agentica-org/DeepSWE-Verifier

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

<div align="center"> <span style="font-family: default; font-size: 1.5em;">DeepSWE-Verifier</span> <div> 🚀 Democratizing Reinforcement Learning for LLM Agents (RLLM) 🌟 </div> </div> <br> <div align="center" style="line-height: 1;"> <a href="https://github.com/agentica-project/rllm" style="margin: 2px;"> <img alt="Code" src="https://img.shields.io/badge/rLLM-000000?style=for-the-badge&logo=github&logoColor=000&logoColor=white" style="display: inline-block; vertical-align: middle;"/> </a> <a href="www.google.com" target="blank" style="margin: 2px;"> <img alt="Blog" src="https://img.shields.io/badge/Notion-%23000000.svg?style=for-the-badge&logo=notion&logoColor=white" style="display: inline-block; vertical-align: middle;"/> </a> <a href="https://x.com/Agentica" style="margin: 2px;"> <img alt="X.ai" src="https://img.shields.io/badge/Agentica-white?style=for-the-badge&logo=X&logoColor=000&color=000&labelColor=white" style="display: inline-block; vertical-align: middle;"/> </a> <a href="https://huggingface.co/agentica-org" style="margin: 2px;"> <img alt="Hugging Face" src="https://img.shields.io/badge/Agentica-fcd022?style=for-the-badge&logo=huggingface&logoColor=000&labelColor" style="display: inline-block; vertical-align: middle;"/> </a> </div> </div> </div>

DeepSWE-Verifier Overview

DeepSWE-Verifier is "critic model" that aids DeepSWE-Preview, a coding agent, for test-time scaling. For each SWE-Bench problem, DeepSWE-Preview generates multiple solutions, which produces multiple code patches, while DeepSWE-Verifier chooses the best code patch.Pairing DeepSWE-Preview with DeepSWE-Verifier can increases SWE-Bench-Verified score by +10% (See Figure 1, Execution-Free Verifier).

DeepSWE-Verifier is a fine-tuned/SFT version of Qwen/Qwen3-14B

Discover more about DeepSWE-Preview's development and capabilities in our technical blog post.

<div style="margin: 0 auto;"> <img src="https://cdn-uploads.huggingface.co/production/uploads/654037be97949fd2304aab7f/a7urAV3isk73ZkIbu3d7s.png" style="width: 100%;" /> <p align="center" style="margin-top: 8px; font-style: italic; color: #666;"> Figure 1: SWE-Bench Verified Performance w.r.t. different TTS strategies. With hybrid TTS, DeepSWE-Preview achieves 59%, beating the current SOTA open-weights model (SkyWork + TTS, 47%) by 12%. We note that only using execution-based and execution-free verifiers is still effective and can bring 10+% performance. </p> </div>

Usage

See our reproduction script for DeepSWE's test-time scaling.

Serving DeepSWE-Verifier

We suggest using vLLM to serve:

# Stop previous server and start verifier model
export MAX_CONTEXT_LEN=76800
vllm serve Qwen/Qwen3-14B \
    --max-model-len $MAX_CONTEXT_LEN \
    --hf-overrides '{"max_position_embeddings": '$MAX_CONTEXT_LEN'}' \
    --enable-lora \
    --lora-modules verifier=agentica-org/DeepSWE-Preview \
    --port 8000 \
    --dtype bfloat16 \
    --max-lora-rank 64 \
    --tensor-parallel-size 8

Training

Hyperparameters

The following hyperparameters were used during training:

• —learning_rate: 1e-05
• —trainbatchsize: 1
• —evalbatchsize: 8
• —seed: 42
• —distributed_type: multi-GPU
• —num_devices: 8
• —totaltrainbatch_size: 8
• —totalevalbatch_size: 64
• —optimizer: Use OptimizerNames.ADAMWTORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizerargs=No additional optimizer arguments
• —lrschedulertype: cosine
• —lrschedulerwarmup_ratio: 0.05
• —num_epochs: 2.0

Framework versions

• —PEFT 0.12.0
• —Transformers 4.51.3
• —Pytorch 2.7.1+cu126
• —Datasets 3.1.0
• —Tokenizers 0.21.2