Skip to content
← Projects

01 / Local intelligence

Jazari · a Turkish language model

A Turkish adaptation of Qwen3.5-4B, with published weights and evaluation notes on Hugging Face.

Open model · Updated · October 2026

My contribution

Training, evaluating and publishing the Turkish adaptation.

The problem

Using a general model for Turkish tasks means checking its failure modes as well as its fluency.

My decisions

  1. Use continued pretraining on Turkish text, then supervised fine-tuning.

  2. Publish the weights alongside evaluation notes so the experiment can be inspected.

  3. Document weaknesses including overconfidence, translation and mathematics.

Working result

The model weights, tokenizer and configuration are available to download. The model card also documents its use in the ailm Linux monitoring tool.

Model-card tests do not establish universal accuracy. The record below verifies published files; it is not a new inference test.

Evidence & sources

Read-only public model inspection · 2026-10-02
GET https://huggingface.co/api/models/mahsum/jazari-4b-sft-tr
Selected response fields (JSON):
{
  "id": "mahsum/jazari-4b-sft-tr",
  "sha": "83a239fdb7f8277b1a8555de9244a1962f5b8ce1",
  "pipeline_tag": "text-generation",
  "files": [
    "config.json",
    "model.safetensors",
    "tokenizer.json"
  ]
}
2 October 2026: read-only inspection of public files. This is source evidence, not an application run or performance test. Text record ↗

Qwen3.5-4B / Continued pretraining / Supervised fine-tuning / Hugging Face