a reinforced self-training approach that integrates process reward guidance with tree search MCTS to collect higher-quality reasoning traces and per-step value for training policy and reward models. It eliminates the need for manual per-step annotation by estimating the probability of steps leading to correct answers. The inferred rewards refine the process reward model and aid in selecting high-quality traces for policy model self-training.
storm
STORM is a LLM system that writes Wikipedia-like articles from scratch based on Internet search. While the system cannot produce publication-ready articles that often require a significant number of edits, experienced Wikipedia editors have found it helpful in their pre-writing stage. **Try out our [live research preview](https://storm.genie.stanford.edu/) to see how STORM can help your knowledge exploration journey and please provide feedback to help us improve the system 🙏!**
giskard
Giskard is an open-source Python library that automatically detects performance, bias & security issues in AI applications. The library covers LLM-based applications such as RAG agents, all the way to traditional ML models for tabular data.
RepoAgent
RepoAgent is an LLM-powered framework designed for repository-level code documentation generation. It automates the process of detecting changes in Git repositories, analyzing code structure through AST, identifying inter-object relationships, replacing Markdown content, and executing multi-threaded operations. The tool aims to assist developers in understanding and maintaining codebases by providing comprehensive documentation, ultimately improving efficiency and saving time.
VITA
VITA is an open-source interactive omni multimodal Large Language Model (LLM) capable of processing video, image, text, and audio inputs simultaneously. It stands out with features like Omni Multimodal Understanding, Non-awakening Interaction, and Audio Interrupt Interaction. VITA can respond to user queries without a wake-up word, track and filter external queries in real-time, and handle various query inputs effectively. The model utilizes state tokens and a duplex scheme to enhance the multimodal interactive experience.
KnowAgent
KnowAgent is a tool designed for Knowledge-Augmented Planning for LLM-Based Agents. It involves creating an action knowledge base, converting action knowledge into text for model understanding, and a knowledgeable self-learning phase to continually improve the model's planning abilities. The tool aims to enhance agents' potential for application in complex situations by leveraging external reservoirs of information and iterative processes.
rosa
ROSA is an AI Agent designed to interact with ROS-based robotics systems using natural language queries. It can generate system reports, read and parse ROS log files, adapt to new robots, and run various ROS commands using natural language. The tool is versatile for robotics research and development, providing an easy way to interact with robots and the ROS environment.
llm-compressor
llm-compressor is an easy-to-use library for optimizing models for deployment with vllm. It provides a comprehensive set of quantization algorithms, seamless integration with Hugging Face models and repositories, and supports mixed precision, activation quantization, and sparsity. Supported algorithms include PTQ, GPTQ, SmoothQuant, and SparseGPT. Installation can be done via git clone and local pip install. Compression can be easily applied by selecting an algorithm and calling the oneshot API. The library also offers end-to-end examples for model compression. Contributions to the code, examples, integrations, and documentation are appreciated.
tonic_validate
Tonic Validate is a framework for the evaluation of LLM outputs, such as Retrieval Augmented Generation (RAG) pipelines. Validate makes it easy to evaluate, track, and monitor your LLM and RAG applications. Validate allows you to evaluate your LLM outputs through the use of our provided metrics which measure everything from answer correctness to LLM hallucination. Additionally, Validate has an optional UI to visualize your evaluation results for easy tracking and monitoring.
embodied-agents
Embodied Agents is a toolkit for integrating large multi-modal models into existing robot stacks with just a few lines of code. It provides consistency, reliability, scalability, and is configurable to any observation and action space. The toolkit is designed to reduce complexities involved in setting up inference endpoints, converting between different model formats, and collecting/storing datasets. It aims to facilitate data collection and sharing among roboticists by providing Python-first abstractions that are modular, extensible, and applicable to a wide range of tasks. The toolkit supports asynchronous and remote thread-safe agent execution for maximal responsiveness and scalability, and is compatible with various APIs like HuggingFace Spaces, Datasets, Gymnasium Spaces, Ollama, and OpenAI. It also offers automatic dataset recording and optional uploads to the HuggingFace hub.
premsql
PremSQL is an open-source library designed to help developers create secure, fully local Text-to-SQL solutions using small language models. It provides essential tools for building and deploying end-to-end Text-to-SQL pipelines with customizable components, ideal for secure, autonomous AI-powered data analysis. The library offers features like Local-First approach, Customizable Datasets, Robust Executors and Evaluators, Advanced Generators, Error Handling and Self-Correction, Fine-Tuning Support, and End-to-End Pipelines. Users can fine-tune models, generate SQL queries from natural language inputs, handle errors, and evaluate model performance against predefined metrics. PremSQL is extendible for customization and private data usage.
