Llama-3.2-3B-Instruct-GRPO-NuminaMath-10K-GGUF

Llama-3.2-3B-Instruct-GRPO-NuminaMath-10K-GGUF

GGUF quantizations of a LoRA fine-tune of meta-llama/Llama-3.2-3B-Instruct, trained with GRPO (reinforcement learning) on AI-MO/NuminaMath-CoT.

Quantized from ermiaazarkhalili/Llama-3.2-3B-Instruct-GRPO-NuminaMath-10K. See that repository for the full-precision weights.

Available quantizations

Usage

llama.cpp

huggingface-cli download ermiaazarkhalili/Llama-3.2-3B-Instruct-GRPO-NuminaMath-10K-GGUF llama-3.2-3b-instruct-grpo-numinamath-10k.q4_k_m.gguf --local-dir .
llama-cli -m llama-3.2-3b-instruct-grpo-numinamath-10k.q4_k_m.gguf -p "Explain gradient checkpointing in two sentences." -n 256

Ollama

echo 'FROM ./llama-3.2-3b-instruct-grpo-numinamath-10k.q4_k_m.gguf' > Modelfile
ollama create llama-3.2-3b-instruct-grpo-numinamath-10k-gguf -f Modelfile
ollama run llama-3.2-3b-instruct-grpo-numinamath-10k-gguf

Limitations

• No benchmark evaluation has been run on this checkpoint. The only reported numbers are training-loss observations.
• Inherits the biases, knowledge cutoff and failure modes of the base model.
• Fine-tuned on a single instruction-following dataset; behaviour outside that distribution is untested.
• LoRA adapters were merged into the base weights, so the merged model cannot be detached from this fine-tune.

Reproducing

Produced by an Unsloth + TRL LoRA training pipeline, executed non-interactively with papermill on a SLURM H100 partition.

---

Card generated from the training run's own configuration and logs by scripts/generate_hub_model_card.py.