Tuğberk AkbulutGet in touch
Back to work

Qwen3.5-2B-Turkish-SFT

2025

A Turkish-tuned small language model that gains ground on benchmarks without forgetting what it knew.

UnslothTRLLoRAHugging FaceGGUF
Live

Fine-tuned Qwen3.5-2B with LoRA on Turkish instructions, chasing a specific and harder-than-it-sounds goal: measurable gains on Turkish benchmarks without catastrophic forgetting of the base model’s general ability.

Results held up: GSM8K-TR +3.34, TruthfulQA +1.63, ARC-TR +1.20, HellaSwag-TR +1.53 over base, with no regression on MMLU-TR. Published as both Safetensors and GGUF quantizations, plus the LoRA adapters on their own, so the tuning is usable whether someone wants the full merged model or just the delta.

GSM8K-TR +3.34 TruthfulQA +1.63 HellaSwag-TR +1.53 ARC-TR +1.20 MMLU-TR ±0.0, no regression