Kelnux/Qwen3-0.6B-seo-uczciweseo-GGUF

Qwen3-0.6B SEO uczciweseo.pl (GGUF) (Experimental)

Domain-specific fine-tuned version of Kelnux/Qwen3-0.6B-seo-bilingual for the uczciweseo.pl brand.

Note: This is an experimental model. The 0.6B parameter size limits domain knowledge injection - Polish text generation quality may degrade on brand-specific questions. English SEO knowledge is well preserved.

Training Pipeline

1. Stage 1: Qwen3-0.6B + bilingual SEO LoRA (r=16, 7,880 examples)
2. Stage 2: Domain adaptation LoRA (r=16, attention-only) on uczciweseo.pl brand knowledge

Stage 2 Details

• Method: LoRA (r=16, alpha=32, q_proj+v_proj only) on merged bilingual model
• Dataset: 2,312 examples (925 domain-specific 5x oversampled + 1,387 bilingual SEO control)
• Domain data: 5x oversampled from 185 unique Q&A pairs
• Mixing ratio: 60% bilingual / 40% domain
• Epochs: 3
• Learning rate: 1.5e-5 (cosine scheduler)
• Gradient clipping: 0.3
• Training loss: 1.55
• Best eval loss: 1.36

Brand Knowledge Target

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

• Company values: no long-term contracts, full transparency, client ownership of assets
• Services: SEO, Google Ads, Bing Ads, AI SEO, CRO, automation
• Industry experience: construction, legal, industrial, automotive, furniture, e-commerce
• Case studies: Elektrobim, Budmater, Higo, Sushi-sklep, ShopGracz, Tulisie, Meble Kukulka
• Educational content: choosing SEO agency, fair contracts, common scams (#nabiciwseo)

Limitations

• 0.6B model has limited capacity for domain-specific knowledge injection
• Polish text quality may degrade on brand-specific questions
• English SEO performance is well preserved
• Recommended for research/experimentation, not production use

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Kelnux/Qwen3-0.6B-seo-uczciweseo")
tokenizer = AutoTokenizer.from_pretrained("Kelnux/Qwen3-0.6B-seo-uczciweseo")

prompt = '<|im_start|>system\nJestes asystentem SEO firmy Uczciwe SEO.<|im_end|>\n<|im_start|>user\nCzym jest Uczciwe SEO?<|im_end|>\n<|im_start|>assistant\n'
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=300)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))