Downloads · 30 days
87
24% of all-time downloads
intisari-indonesian-chat-v4
INTISARI/intisari-indonesian-chat-v4
intisari-indonesian-chat-v4 is a machine learning model from INTISARI. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
INTISARI-CHAT-V4 adalah LLM - LARGE LANGUAGE MODEL BAHASA INDONESIA Model chat Bahasa Indonesia berukuran kecil (~58.5M params) untuk percakapan natural casual.
All-time downloads
360
Public
Repo size
282 MB
Likes
1
Public
Repo makeup
Click a slice to open those files.
.gguf282 MB · 100%
At a glance
- License
- apache-2.0
- Model type
- llama
- Access
- Public
- Created
- Aug 14, 2026
- Updated
- Aug 16, 2026
- SHA
- 64ad2ffa
Model card
From the Hugging Face model README
INTISARI-CHAT-V4
INTISARI-CHAT-V4 adalah LLM - LARGE LANGUAGE MODEL BAHASA INDONESIA Model chat Bahasa Indonesia berukuran kecil (~58.5M params) untuk percakapan natural casual.
Demo
Coba model secara langsung:
Statistik
| Parameter | Nilai |
|---|---|
| Parameters | 58.5M |
| Training Samples | 50,047 |
| Context Length | 512 tokens |
| Epochs | 3 |
| Validation Loss | 1.236 |
| Validation PPL | 3.44 |
| Validation Accuracy | 71.0% |
Evaluasi
| Metric | Value |
|---|---|
| Avg tokens | 13.21 |
| Distinct-2 | 73.8% |
| Distinct-3 | 89.0% |
| Degeneration Rate | 3.0% |
| Topic Drift | 15.0% |
Arsitektur
vocab_size: 8196
n_layers: 20
n_heads: 8
n_kv_heads: 2
hidden_dim: 512
ffn_dim: 1408
context_length: 512
rope_theta: 10000
activation: SwiGLU
normalization: RMSNorm
Training
Model dilatih menggunakan Flatbuild dengan 50,047 samples percakapan casual Bahasa Indonesia.
Topik percakapan:
- Greeting & Opening
- Small Talk
- Question → Answer
- Follow-up
- Acknowledgement
- Empathy & Emotional Support
- Humor & Banter
- Requests & Refusal
- Topic Management
- Conversation Ending
Training History
| Epoch | Val Loss | PPL | Accuracy |
|---|---|---|---|
| 1 | 2.296 | 9.94 | 49.0% |
| 2 | 1.665 | 5.28 | 61.1% |
| 3 | 1.236 | 3.44 | 71.0% |
Turn Distribution
| Turns | Percentage |
|---|---|
| 2 turns | 64.5% |
| 3 turns | 7.0% |
| 4 turns | 22.0% |
| 5+ turns | 6.5% |
Evaluasi Generation
Prinsip Dataset
Teach the model how to converse naturally in Indonesian, not how to answer knowledge questions.
Prioritas: Naturalness > Coherence > Appropriateness > Diversity
Contoh Generation
Greeting
Prompt: <|user|>\nHalo\n<|assistant|>\n
Output: Halo, aku baik. Kamu?
Food
Prompt: <|user|>\nUdah makan?\n<|assistant|>\n
Output: Udah, kamu?
Emotion
Prompt: <|user|>\nGue capek banget\n<|assistant|>\n
Output: Parah juga ya. Banyak kerjaan?
Casual
Prompt: <|user|>\nWeekend mau ngapain?\n<|assistant|>\n
Output: Nggak tau nih. Lo?
Keterbatasan
- Model kecil (~58.5M params), kemampuan terbatas
- Tidak punya pengetahuan dunia yang luas
- Generasi masih pendek
Lisensi
Apache 2.0
Dilatih menggunakan Flatbuild
- Type
- llama
- License
- apache-2.0
- Languages
- id
- Created
- Aug 14, 2026
- Updated
- Aug 16, 2026