Views
No views yet
1from transformers import AutoModelForCausalLM, AutoTokenizer
2
3model = AutoModelForCausalLM.from_pretrained("Agytai/qwen3-4b-history_kz")
4tokenizer = AutoTokenizer.from_pretrained("Agytai/qwen3-4b-history_kz")
5
6messages = [
7 {"role": "system", "content": "Ты эксперт по Истории Казахстана."},
8 {"role": "user", "content": "Кто такой Толе би?"}
9]
10
11text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
12inputs = tokenizer(text, return_tensors="pt")
13outputs = model.generate(**inputs, max_new_tokens=256)
14print(tokenizer.decode(outputs[0], skip_special_tokens=True))| Metric | Base Model | Fine-tuned | Improvement |
|---|---|---|---|
| ROUGE-1 | 0.0% | 0.0% | 0.00% |
| ROUGE-2 | 0.0% | 0.0% | 0.00% |
| ROUGE-L | 0.0% | 0.0% | 0.00% |
| BLEU | 0.0% | 8.92% | +8.92% |
| BERTScore-P | 63.06% | 72.55% | +9.49% |
| BERTScore-R | 66.66% | 69.1% | +2.44% |
| BERTScore-F1 | 64.78% | 70.75% | +5.97% |
| Parameter | Value |
|---|---|
| Base Model | Qwen/Qwen3-4B |
| Method | SFT + LoRA |
| Epochs | 3 |
| Learning Rate | 0.0002 |
| LoRA r | 32 |
| LoRA alpha | 64 |
| Batch Size | 2 x 4 |