Guilherme-Coimbra/dr-tulu-8b-gguf

DR-Tulu-8B GGUF (Qwen3-based)

This repository contains GGUF-converted versions of the DR-Tulu-8B model, originally based on the Qwen3 architecture. The model was converted from Hugging Face safetensors format using llama.cpp and is intended for local inference and deployment via GGUF-compatible runtimes such as llama.cpp and Ollama.

🧠 Model Overview

• —Base model: rl-research/DR-Tulu-8B
• —Architecture: Qwen3ForCausalLM (Qwen3 family)
• —Original precision: BF16
• —Context length: 40,960 tokens
• —RoPE scaling: θ = 1,000,000
• —Vocabulary size: 151,936 tokens

🔄 Conversion Pipeline

The model was converted using the following pipeline:

1. 1.Hugging Face safetensors (BF16)
2. 2.Conversion to GGUF using llama.cpp (FP16)
3. 3.Quantization to lower-bit GGUF variants using llama-quantize
4. 4.Deployment via llama.cpp / Ollama

llama.cpp commit

253ba110bcd372207ca7b0bb56f1ea10d60d53fd

💾 Model Files

⚙️ System Configuration (Reproducibility)

Operating System

• —Linux (WSL2 kernel 6.6.87)

Hardware

• —GPU: NVIDIA GeForce RTX 3050 (6GB VRAM)

CUDA

• —CUDA Version: 13.0
• —Driver Version: 581.83

Build Tools

• —GCC: 13.3.0
• —G++: 13.3.0
• —CMake: 3.28.3

Python Environment

• —Python: 3.12.3

Runtime

• —Ollama: 0.20.5

🚀 Usage

Using llama.cpp

./llama-cli -m dr-tulu-8b-f16.gguf -p "Hello!"

Using Ollama

Create a Modelfile, pointing FROM at whichever GGUF file from the table above you downloaded:

FROM ./dr-tulu-8b-q3_k_m.gguf

PARAMETER num_ctx 40960
PARAMETER temperature 0.7

Then:

ollama create dr-tulu -f Modelfile
ollama run dr-tulu

📊 Notes on Performance

• —FP16 model requires ~16GB RAM/VRAM
• —Q3KM requires ~4GB RAM/VRAM, with reduced quality vs FP16
• —Q2_K requires ~3.2GB RAM/VRAM, with a noticeable quality drop, especially on reasoning-heavy tool-use tasks

📚 Reproducibility

To ensure reproducibility, the following were recorded:

• —llama.cpp commit hash
• —system hardware specifications
• —CUDA version
• —Python environment versions
• —conversion pipeline

⚠️ Limitations

• —Large context (40K) may exceed memory constraints on low-end GPUs
• —FP16 model is not suitable for low-resource devices
• —Performance depends heavily on quantization level

📜 License

This repository inherits the license of the original DR-Tulu-8B model. Please refer to the upstream model for details.

[!NOTE] For full information, go check out the Dr Tulu paper here. We have recently (23/02/2026) updated the model, please check branches for older versions of the model. Our initial release was step_1000, our second release was step_1900. Our final release is step_4000 (main and branch). Some checkpoint steps are missing, apologies, we simply do not have those checkpoints.

<img src="https://huggingface.co/rl-research/DR-Tulu-SFT-8B/resolve/main/drtululogo.png" alt="Figure 1" width="500"/>

DR Tulu-8B

This is the RL checkpoint of DR Tulu, an open deep research agent trained on top of rl-research/DR-Tulu-SFT-8B.

This model has undergone RL training on this dataset. For more details on DR Tulu please read our [paper](https://allenai.org/papers/drtulu)!

Inference and Usage

This model has been trained for tool-use using the dr-agent-lib framework. As such, running it out of the box with HuggingFace or vLLM will not work well!

See our github for more details on installation and how to run our model. Or check out our demo!

Evaluation Results

We provide evaluation instructions in our github.

For more baselines, explanations of this table, and analysis of results, check out the Dr Tulu paper!

Intended uses & limitations

This model is licensed under Apache 2.0. It is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines.

Training

The script used to train this model can be found here.

For hyperparameter details, check out the Dr Tulu paper.

Links

• —📝 DR Tulu Paper
• —⚙️ DR Tulu demo
• —💻 DR Tulu code
• —🤖 DR Tulu collection

Citation

@article{shao2025dr,
  title={DR Tulu: Reinforcement Learning with Evolving Rubrics for Deep Research},
  author={Shao, Rulin and Asai, Akari and Shen, Shannon Zejiang and Ivison, Hamish and Kishore, Varsha and Zhuo, Jingming and Zhao, Xinran and Park, Molly and Finlayson, Samuel G and Sontag, David and others},
  journal={arXiv preprint arXiv:2511.19399},
  year={2025}
}