
Base model	`Qwen/Qwen3-0.6B`
Training data	`AI-MO/NuminaMath-CoT`
Method	LoRA GRPO (reinforcement learning) via Unsloth + TRL
License	`apache-2.0` (inherited from the base model)

Available quantizations

File	Size
`qwen3-0.6b-grpo-numinamath-100k.q4_k_m.gguf`	397 MB
`qwen3-0.6b-grpo-numinamath-100k.q5_k_m.gguf`	444 MB
`qwen3-0.6b-grpo-numinamath-100k.q8_0.gguf`	639 MB

Usage

llama.cpp

bash

huggingface-cli download ermiaazarkhalili/Qwen3-0.6B-GRPO-NuminaMath-100K-GGUF qwen3-0.6b-grpo-numinamath-100k.q4_k_m.gguf --local-dir .
llama-cli -m qwen3-0.6b-grpo-numinamath-100k.q4_k_m.gguf -p "Explain gradient checkpointing in two sentences." -n 256

Ollama

bash

echo 'FROM ./qwen3-0.6b-grpo-numinamath-100k.q4_k_m.gguf' > Modelfile
ollama create qwen3-0.6b-grpo-numinamath-100k-gguf -f Modelfile
ollama run qwen3-0.6b-grpo-numinamath-100k-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.