Vitron
Vitron is a unified pixel-level vision LLM designed for comprehensive understanding, generating, segmenting, and editing static images and dynamic videos. It addresses challenges in existing vision LLMs such as superficial instance-level understanding, lack of unified support for images and videos, and insufficient coverage across various vision tasks. The tool requires Python >= 3.8, Pytorch == 2.1.0, and CUDA Version >= 11.8 for installation. Users can deploy Gradio demo locally and fine-tune their models for specific tasks.
allms
allms is a versatile and powerful library designed to streamline the process of querying Large Language Models (LLMs). Developed by Allegro engineers, it simplifies working with LLM applications by providing a user-friendly interface, asynchronous querying, automatic retrying mechanism, error handling, and output parsing. It supports various LLM families hosted on different platforms like OpenAI, Google, Azure, and GCP. The library offers features for configuring endpoint credentials, batch querying with symbolic variables, and forcing structured output format. It also provides documentation, quickstart guides, and instructions for local development, testing, updating documentation, and making new releases.
jina
Jina is a tool that allows users to build multimodal AI services and pipelines using cloud-native technologies. It provides a Pythonic experience for serving ML models and transitioning from local deployment to advanced orchestration frameworks like Docker-Compose, Kubernetes, or Jina AI Cloud. Users can build and serve models for any data type and deep learning framework, design high-performance services with easy scaling, serve LLM models while streaming their output, integrate with Docker containers via Executor Hub, and host on CPU/GPU using Jina AI Cloud. Jina also offers advanced orchestration and scaling capabilities, a smooth transition to the cloud, and easy scalability and concurrency features for applications. Users can deploy to their own cloud or system with Kubernetes and Docker Compose integration, and even deploy to JCloud for autoscaling and monitoring.
cognee
Cognee is an open-source framework designed for creating self-improving deterministic outputs for Large Language Models (LLMs) using graphs, LLMs, and vector retrieval. It provides a platform for AI engineers to enhance their models and generate more accurate results. Users can leverage Cognee to add new information, utilize LLMs for knowledge creation, and query the system for relevant knowledge. The tool supports various LLM providers and offers flexibility in adding different data types, such as text files or directories. Cognee aims to streamline the process of working with LLMs and improving AI models for better performance and efficiency.
humanoid-gym
Humanoid-Gym is a reinforcement learning framework designed for training locomotion skills for humanoid robots, focusing on zero-shot transfer from simulation to real-world environments. It integrates a sim-to-sim framework from Isaac Gym to Mujoco for verifying trained policies in different physical simulations. The codebase is verified with RobotEra's XBot-S and XBot-L humanoid robots. It offers comprehensive training guidelines, step-by-step configuration instructions, and execution scripts for easy deployment. The sim2sim support allows transferring trained policies to accurate simulated environments. The upcoming features include Denoising World Model Learning and Dexterous Hand Manipulation. Installation and usage guides are provided along with examples for training PPO policies and sim-to-sim transformations. The code structure includes environment and configuration files, with instructions on adding new environments. Troubleshooting tips are provided for common issues, along with a citation and acknowledgment section.
For similar tasks
Xwin-LM
Xwin-LM is a powerful and stable open-source tool for aligning large language models, offering various alignment technologies like supervised fine-tuning, reward models, reject sampling, and reinforcement learning from human feedback. It has achieved top rankings in benchmarks like AlpacaEval and surpassed GPT-4. The tool is continuously updated with new models and features.
LLM-Tool-Survey
This repository contains a collection of papers related to tool learning with large language models (LLMs). The papers are organized according to the survey paper 'Tool Learning with Large Language Models: A Survey'. The survey focuses on the benefits and implementation of tool learning with LLMs, covering aspects such as task planning, tool selection, tool calling, response generation, benchmarks, evaluation, challenges, and future directions in the field. It aims to provide a comprehensive understanding of tool learning with LLMs and inspire further exploration in this emerging area.
