Tuguberk/Qwen3.5-2B-Turkish-SFT
Tuguberk/Qwen3.5-2B-Turkish-SFT is a 2.3 billion parameter language model fine-tuned from the Qwen3.5-2B base model. It is specifically optimized for Turkish instruction-following tasks, leveraging a 32768 token context length. This model demonstrates improved performance in mathematical reasoning, truthfulness, and general reasoning benchmarks for Turkish, making it suitable for applications requiring robust Turkish language understanding and generation.
Loading preview...
Model Overview
Tuguberk/Qwen3.5-2B-Turkish-SFT is a 2.3 billion parameter language model built upon the Qwen3.5-2B base model. It has been fine-tuned using LoRA (bf16) with the AlicanKiraz0/Turkish-SFT-Dataset-v1.0, a general-purpose Turkish instruction-following dataset comprising 5,579 examples. The fine-tuning process, conducted over 3 epochs, significantly reduced the training loss from 1.47 to 0.82.
Key Capabilities & Performance
This model is specifically optimized for Turkish instruction-following. Evaluation using lm-evaluation-harness-turkish shows notable improvements over its base model:
- Mathematical Reasoning (GSM8K-TR): Achieved a +3.34 point increase, reaching 41.38%.
- Truthfulness (TruthfulQA): Improved by +1.63 points to 49.08%.
- General Reasoning (ARC-TR, HellaSwag-TR): Saw increases of +1.20 and +1.53 points respectively.
- Catastrophic Forgetting: General knowledge (MMLU) was largely preserved, indicating no significant loss of prior capabilities.
Use Cases & Limitations
This model is particularly well-suited for applications requiring accurate and contextually relevant responses in Turkish. It offers a good balance of quality and size, with Q4_K_M GGUF quantization recommended for most uses. However, due to training on a relatively small dataset (5.5K examples), its performance in complex mathematical reasoning or tasks requiring extensive chain-of-thought may be limited compared to larger, more extensively trained models. It is primarily optimized for Turkish instruction-following.