---

license: apache-2.0

language:

- en

library_name: transformers

pipeline_tag: text-generation

tags:

- text-generation

- trl

- grpo

- reinforcement-learning

- math

- reasoning

- numina

- gguf

- llama.cpp

- ollama

- lm-studio

- quantized

---

Qwen2.5-3B-GRPO-NuminaMath-100K-GGUF

GGUF quantized versions of Qwen2.5-3B-GRPO-NuminaMath-100K, fine-tuned using GRPO (Group Relative Policy Optimization) on NuminaMath-CoT.

Overview

| Property | Value |

|----------|-------|

| Developed by | Behrooz Azarkhalili |

| License | APACHE-2.0 |

| Language | English |

| Training Method | GRPO (Group Relative Policy Optimization) |

| Dataset | NuminaMath-CoT |

| Format | GGUF (llama.cpp quantized) |

Available Quantizations

This repository provides GGUF variants at multiple quantization levels (Q4_K_M, Q5_K_M, Q8_0). See the repository file list for exact filenames.

Ollama

ollama pull hf.co/ermiaazarkhalili/Qwen2.5-3B-GRPO-NuminaMath-100K-GGUF:Q4_K_M

llama.cpp

llama-cli -m qwen2.5-3b-grpo-numinamath-100k-gguf-q4_k_m.gguf -p "Your prompt here" -n 256

Limitations

• Language: Primarily trained on English data
• Knowledge Cutoff: Limited to base model's training data cutoff
• Hallucinations: May generate plausible-sounding but incorrect information
• Safety: Not extensively safety-tuned; use with appropriate guardrails

Intended Use

• Research on language model fine-tuning
• Educational purposes and experimentation
• Prototyping — not recommended for production without additional safety measures

Citation

If you use this model in your research or applications, please cite:

@misc{azarkhalili2026_qwen2_5_3b_grpo_numinamath_100k_gguf,
    author = {Azarkhalili, Behrooz},
    title = {Qwen2.5-3B-GRPO-NuminaMath-100K-GGUF},
    year = {2026},
    publisher = {Hugging Face},
    url = {https://huggingface.co/ermiaazarkhalili/Qwen2.5-3B-GRPO-NuminaMath-100K-GGUF}
}

> To generate a citable DOI, click "Cite this model" on the model page.

Training Method

Trained using GRPO (Group Relative Policy Optimization) on NuminaMath-CoT.

Base (merged) model: Qwen2.5-3B-GRPO-NuminaMath-100K

Acknowledgments

• Hugging Face TRL Team for the training library
• llama.cpp for the GGUF quantization format
• Compute Canada / DRAC for HPC resources
• Base model developers for making their weights openly available