Instructions to use neosantara/wader-100m-base with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use neosantara/wader-100m-base with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="neosantara/wader-100m-base", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("neosantara/wader-100m-base", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("neosantara/wader-100m-base", trust_remote_code=True, device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use neosantara/wader-100m-base with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "neosantara/wader-100m-base"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "neosantara/wader-100m-base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/neosantara/wader-100m-base

SGLang

How to use neosantara/wader-100m-base with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "neosantara/wader-100m-base" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "neosantara/wader-100m-base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "neosantara/wader-100m-base" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "neosantara/wader-100m-base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use neosantara/wader-100m-base with Docker Model Runner:
```
docker model run hf.co/neosantara/wader-100m-base
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

wader-100m-base 🐟

[English Version: README.en.md]

Model bahasa Indonesia ringan berukuran ~110 Juta parameter yang mengimplementasikan arsitektur DeepSeek-V4 (MLA, MoE, Hyper-Connections), dilatih dari nol menggunakan korpus percakapan bahasa Indonesia berskala besar. Ini adalah model dasar (pretrained base model) — untuk versi Chat/SFT, silakan gunakan neosantara/wader-100m.

Arsitektur

Model ini mengimplementasikan inovasi kunci DeepSeek-V4 pada skala miniatur:

Komponen	Detail
Parameter	~110M total (41M embedding, 69M non-embedding)
Hidden size	320
Layer	8
Attention heads	8 (1 KV head — gaya MQA)
Head dim	96 (32 RoPE + 64 NoPE)
MLA	q_lora_rank=160, o_groups=2, o_lora_rank=80
MoE	4 routed experts + 1 shared, top-2 routing
Expert FFN	SwiGLU, intermediate_size=640
Routing	sqrtsoftplus scoring, noaux_tc method
Hyper-Connections	hc_mult=4, Sinkhorn routing (2 iters)
Vocab	128.000 (Tokenizer DeepSeek-V4)
Konteks	2.048 token

Fitur DeepSeek-V4 yang Diimplementasikan

Multi-head Latent Attention (MLA): Kompresi KV cache menggunakan proyeksi laten.
Mixture of Experts (MoE): Aktivasi jarang (sparse) — hanya 2 dari 4 expert yang aktif per token.
Hyper-Connections: Status tersembunyi (hidden states) salinan ganda dengan rute Sinkhorn yang dapat dipelajari (menggantikan koneksi residual standar).
SwiGLU FFN: Blok feed-forward berkapasitas tinggi.
Grouped output projection: Rute keluaran yang terstruktur dan efisien.

Pelatihan (Training)

Dataset: neosantara/IndoTalk (396k baris penjelasan mendalam, penalaran, dan obrolan kasual berbahasa Indonesia)
Langkah (Steps): 5.000 langkah
Ukuran Batch: 8 × 4 gradient accumulation = 32 efektif
Panjang Sekuens: 2.048 token
Learning rate: 6e-4 (Cosine decay dengan warmup)
Optimizer: AdamW (β1=0.9, β2=0.95, weight_decay=0.1)
Presisi: bf16 mixed precision
Perangkat Keras (Hardware): NVIDIA GeForce RTX 5090 Blackwell (Daytona Cloud) via PyTorch Nightly cu128

Metrik Pelatihan

Metrik	Nilai
Final Train Loss	~2.83 (Cross-Entropy pada Step 5000)
Token Accuracy	~49.38% (Prediksi kata berikutnya dari awal)

Penggunaan (Usage)

import torch
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

# PENTING: Model ini mewajibkan penggunaan GPU modern (Ampere/Ada/Hopper) yang mendukung bfloat16.
# Jika dijalankan di CPU atau GPU tua (seperti T4), akan terjadi error NaN.
device = "cuda" if torch.cuda.is_available() else "cpu"

# 1. Load config and model
repo_id = "neosantara/wader-100m-base"
config = AutoConfig.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_config(config, trust_remote_code=True).to(torch.bfloat16).to(device)

# 2. Download and load weights
weights_path = hf_hub_download(repo_id, "model.safetensors")
state_dict = load_file(weights_path)
model.load_state_dict(state_dict, strict=True)
model.eval()

# 3. Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)

# 4. Generate Indonesian text completion
prompt = "Indonesia adalah negara kepulauan yang memiliki"
input_ids = tokenizer.encode(prompt, return_tensors="pt").to(device)

with torch.no_grad():
    output = model.generate(
        input_ids,
        max_new_tokens=100,
        temperature=0.7,
        top_p=0.9,
        repetition_penalty=1.1,
        pad_token_id=tokenizer.eos_token_id,
    )

print(tokenizer.decode(output[0], skip_special_tokens=True))

Keterbatasan (Limitations)

Ukuran kecil: Hanya 110 Juta parameter, didesain untuk eksperimen kilat.
Hanya Base Model: Ini adalah model dasar untuk meneruskan kalimat (text completion). Untuk interaksi chat/tanya-jawab, harap gunakan neosantara/wader-100m.
Custom architecture: Membutuhkan trust_remote_code=True.
Wajib GPU Modern: Komponen DeepSeek-V4 seperti MLA/MoE dan Hyper-Connections dikembangkan khusus untuk GPU dengan presisi bfloat16. Berjalan di CPU murni (float32) atau GPU lama (T4) akan meluapkan limit kalkulasi (overflow) yang menyebabkan NaN error.

Lisensi

Apache-2.0

Downloads last month: 36

Safetensors

Model size

0.1B params

Tensor type

F32

Model tree for neosantara/wader-100m-base

Finetunes

2 models

Quantizations

1 model

neosantara
/

wader-100m-base