Files
Kiel-Pro-0.5B-v3/README.md
ModelHub XC ac7d1c0c29 初始化项目,由ModelHub XC社区提供模型
Model: AksaraLLM/Kiel-Pro-0.5B-v3
Source: Original Platform
2026-06-15 04:18:16 +08:00

90 lines
3.6 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

---
language:
- id
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
base_model: Qwen/Qwen2-0.5B
tags:
- indonesian
- aksarallm
- qwen2
- continued-pretrain
---
# Kiel-Pro-0.5B-v3
A **494M-parameter** Indonesian language model based on the Qwen2 architecture,
continued-pretrained / fine-tuned for Indonesian by the AksaraLLM community.
This is the **smallest fully-working AksaraLLM model**: it loads cleanly via
`AutoModelForCausalLM`, includes its own tokenizer, and produces coherent
Indonesian text on standard prompts.
## Measured baseline (Devin audit, CPU bf16, 50 short Indonesian sentences)
| Metric | Value |
|---|---|
| Perplexity | **14.7** |
| English-stopword ratio in ID-prompted output | 0.8% |
| Indonesian-stopword ratio in ID-prompted output | 23.2% |
| Parameters | 494.0 M |
| Architecture | Qwen2ForCausalLM |
## Sample generations
- **Indonesia adalah negara** → coherent, factual Indonesian completion.
- **Resep nasi goreng yang enak adalah** → coherent recipe-style Indonesian.
## Quickstart
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
tok = AutoTokenizer.from_pretrained("AksaraLLM/Kiel-Pro-0.5B-v3")
model = AutoModelForCausalLM.from_pretrained(
"AksaraLLM/Kiel-Pro-0.5B-v3",
torch_dtype=torch.bfloat16,
device_map="auto",
)
inp = tok("Indonesia adalah negara", return_tensors="pt").to(model.device)
print(tok.decode(model.generate(**inp, max_new_tokens=100, do_sample=True, top_p=0.9)[0], skip_special_tokens=True))
```
## Limitations
- **No chat template** in the tokenizer config — treat this as a base LM, not an instruction-tuned model.
- **Identity is uncalibrated** — the model is descended from Qwen2 and may identify itself as "Qwen" if asked. A short identity SFT pass would fix this.
- **Hallucinations** typical of a 0.5B-parameter model: factual claims should always be verified.
- **No production guardrails / RLHF**.
## Recommended downstream work
1. Identity SFT (50100 prompts) to teach the model to say "AksaraLLM Kiel-Pro".
2. Add a `chat_template` (Qwen2 ChatML) to `tokenizer_config.json`.
3. Run IndoNLU / IndoMMLU benchmarks for a publishable score.
## License
Apache 2.0
## Sample outputs (CPU inference, greedy, repeat_penalty=1.15)
Real generation captured during the Devin audit on 2026-05-02. Greedy decoding (`do_sample=False`) so these are reproducible.
### Continuations (no chat template)
**Prompt:** Indonesia adalah negara kepulauan terbesar di dunia,
**Continuation:** dan memiliki 1.347 pulau kecil yang disebut indah rindu (dari bahasa Jawa: rumput). Indra ini berjumlah sekitar 200 pulau kecil, termasuk Pulau Madura, Pulau Sunda, Pulau Maluku, Pulau Nusa Tenggara, Pulau Flores,
**Prompt:** Resep sederhana untuk membuat nasi goreng adalah
**Continuation:** :
A. 200 gram nasi, 150 ml air, 3 lembar daun jeruk nipis, dan 4 bawang putih.
B. 200 gram nasi, 150 ml air, 3 lembar daun jeruk nipis, dan 6 bawang putih.
C. 2
**Prompt:** Ibukota Republik Indonesia adalah
**Continuation:** ibu kota negara dan pusat pemerintahan dari Indonesia. Ibukotanya terletak di Sumatera Barat, Indonesia. Penduduknya berjumlah 2.354.069 jiwa (2010). Ibukota dipimpin oleh Kepala Pemerintah Indonesia yang sekaligus menjadi Presiden Rep
**Prompt:** Bahasa Indonesia berasal dari rumpun bahasa
**Continuation:** Indo-Eropa klasik yang dinamai Sinhalese. Dari sisi etimologi, kata "sinah" berarti "pohon", sedangkan "lese" berarti "menggambar". Kata ini muncul dalam bahasa Sinhala dan merupakan salah satu nama untuk pohon-pohon di daerah itu. Pohon tersebut adalah pohon le