Kingcode12/Nusa-Ai

Nusa v1.5 candidate — base checkpoint (experimental)

Nusa is a project-authored Indonesian language-model experiment by its creator. This repository is intended to be kept private while the weights and Q&A data are served through a separate Hugging Face Space.

Critical checkpoint status

The included weights.fp16.bin and manifest.json are the v1.4 experimental expanded checkpoint stored in FP16, with 150,778,944 parameters and 301,557,888 bytes of raw weights. It is a starter checkpoint for the v1.5 training workflow, not a completed v1.5 fine-tune. The v1.5 75k curated dataset has not automatically been learned by these weights. The embedded manifest intentionally retains the underlying checkpoint version for provenance.

Files required by the Space

• manifest.json: model architecture, tensor map, tokenizer, byte count, and checkpoint SHA-256.
• weights.fp16.bin: flat FP16 checkpoint matching the manifest exactly.
• qna.json: original local knowledge/Q&A entries.
• qna-addon.json: additional curated Q&A entries.
• DATA-LICENSES.md: source and license notes carried forward from the project package.
• BASE-CHECKPOINT-REPORT.json: file integrity and provenance summary.

Architecture and use

• Custom decoder-only causal Transformer implemented in nusa_model.py in the source Space.
• 150.8M parameters; configuration includes 8 layers, 8 attention heads, model width 448, context size 512 and FFN width 20,000.
• The runtime converts FP16 storage to FP32 parameters for CPU inference.
• Model is custom; do not assume compatibility with Llama, Transformers AutoModel, or standard GGUF converters.

Evaluation status

FP16 storage conversion reduced the raw weight file from approximately 603 MB to approximately 302 MB. A short numerical test matched top-token decisions for three tested inputs, but that small check is not broad semantic evaluation. The current checkpoint should remain experimental until it is fine-tuned on curated data, compared on held-out prompts, and manually reviewed for Indonesian conversation, reasoning, arithmetic, and code quality.

Intended usage

Research and controlled experiments. Do not rely on it for high-stakes advice. Keep the repository private and expose only the inference API. Do not expose the Hub token in browser code.