A language model tuned to write long-form fiction that doesn't read like a machine wrote it. Ozan is Turkish for storyteller.

The problem

Instruct models tell decent stories in a voice you can smell from across the room: “barely above a whisper,” “a testament to,” the reflexive “not just X, but Y.” The community calls it slop, and it comes from training on synthetic data that already contained it. The base model scores 9.47 on EQ-Bench's slop metric. Ozan scores 5.30, a 44% cut, with the cleanest repetition score in the whole comparison field and zero degenerate outputs across 96 test generations.

Slop scores across models: Ozan-v1-12B scores 5.30, lower than every other 12B tested and lower than gemma-3-27b

How it was made

Everything is public and reproducible. Six open datasets were reduced from 305K records to 17K by semantic dedup, a slop-density filter, and 13-gram decontamination against the benchmark prompts. Supervised fine-tuning ran as 4-bit QLoRA with completion-only loss, stopped early when the slop metric plateaued. Then a preference pass, DPO toward real public-domain book prose over AI prose, chosen from a measured checkpoint curve, because one checkpoint further and the gains reversed. Merged, quantized to imatrix GGUFs, shipped.

Received

It launched to the #1 trending spot on Hugging Face and held it there for a week.

“A genuine improvement on both slop reduction & writing quality.” — Sam Paech, author of EQ-Bench, the benchmark it was measured on

Try it

ollama run hf.co/arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M

Weights: Ozan-v1-12B · quants: GGUF · recipe & code: github.com/arbazsiddiqui/Ozan