Viena 60M (SFT)

Model details

• Developed by: Vietrix
• Model type: decoder-only causal LM (Llama-style)
• Parameters: ~60M
• Layers: 16
• Hidden size: 512
• Attention heads: 8 (KV heads: 4)
• Max sequence length: 1024
• RoPE theta: 10000
• Normalization/MLP: RMSNorm + SwiGLU
• Precision: BF16 training

Tokenizer

• SentencePiece BPE
• Target vocab in config: 32k
• Actual vocab in tokenizer.model: 2105 (trained on a small corpus)
• Note: embeddings are sized for 32k; only the first 2105 tokens are used by the tokenizer.

Training data

• Internal synthetic Vietnamese instruction/chat data.
• Train/val split: 2,000 / 200 JSONL records.
• Format: messages with roles (system/user/assistant/tool).
• PII: best-effort redaction applied during dataset preparation.

Fine-tuning procedure

• Initialized from: vietrix/viena-60m-pretrain.
• Objective: token-level cross-entropy, prompt loss disabled.
• Sequence length: 1024.
• Global batch size: 32 (batch 8 x grad_accum 4).
• Optimizer: AdamW, lr 2e-4, weight decay 0.01, cosine decay with warmup.
• Steps: 1,000.
• Validation every 200 steps (10 batches).

Intended use

• Vietnamese chat/instruction-following use cases.
• Research and prototyping; not a production-grade safety model.

Limitations

• Trained on a small synthetic corpus; may hallucinate or respond incorrectly.
• Not safety-tuned for sensitive domains.
• Tokenizer vocab is small; lexical coverage is limited.

How to use