yusufbaykaloglu/qwen3.5-4b-turkish-sft

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Mar 8, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The yusufbaykaloglu/qwen3.5-4b-turkish-sft model is a 4.5 billion parameter language model developed by Yusuf Baykaloglu, fine-tuned for Turkish language tasks. Built upon the Qwen3.5-4B architecture, it leverages Supervised Fine-Tuning (SFT) on the helpsteer3-tr dataset's edit subset. This model is optimized for generating Turkish text and understanding Turkish instructions, supporting a context length of 262,144 tokens.

Loading preview...

Model Overview

The yusufbaykaloglu/qwen3.5-4b-turkish-sft is a 4.5 billion parameter language model developed by Yusuf Baykaloglu, specifically fine-tuned for the Turkish language. It is based on the Qwen3.5-4B architecture and utilizes Supervised Fine-Tuning (SFT) on the helpsteer3-tr dataset's edit subset.

Key Capabilities & Features

  • Turkish Language Proficiency: Optimized for generating and understanding Turkish text through dedicated SFT.
  • Base Architecture: Built on the robust Qwen3.5-4B model, featuring a Qwen3_5ForConditionalGeneration architecture.
  • Extended Context Length: Supports a substantial context window of 262,144 tokens, enabling processing of longer Turkish inputs.
  • Efficient Fine-Tuning: Trained using LoRA (bf16) via Unsloth, with only 0.47% of parameters being trainable, making the fine-tuning process efficient.
  • Bilingual Support: While primarily focused on Turkish, the base model's capabilities suggest potential for English understanding.

Training Details

The model was fine-tuned on 13,740 samples from the helpsteer3-tr dataset over 2 epochs, achieving a final loss of 1.0971. The training utilized an AdamW 8-bit optimizer and a cosine learning rate scheduler with 3% warmup.

Usage Considerations

Users should note that the base Qwen3.5 model has a "Thinking mode" enabled by default. For direct responses without this mode, set enable_thinking=False in the chat template application. Quantized GGUF versions are also available for more efficient deployment.