Status: Experimental. Training completed with low loss but generation quality degraded on Polish domain questions. English SEO knowledge preserved.

Training Details

Base model: Qwen3-4B (3.09B params)
Method: LoRA (r=16, alpha=32) on all linear layers
Trainable params: 33,030,144 (0.81%)
Dataset: 2,312 examples (925 domain 5x-oversampled + 1,387 bilingual SEO)
Epochs: 2
Learning rate: 2e-5 (cosine scheduler)
Training loss: 0.4322
Best eval loss: 1.014
Hardware: Apple Silicon MPS (24GB), fp16 LoRA

Brand Knowledge Target

Training data covers uczciweseo.pl (EXELMEDIA sp. z o.o.):

Company values: no long-term contracts, full transparency
Services: SEO, Google Ads, Bing Ads, AI SEO, CRO, automation
Industry experience: construction, legal, industrial, automotive, furniture, e-commerce

Known Issues

Polish domain answers show quality degradation (URL-like artifacts)
English general SEO knowledge well preserved
Recommended for research/experimentation only

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Kelnux/Qwen3-4B-seo-uczciweseo")
tokenizer = AutoTokenizer.from_pretrained("Kelnux/Qwen3-4B-seo-uczciweseo")

Kelnux/Qwen3-4B-seo-uczciweseo-GGUF

Qwen3-4B SEO uczciweseo.pl (GGUF - Experimental)

Training Details

Brand Knowledge Target

Known Issues

Usage

Kelnux/
Qwen3-4B-seo-uczciweseo-GGUF