---
title: BFS-Prover-V2-7B
canonical_url: "https://www.modelscope.cn/models/bytedance-community/BFS-Prover-V2-7B"
md_url: "https://www.modelscope.cn/models/bytedance-community/BFS-Prover-V2-7B.md"
repository: bytedance-community/BFS-Prover-V2-7B
last_updated: 2026-06-05
license: apache-2.0
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - qwen2
architectures:
  - Qwen2ForCausalLM
base_model:
  - Qwen/Qwen2.5-Math-7B
base_model_relation: finetune
parameters: 7.6B
tensor_type:
  - BF16
library_name:
  - transformer
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 14
stars: 0
tags:
  - lean4
  - step-prover
---

# BFS-Prover-V2-7B

> BFS-Prover-V2-7B - bytedance-community 在 ModelScope 开源的模型。BFS-Prover-V2: Scaling up Multi-Turn Off-Policy RL and Multi-Agent Tree Search for LLM Step-Provers Introduction

bytedance-community/BFS-Prover-V2-7B 是 ModelScope 魔搭社区上的 7.6B 参数text-generation模型，采用 apache-2.0 许可，基于 Qwen/Qwen2.5-Math-7B 构建。

- **Repository**: bytedance-community/BFS-Prover-V2-7B
- **License**: apache-2.0
- **Tasks**: text-generation
- **Parameters**: 7.6B
- **Base model**: Qwen/Qwen2.5-Math-7B
- **Tags**: lean4, step-prover
- **Downloads**: 14
- **Stars**: 0
- **Last updated**: 2026-06-05

Source: https://www.modelscope.cn/models/bytedance-community/BFS-Prover-V2-7B

---

<div align="center">
  <h1 style="font-size: 2.0em;">BFS-Prover-V2: Scaling up Multi-Turn Off-Policy RL and Multi-Agent Tree Search for LLM Step-Provers</h1>
  <div align="center" style="line-height: 1;">
    <a href="https://bfs-prover.github.io/V2/">
      <img src="https://img.shields.io/badge/Homepage-BFS--Prover--V2-78DED4?style=flat-square&labelColor=2E5AA8">
    </a>
    <a href="https://arxiv.org/abs/2509.06493">
      <img src="https://img.shields.io/badge/arXiv-2509.06493-b31b1b.svg?style=flat-square&labelColor=2E5AA8">
    </a>
    <a href="https://github.com/ByteDance-Seed/BFS-Prover-V2">
      <img src="https://img.shields.io/badge/GitHub-BFS--Prover--V2-808080?&style=flat-square&labelColor=2E5AA8">
    </a>
    <a href="https://github.com/cmu-l3/llmlean">
      <img src="https://img.shields.io/badge/Integration-LLMLean-black?style=flat-square&labelColor=2E5AA8">
    </a>
    <a href="https://www.apache.org/licenses/LICENSE-2.0.txt">
      <img src="https://img.shields.io/badge/License-Apache%202.0-purple.svg?style=flat-square&labelColor=2E5AA8">
    </a>
  </div>
</div>
  
## Introduction

We introduce **BFS-Prover-V2**, the state-of-the-art open-source step-level theorem proving system for Lean4, designed to address the dual challenges of scaling both training and inference in neural theorem proving. BFS-Prover-V2 introduces novel solutions to overcome these limitations through:

1. **Training-time scaling**: A novel multi-stage expert iteration framework with adaptive tactic-level data filtering and periodic retraining to surmount the performance plateaus that typically curtail long-term post training

2. **Inference-time scaling**: A planner-enhanced multi-agent tree search system for hierarchical reasoning that scales performance at inference time

**BFS-Prover-V2** achieves 95.08\% and 41.4\% on the miniF2F and ProofNet test sets respectively, setting a new state-of-the-art for step-level provers.

This repo contains the **BFS-Prover-V2-7B** model, with the following features:

- Base Model: Qwen2.5-Math-7B
- Training Approach: Multi-stage expert iteration with best-first tree search
- Training Data Sources:
  - Mathlib (via LeanDojo)
  - Lean-Github repositories
  - Autoformalized NuminaMath datasets
  - Goedel-Pset

## Benchmark Performance

<div align="center">

| Model | miniF2F-test | miniF2F-valid | ProofNet-test |
|:------|:------------:|:-------------:|:-------------:|
| 👉 **BFS-Prover-V2-7B** | 82.4% | - | - |
| BFS-Prover-V2-32B | 86.1% | 85.5% | 41.4% |
| BFS-Prover-V2-32B w/ Planner | 95.08% | 95.5% | - |
</div>

## Usage
- The model expects input in the format `"{state}:::"` where {state} is a Lean4 tactic state.
- `:::` serves as a special indicator to signal the model to generate a tactic for the given state. 
- The model will echo back the input state followed by the generated tactic.


```python
# Example code for loading and using the tactic generator model
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("ByteDance-Seed/BFS-Prover-V2-7B")
tokenizer = AutoTokenizer.from_pretrained("ByteDance-Seed/BFS-Prover-V2-7B")

# imo_1964_p2 from miniF2F
state = """a b c : ℝ

    h₀ : 0 < a ∧ 0 < b ∧ 0 < c

    h₁ : c < a + b

    h₂ : b < a + c

    h₃ : a < b + c

    ⊢ a ^ 2 * (b + c - a) + b ^ 2 * (c + a - b) + c ^ 2 * (a + b - c) ≤ 3 * a * b * c"""

# Tactic generation
sep = ":::"
prompt = state + sep
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs)
tactic = tokenizer.decode(outputs[0], skip_special_tokens=True).split(sep)[1]
print(tactic)

# Generated tactic: "nlinarith [sq_nonneg (a - b), sq_nonneg (c - a), sq_nonneg (b - c)]"
```

## Citation

```bibtex
@article{xin2025scaling,
  title={Scaling up Multi-Turn Off-Policy RL and Multi-Agent Tree Search for LLM Step-Provers},
  author={Xin, Ran and Zheng, Zeyu and Nie, Yanchen and Yuan, Kun and Xiao, Xia},
  journal={arXiv preprint arXiv:2509.06493},
  year={2025}
}
```

## License

This project is licensed under the [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0.txt).

## Contact

For questions and feedback about the tactic generator model, please contact:
- Ran Xin (ran.xin@bytedance.com)
- Zeyu Zheng (zeyuzhen@andrew.cmu.edu)