Awesome-AI-Agents
Awesome-AI-Agents is a curated list of projects, frameworks, benchmarks, platforms, and related resources focused on autonomous AI agents powered by Large Language Models (LLMs). The repository showcases a wide range of applications, multi-agent task solver projects, agent society simulations, and advanced components for building and customizing AI agents. It also includes frameworks for orchestrating role-playing, evaluating LLM-as-Agent performance, and connecting LLMs with real-world applications through platforms and APIs. Additionally, the repository features surveys, paper lists, and blogs related to LLM-based autonomous agents, making it a valuable resource for researchers, developers, and enthusiasts in the field of AI.
ReST-MCTS
ReST-MCTS is a reinforced self-training approach that integrates process reward guidance with tree search MCTS to collect higher-quality reasoning traces and per-step value for training policy and reward models. It eliminates the need for manual per-step annotation by estimating the probability of steps leading to correct answers. The inferred rewards refine the process reward model and aid in selecting high-quality traces for policy model self-training.
LongCite
LongCite is a tool that enables Large Language Models (LLMs) to generate fine-grained citations in long-context Question Answering (QA) scenarios. It provides models trained on GLM-4-9B and Meta-Llama-3.1-8B, supporting up to 128K context. Users can deploy LongCite chatbots, generate accurate responses, and obtain precise sentence-level citations. The tool includes components for model deployment, Coarse to Fine (CoF) pipeline for data construction, model training using LongCite-45k dataset, evaluation with LongBench-Cite benchmark, and citation generation.
For similar jobs
weave
Weave is a toolkit for developing Generative AI applications, built by Weights & Biases. With Weave, you can log and debug language model inputs, outputs, and traces; build rigorous, apples-to-apples evaluations for language model use cases; and organize all the information generated across the LLM workflow, from experimentation to evaluations to production. Weave aims to bring rigor, best-practices, and composability to the inherently experimental process of developing Generative AI software, without introducing cognitive overhead.
LLMStack
LLMStack is a no-code platform for building generative AI agents, workflows, and chatbots. It allows users to connect their own data, internal tools, and GPT-powered models without any coding experience. LLMStack can be deployed to the cloud or on-premise and can be accessed via HTTP API or triggered from Slack or Discord.
VisionCraft
The VisionCraft API is a free API for using over 100 different AI models. From images to sound.
kaito
Kaito is an operator that automates the AI/ML inference model deployment in a Kubernetes cluster. It manages large model files using container images, avoids tuning deployment parameters to fit GPU hardware by providing preset configurations, auto-provisions GPU nodes based on model requirements, and hosts large model images in the public Microsoft Container Registry (MCR) if the license allows. Using Kaito, the workflow of onboarding large AI inference models in Kubernetes is largely simplified.
PyRIT
PyRIT is an open access automation framework designed to empower security professionals and ML engineers to red team foundation models and their applications. It automates AI Red Teaming tasks to allow operators to focus on more complicated and time-consuming tasks and can also identify security harms such as misuse (e.g., malware generation, jailbreaking), and privacy harms (e.g., identity theft). The goal is to allow researchers to have a baseline of how well their model and entire inference pipeline is doing against different harm categories and to be able to compare that baseline to future iterations of their model. This allows them to have empirical data on how well their model is doing today, and detect any degradation of performance based on future improvements.
tabby
Tabby is a self-hosted AI coding assistant, offering an open-source and on-premises alternative to GitHub Copilot. It boasts several key features: * Self-contained, with no need for a DBMS or cloud service. * OpenAPI interface, easy to integrate with existing infrastructure (e.g Cloud IDE). * Supports consumer-grade GPUs.
spear
SPEAR (Simulator for Photorealistic Embodied AI Research) is a powerful tool for training embodied agents. It features 300 unique virtual indoor environments with 2,566 unique rooms and 17,234 unique objects that can be manipulated individually. Each environment is designed by a professional artist and features detailed geometry, photorealistic materials, and a unique floor plan and object layout. SPEAR is implemented as Unreal Engine assets and provides an OpenAI Gym interface for interacting with the environments via Python.
Magick
Magick is a groundbreaking visual AIDE (Artificial Intelligence Development Environment) for no-code data pipelines and multimodal agents. Magick can connect to other services and comes with nodes and templates well-suited for intelligent agents, chatbots, complex reasoning systems and realistic characters.