Qwen3-4B-SFT-Multitask/README.md

base_model, library_name, language, license, pipeline_tag, tags, datasets

base_model	library_name	language	license	pipeline_tag	tags	datasets
aitf-kpm-ugm/Qwen3-4B-CPT-Base	transformers	id	apache-2.0	text-generation	qwen3 bahasa-indonesia sft multitask sentiment-analysis kategorisasi alpaca lora-merged	custom

Qwen3-4B SFT — Multitask Bahasa Indonesia

Model hasil fine-tuning (LoRA, sudah di-merge ke base weights) dari aitf-kpm-ugm/Qwen3-4B-CPT-Base untuk berbagai tugas NLP Bahasa Indonesia.

---

Detail Model

Atribut	Nilai
Base model	aitf-kpm-ugm/Qwen3-4B-CPT-Base
Metode fine-tune	LoRA (r=64, alpha=128)
Status adapter	Merged ke base weights
Bahasa output	Bahasa Indonesia
Format chat	Alpaca
Precision	bfloat16
Max seq length	2048
Training epochs	3
Best checkpoint	step 2400
Best val loss	0.3566

---

Task yang Didukung

1. Sentimen Analysis — klasifikasi positif / netral / negatif
2. Justifier — klasifikasi is_relevant: true / false
3. Kategorisasi Issue — klasifikasi sub_category_label

---

Cara Pakai

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

REPO = "Atikarahmanda/Qwen3-4B-SFT-Multitask"

tokenizer = AutoTokenizer.from_pretrained(REPO)
model = AutoModelForCausalLM.from_pretrained(
    REPO,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model.eval()

messages = [
    {"role": "system", "content": "Sistem prompt sesuai task."},
    {"role": "user",   "content": "Input artikel di sini."},
]

text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
input_len = inputs.input_ids.shape[1]

with torch.no_grad():
    out = model.generate(
        **inputs,
        max_new_tokens=128,
        do_sample=False,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.pad_token_id,
    )

print(tokenizer.decode(out[0, input_len:], skip_special_tokens=True))
```

---

## Training Details

- **Framework**: Unsloth + TRL SFTTrainer
- **LoRA config**: r=64, alpha=128
- **Optimizer**: AdamW 8-bit
- **LR scheduler**: Cosine, warmup ratio 0.03
- **Effective batch size**: 6 x 8 = 48
- **train_on_responses_only**: Ya

---

## Lisensi

Mengikuti lisensi base model: [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0).