YsK-dev/TurkWeb-Edu-Student-Qwen1.5B-SOTA

TurkWeb-Edu Student (Reasoning) 🇹🇷🇹🇷🇹🇷🇹🇷🇹🇷🇹🇷

A Turkish educational content scorer that generates reasoning before scoring. This is the Turkish equivalent of FineWeb-Edu classifier, but using Generative Reasoning Distillation.

How It Works

1. You send Turkish text
2. The model thinks (generates reasoning in Turkish)
3. Then outputs an educational quality score (0-5)

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("YsK-dev/TurkWeb-Edu-Student-Qwen1.5B-SOTA", dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("YsK-dev/TurkWeb-Edu-Student-Qwen1.5B-SOTA")

messages = [
    {"role": "system", "content": "You are an educational quality classifier."},
    {"role": "user", "content": "Analyze the following Turkish text for educational value (0-5):\n\n<your text>\n\nProvide your reasoning and final score."}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
enc = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**enc, max_new_tokens=300, temperature=0.1, do_sample=True)
print(tokenizer.decode(output[0][enc["input_ids"].shape[1]:], skip_special_tokens=True))

Training Details

Component	Value
Teacher	Qwen3-30B-A3B-Instruct-2507
Student	Qwen/Qwen2.5-1.5B-Instruct
Method	SFT with reasoning distillation (LoRA r=64)
Data	660K Turkish web samples from FineWeb-2
Hardware	1x NVIDIA H100 80GB
Steps	20,000