Cbgcbg/gasing-qwen3-1.7b-curriculum-v1

GASING Qwen3 1.7B - Curriculum Learning

This model was trained using curriculum learning on the GASING dataset for Indonesian mathematical reasoning.

Model Details

• Base Model: unsloth/Qwen3-1.7B
• Training Method: Curriculum Learning (6 epochs with progressive difficulty)
• Dataset: GASING (Indonesian mathematical problems)
• Fine-tuning: LoRA (r=8, alpha=32) → Merged to full weights

Training Results

• Best Training Loss: 0.0026 (Epoch 6)
• Training Strategy: Progressive difficulty curriculum

Curriculum Schedule

Epoch	Easy	Medium	Hard
1	5%	0%	0%
2	30%	65%	5%
3	10%	80%	10%
4	5%	80%	15%
5	5%	75%	20%
6	5%	70%	25%

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Cbgcbg/gasing-qwen3-1.7b-curriculum-v1",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

tokenizer = AutoTokenizer.from_pretrained(
    "Cbgcbg/gasing-qwen3-1.7b-curriculum-v1",
    trust_remote_code=True
)

# Example usage
question = "Bagaimana cara mencari panjang sisi segitiga jika diketahui sudut alpha dan sisi miringnya 1?"
messages = [
    {"role": "system", "content": "Mulai sekarang anda adalah AI Asisten bernama 'GASING'..."},
    {"role": "user", "content": question}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=2048,
        temperature=0.7,
        do_sample=True
    )

response = tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True)
print(response)

Training Configuration

• Learning Rate: 0.0001
• Batch Size: 16
• Gradient Accumulation: 8
• LoRA r: 8
• LoRA alpha: 32
• Max Sequence Length: 8192

Created by Institut Teknologi Del (IT Del)