vietrix/viena-tiny-demo

Viena Tiny Demo (SFT)

This is a tiny, demo-only Viena checkpoint fine-tuned for instruction following. It is not production quality. It is intended for smoke tests and workflow validation.

Model description

• Architecture: decoder-only Transformer (VienaModel) with RMSNorm, RoPE, SwiGLU, GQA.
• Parameters: ~10M (tiny config).
• Tokenizer: SentencePiece BPE (target vocab 2000; actual vocab may be smaller due to tiny data).
• Training: small offline synthetic dataset shipped with the repo.

Training data

• Pretrain: viena_data/examples/pretrain_offline.jsonl
• SFT: viena_data/examples/sft_offline_train.jsonl
• Validation: viena_data/examples/sft_offline_val.jsonl

All datasets are synthetic and intended for offline tests.

Training recipe (tiny)

• Config: configs/viena_tiny.yaml
• Pretrain: 50 steps
• SFT: 20 steps

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "vietrix/viena-tiny-demo"
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    device_map="auto",
)

prompt = "<|system|>
You are Viena.
<|user|>
Xin chao!
<|assistant|>
"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=128, do_sample=True, temperature=0.7, top_p=0.9)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Limitations

• Very small dataset and very few steps.
• Not suitable for real use or evaluation.
• Likely to hallucinate or be inconsistent.

License

MIT (code + demo weights). See repository license for details.