初始化项目,由ModelHub XC社区提供模型
Model: aitf-komdigi/KomdigiUB-8B-Base Source: Original Platform
This commit is contained in:
182
README.md
Normal file
182
README.md
Normal file
@@ -0,0 +1,182 @@
|
||||
---
|
||||
base_model:
|
||||
- Qwen/Qwen3-8B
|
||||
library_name: peft
|
||||
pipeline_tag: text-generation
|
||||
anguage:
|
||||
- id
|
||||
tags:
|
||||
- base_model:Qwen/Qwen3-8B
|
||||
- lora
|
||||
- sft
|
||||
- transformers
|
||||
- trl
|
||||
- lm-eval
|
||||
- bakat
|
||||
- indonesian
|
||||
license: apache-2.0
|
||||
datasets:
|
||||
- internal-curated
|
||||
---
|
||||
|
||||
# Bakat-8B-Base
|
||||
|
||||
## Model Details
|
||||
|
||||
### Model Description
|
||||
|
||||
**Bakat-8B-Base** adalah base model bahasa Indonesia yang dirancang untuk **Continued Pre-Training (CPT)** pada domain kebijakan dan pengawasan ruang digital. Model ini dibangun di atas arsitektur **Qwen3-8B**, dengan pendekatan **LoRA (Low-Rank Adaptation)** dan **4-bit quantization** untuk efisiensi memori dan komputasi.
|
||||
|
||||
* **Developed by**: Tim 1 AITF
|
||||
* **Model type**: Causal Language Model (LoRA Adapter)
|
||||
* **Base architecture**: Qwen3-8B
|
||||
* **Primary language**: Indonesian (id)
|
||||
* **License**: Apache-2.0
|
||||
|
||||
---
|
||||
|
||||
## Training Data Composition
|
||||
|
||||
| Kategori | Elemen | Jumlah Token (M) | Persentase |
|
||||
| ---------------- | ----------------------------------------------------------------------------------------------------- | ---------------- | ---------- |
|
||||
| **DTP** | Okupasi PON TIK, Tren Pekerjaan, Kompetensi & SDM, Kebijakan & Regulasi DTP, Teknologi Digital Talent | 94 | 43.9% |
|
||||
| **PRD** | Judi Online, Hoax, Perlindungan Anak, Konten Edukasi, Kebijakan & Regulasi PRD, Kekerasan Masyarakat | 92 | 42.9% |
|
||||
| **Wikipedia ID** | Pengetahuan Umum Berbahasa Indonesia | 28.2 | 13.2% |
|
||||
| **Total** | – | **214.2** | **100%** |
|
||||
|
||||
---
|
||||
|
||||
## Intended Use
|
||||
|
||||
### Direct Use (Recommended)
|
||||
|
||||
Model ini **ditujukan untuk Continued Pre-Training**, khususnya untuk:
|
||||
|
||||
* Adaptasi domain kebijakan publik dan regulasi digital
|
||||
* Pengayaan pengetahuan spesifik Indonesia
|
||||
* Pre-adaptation sebelum Instruction Tuning atau SFT
|
||||
|
||||
### Out-of-Scope Use
|
||||
|
||||
* **Long-context conversations** (belum dioptimalkan)
|
||||
* **High-stakes decision making** (legal, medis, finansial)
|
||||
* **Chat-oriented instruction following** tanpa fine-tuning lanjutan
|
||||
|
||||
---
|
||||
|
||||
## Bias, Risks, and Limitations
|
||||
|
||||
* Dataset didominasi oleh domain kebijakan dan pengawasan ruang digital, sehingga bias topikal dapat muncul pada domain non-terkait.
|
||||
* Model belum melalui tahap preference alignment (RLHF/DPO).
|
||||
* Konten Wikipedia digunakan sebagai penyeimbang, namun tidak menjamin netralitas penuh.
|
||||
|
||||
Pengguna disarankan melakukan evaluasi tambahan sebelum penggunaan produksi.
