---

base_model: ermiaazarkhalili/granite-4.0-micro-GRPO-NuminaMath-20K

tags:

• gguf
• llama.cpp
• quantized
• grpo
• reasoning

license: apache-2.0

language:

• en

pipeline_tag: text-generation

---

granite-4.0-micro-GRPO-NuminaMath-20K — GGUF

GGUF quantizations of ermiaazarkhalili/granite-4.0-micro-GRPO-NuminaMath-20K,

a GRPO (Group Relative Policy Optimization) reinforcement-learning fine-tune, converted

with llama.cpp.

| Field | Value |

|---|---|

| Source checkpoint | ermiaazarkhalili/granite-4.0-micro-GRPO-NuminaMath-20K |

| Base model | ibm-granite/granite-4.0-micro |

| Training | GRPO (group-relative RL) on mathematical chain-of-thought reasoning (NuminaMath) |

| Dataset / environment | AI-MO/NuminaMath-CoT (20K subset) |

| Quantization tool | llama.cpp convert_hf_to_gguf.py + llama-quantize |

Available quantizations

| File | Size | Notes |

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

| granite-4.0-micro-grpo-numinamath-20k.q4_k_m.gguf | 2.10 GB (recommended) | 4-bit K-quant medium; best size/quality balance |

| granite-4.0-micro-grpo-numinamath-20k.q5_k_m.gguf | 2.44 GB (balanced) | 5-bit K-quant medium; near-full quality |

| granite-4.0-micro-grpo-numinamath-20k.q8_0.gguf | 3.62 GB (largest) | 8-bit; closest to the source precision |

Recommended default: Q4_K_M. For maximum fidelity use Q8_0.

Usage

llama.cpp

# One-shot
llama-cli -hf ermiaazarkhalili/granite-4.0-micro-GRPO-NuminaMath-20K-GGUF --jinja -p "Your prompt here" -n 256

# Interactive chat
llama-cli -hf ermiaazarkhalili/granite-4.0-micro-GRPO-NuminaMath-20K-GGUF --jinja -cnv

Ollama

ollama run hf.co/ermiaazarkhalili/granite-4.0-micro-GRPO-NuminaMath-20K-GGUF:Q4_K_M

llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="ermiaazarkhalili/granite-4.0-micro-GRPO-NuminaMath-20K-GGUF",
    filename="*q4_k_m.gguf",
    n_ctx=4096,
)
out = llm.create_chat_completion(
    messages=[{"role": "user", "content": "Your prompt here"}],
    max_tokens=256,
)
print(out["choices"][0]["message"]["content"])

Intended use

Research and non-commercial experimentation. This model was RL-tuned on mathematical chain-of-thought reasoning (NuminaMath);

it is a reasoning demonstrator, not a general-purpose assistant. Verify outputs before

any downstream use.

Limitations

• GGUF quantizations carry unavoidable quality loss relative to the source weights;

prefer Q8_0 when fidelity matters.

• Inherits every limitation of the source checkpoint

(ermiaazarkhalili/granite-4.0-micro-GRPO-NuminaMath-20K).

• Optimized for mathematical chain-of-thought reasoning (NuminaMath); capability on unrelated tasks is not guaranteed.

Citation

@misc{granite_4_0_micro_grpo_numinamath_20k_gguf,
  author       = {Ermia Azarkhalili},
  title        = {granite-4.0-micro-GRPO-NuminaMath-20K — GGUF quantized},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/ermiaazarkhalili/granite-4.0-micro-GRPO-NuminaMath-20K-GGUF}}
}