You can deploy this model and start earning money today!

barismenderes/
aktapokus-gmail-v3

Name: aktapokus-gmail-v3
Availability: InStock

Qwen/Qwen2.5-3B-Instruct

Views

No views yet

Aktapokus Gmail v3 — Turkish email classification (Qwen2.5-3B, GGUF)

Qwen2.5-3B-Instruct fine-tuned with QLoRA (via Unsloth on a free Google Colab T4 GPU) to classify Turkish emails into a fixed set of categories. Quantized to GGUF (Q4_K_M) for local inference via Ollama.

Built for the Aktapokus Gmail Düzenleyici tool — a local-first digital operations assistant.

Training

Base model: unsloth/Qwen2.5-3B-Instruct-bnb-4bit
Method: QLoRA (r=16, targeting q/k/v/o/gate/up/down proj), ~1-2% of parameters trained
Data: synthetic Turkish email examples with varied instruction phrasing, so the model learns the task (classify an email) rather than one fixed prompt template
3 epochs, batch size 2, lr 2e-4

Usage with Ollama

Download Qwen2.5-3B-Instruct.Q4_K_M_V3.gguf from this repo, then in the same folder create a file named Modelfile:

FROM ./Qwen2.5-3B-Instruct.Q4_K_M_V3.gguf

ollama create aktapokus-gmail-v3 -f Modelfile

Prompt pattern the model was trained on:

Aşağıdaki e-postayı incele ve şu kategorilerden birine ata: reklam_pazarlama, bulten, fatura, is, kisisel, spam_supheli, diger.

Kimden: <gönderen>
Konu: <konu>

Sadece kategori adını yaz, başka hiçbir şey ekleme.

License

Same provisional terms as the Aktapokus core project — personal/educational use permitted, commercial use requires a separate agreement with the author.

barismenderes/aktapokus-gmail-v3

Aktapokus Gmail v3 — Turkish email classification (Qwen2.5-3B, GGUF)

Categories

Training

Usage with Ollama

License

barismenderes/
aktapokus-gmail-v3