qwen3-tiny-v2-finetuned/README.md

license, language, base_model, tags, pipeline_tag, library_name, quantized_by

license	language	base_model	tags	pipeline_tag	library_name	quantized_by
apache-2.0	en	g023/qwen3-tiny-v2	qwen3 gguf q8_0 finetuned grpo lora-merged text-generation	text-generation	llama.cpp	g023

Qwen3-g023-tiny-v2-FT-Q8_0 - GRPO Finetuned Q8_0 GGUF Export

https://huggingface.co/g023/qwen3-tiny-v2-finetuned/

Q8_0 GGUF export of a GRPO finetuned Qwen3 model to achieve improved reasoning and reduced repetition. Original SRC Model: https://huggingface.co/g023/qwen3-tiny-v2

THIS IS A WIP (WORK IN PROGRESS)

Files

• Qwen3-g023-tiny-v2-FT-Q8_0.gguf: Q8_0 GGUF model (~1.81 GB)
• Modelfile: Ollama template + tested default sampling settings
• params_best.json: Best sampled parameters from automated sweep
• sweep_results.json: Full sweep results and per-test outcomes

Tested Best Parameters (Default in Modelfile)

• temperature: 0.65
• top_p: 0.9
• top_k: 20
• min_p: 0.0
• repeat_penalty: 1.05
• presence_penalty: 0.1
• frequency_penalty: 0.1
• num_ctx: 40000

Usage (Ollama)

ollama create qwen3-g023-tiny-v2-FT-Q8_0 -f Modelfile
ollama run qwen3-g023-tiny-v2-FT-Q8_0

# thinking on
ollama run qwen3-g023-tiny-v2-FT-Q8_0 --think "Explain why the sky is blue"

# thinking off
ollama run qwen3-g023-tiny-v2-FT-Q8_0 --think=false "Explain why the sky is blue"

or pull from huggingface directly to ollama:

ollama run hf.co/g023/qwen3-tiny-v2-finetuned:Q8_0

Notes

• Template is the Qwen3-compatible template with think/no_think handling.
• If you want stricter non-thinking behavior, compare alternatives in sweep_results.json.