Instructions to use AksaraLLM/aksarallm-1.5b-native with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use AksaraLLM/aksarallm-1.5b-native with Transformers:

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

pipe = pipeline("text-generation", model="AksaraLLM/aksarallm-1.5b-native")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("AksaraLLM/aksarallm-1.5b-native")
model = AutoModelForCausalLM.from_pretrained("AksaraLLM/aksarallm-1.5b-native")
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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use AksaraLLM/aksarallm-1.5b-native with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AksaraLLM/aksarallm-1.5b-native"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AksaraLLM/aksarallm-1.5b-native",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/AksaraLLM/aksarallm-1.5b-native

SGLang

How to use AksaraLLM/aksarallm-1.5b-native 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 "AksaraLLM/aksarallm-1.5b-native" \
    --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": "AksaraLLM/aksarallm-1.5b-native",
		"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 "AksaraLLM/aksarallm-1.5b-native" \
        --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": "AksaraLLM/aksarallm-1.5b-native",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use AksaraLLM/aksarallm-1.5b-native with Docker Model Runner:
```
docker model run hf.co/AksaraLLM/aksarallm-1.5b-native
```

Ezekiel999 commited on 5 days ago

Commit

d823769

verified ·

1 Parent(s): 8e8fb99

[Devin Audit] update model card with measured baseline metrics + honest framing

Browse files

See https://huggingface.co/datasets/AksaraLLM/audit-2026-04 (or the AUDIT_REPORT.md attached) for methodology and the full per-model eval results.

Files changed (1) hide show

README.md +75 -0

README.md ADDED Viewed

	@@ -0,0 +1,75 @@

+---
+language:
+- id
+- en
+license: apache-2.0
+tags:
+- aksarallm
+- indonesian
+- llama
+- from-scratch
+- text-generation
+pipeline_tag: text-generation
+library_name: transformers
+---
+# aksarallm-1.5b-native
+The first **fully from-scratch** AksaraLLM 1.5B model (2.04B actual params),
+LLaMA-style architecture. Where the `AksaraLLM-Qwen-1.5B*` line is descended
+from Qwen2, this checkpoint contains no inherited weights — it was trained
+from random init on AksaraLLM's own corpus and tokenizer.
+## Measured baseline (Devin audit, CPU bf16, 50 short Indonesian sentences)
+| Metric | Value |
+|---|---|
+| Perplexity | **113.5** (much higher than Qwen-derived models, see below) |
+| English-stopword ratio in ID-prompted output | 0.0% |
+| Indonesian-stopword ratio in ID-prompted output | **31.3%** (highest of any AksaraLLM model — most Indonesian-saturated) |
+| Parameters | 2039.0 M |
+| Architecture | LlamaForCausalLM |
+| Vocabulary | 151 665 |
+## Why the high perplexity?
+This model started from random init and has been trained on a smaller
+corpus than the Qwen2-derived models, which began with ~5 T tokens of pretraining
+already baked in. PPL ≈ 113 reflects "model is converging on Indonesian
+distribution but not fully there yet". The very high Indonesian-word
+ratio (31%) and zero English leak suggest the model is producing
+Indonesian-only output even when uncertain — a useful signal that the
+language identity is correctly trained, but the lexical / factual quality
+is below the Qwen-derived models.
+This is the **honest from-scratch baseline** for the AksaraLLM project. It is
+the right reference point when measuring how much value continued
+pretraining / from-scratch with a larger corpus delivers (which is exactly
+what the planned 20B aims to demonstrate).
+## Loading notes
+The checkpoint contains legacy `rope.sin_cached` and `rope.cos_cached`
+keys that are unexpected by HF's `LlamaForCausalLM`; HF silently drops
+them on load — this is benign. Same `tie_word_embeddings` config / checkpoint
+mismatch as the Qwen variants; recommend setting `tie_word_embeddings: false`
+in `config.json`.
+## Quickstart
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+import torch
+tok = AutoTokenizer.from_pretrained("AksaraLLM/aksarallm-1.5b-native")
+model = AutoModelForCausalLM.from_pretrained(
+    "AksaraLLM/aksarallm-1.5b-native",
+    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=120, do_sample=True, top_p=0.9)[0], skip_special_tokens=True))
+```
+## License
+Apache 2.0