---
language: id
license: mit
tags:
- indonesian
- language-model
- klinexa-neo
---

# Klinexa Neo v1.0.1

Indonesian LLM — ~588M parameters | Morpheme-aware tokenization

## Training Info

- **Steps** : 119
- **Val loss** : 3.7412

## Load Model

```python
import torch
from model.architecture import IndonesianLLM
from tokenizer.morpheme_tokenizer import MorphemeTokenizer

tokenizer = MorphemeTokenizer(vocab_size=32000)
tokenizer.load('tokenizer.pkl')

model = IndonesianLLM(vocab_size=32000, hidden_size=1280,
    num_hidden_layers=32, num_attention_heads=20,
    intermediate_size=4096, max_position_embeddings=512)
model.load_state_dict(torch.load('KlinexaNeo.pt', map_location='cpu'))
model.eval()
```