--- 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 (50–100 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