|
||||
|
||||
---
|
||||
|
||||
## Recommendations
|
||||
|
||||
* Gunakan **Qwen3 chat template** untuk hasil generasi terbaik.
|
||||
* Lakukan **Instruction Fine-Tuning** atau **Preference Tuning** sebelum deployment ke end-user.
|
||||
* Verifikasi keluaran model untuk informasi kritikal.
|
||||
|
||||
---
|
||||
|
||||
## How to Get Started
|
||||
|
||||
Load the model using **HuggingFace Transformers**:
|
||||
|
||||
```python
|
||||
import torch
|
||||
from transformers import AutoTokenizer, AutoModelForCausalLM
|
||||
|
||||
# 1. Configuration
|
||||
model_id = "aitfindonesia/Bakat-8B-Base" # Replace with your actual Hub ID
|
||||
|
||||
# 2. Load Model
|
||||
# Use bfloat16 for A100/A10G, float16 for T4
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
model = AutoModelForCausalLM.from_pretrained(
|
||||
model_id,
|
||||
torch_dtype=torch.bfloat16,
|
||||
device_map="auto"
|
||||
)
|
||||
|
||||
# 3. Inference Example (Completion)
|
||||
input_text = "Strategi utama untuk mengurangi gap talenta digital di Indonesia adalah"
|
||||
inputs = tokenizer(input_text, return_tensors="pt").to("cuda")
|
||||
|
||||
with torch.no_grad():
|
||||
outputs = model.generate(
|
||||
**inputs,
|
||||
max_new_tokens=100,
|
||||
do_sample=True,
|
||||
temperature=0.7
|
||||
)
|
||||
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Training Details
|
||||
|
||||
### Training Data
|
||||
|
||||
* **Total size**: ~214M tokens
|
||||
* **Domains**: Digital Talent Policy (DTP), Pengawasan Ruang Digital (PRD), Wikipedia Indonesia
|
||||
* **Split**: Train (90%) / Validation (10%)
|
||||
|
||||
### Training Procedure
|
||||
|
||||
Model dilatih menggunakan **Continued Pre-Training (CPT)** dengan LoRA pada HuggingFace Transformers.
|
||||
|
||||
#### Hyperparameters
|
||||
|
||||
* **Precision**: bf16 (mixed precision)
|
||||
* **Quantization**: 4-bit (nf4)
|
||||
* **LoRA Rank (r)**: 8
|
||||
* **LoRA Alpha**: 16
|
||||
* **Target modules**: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
|
||||
* **Batch size**: 4 / device
|
||||
* **Gradient accumulation**: 16 (effective batch size = 32)
|
||||
* **Learning rate**: 2e-4 (linear schedule)
|
||||
* **Warmup ratio**: 0.03
|
||||
* **Epochs**: 1
|
||||
* **Optimizer**: adamw_8bit
|
||||
|
||||
---
|
||||
|
||||
## Evaluation
|
||||
|
||||
### Results
|
||||
|
||||
* **Final Training Loss**: ~1.2685
|
||||
* **Final Validation Loss**: ~1.264
|
||||
* **Training Perplexity**: ~3.56
|
||||
* **Validation Perplexity**: ~3.55
|
||||
|
||||
### Benchmark (General)
|
||||
|
||||
* **MMLU**: ~74.20
|
||||
* **IndoMMLU**: ~65.66
|
||||
* **XCOPA-ID**: ~75.80
|
||||
|
||||
---
|
||||
|
||||
## Environmental Impact
|
||||
|
||||
Estimasi emisi karbon mengikuti metodologi Lacoste et al. (2019).
|
||||
|
||||
* **Hardware**: NVIDIA A100 80GB
|
||||
* **Training time**: ~36 jam
|
||||
* **Compute region**: Indonesia
|
||||
* **Infrastructure**: University / Private Server
|
||||
|
||||
---
|
||||
|
||||
## Framework Versions
|
||||
|
||||
* Transformers: 4.x
|
||||
* PyTorch: 2.x
|
||||
* Datasets: 2.x
|
||||
* Tokenizers: 0.x
|
||||
Reference in New Issue
Block a user