---
library_name: transformers
tags:
- text-generation
- trust-remote-code
- sentencepiece
- custom-architecture
pipeline_tag: text-generation
license: mit
---

## Model Summary

TaoNet-mini-A2 is a 0.5B local-first language model intended for text generation experiments, lightweight instruction following, and research on efficient custom architectures.

This release is organized as a standard Hugging Face model package, while keeping the underlying TaoNet implementation in the repository for transparent loading and export.

## Model Details

### Model Specifications

Specification | Value
--- | ---
Model name | `TaoNet-mini-A2`
Model type | Causal language model
Architecture | `TaoNetForCausalLM`
Vocabulary size | 8,192
Hidden size | 1,024
Number of layers | 16
Number of attention heads | 8
Head dimension | 128
Latent KV dimension | 768
Feed-forward dimension | 3,072
Maximum sequence length | 1,024 tokens
Dropout | 0.02
Embedding type | Factorized embedding
Rope scale | 40.0
Tokenizer | SentencePiece
Special tokens | `<UNK>`, `<BOS>`, `<EOS>`, `<PAD>`

## Hardware

### Hardware

- GPU: 1 x RTX 5090

### Software

- Training framework: TaoTrain

## Quick Start

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_NAME = "TaoTern/TaoNet-mini-A2"

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.bfloat16 if device == "cuda" else torch.float32

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    trust_remote_code=True,
    torch_dtype=dtype,
).to(device)

prompt = "Fruit is now expensive so we should"
inputs = tokenizer(prompt, return_tensors="pt").to(device)

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        max_new_tokens=64,
        temperature=0.7,
        top_p=0.85,
        repetition_penalty=1.2,
        do_sample=True,
        pad_token_id=tokenizer.pad_token_id,
        eos_token_id=tokenizer.eos_token_id,
    )

completion = tokenizer.decode(
    output_ids[0][inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)
print(completion)
```

## Benchmarks

The following scores were reported for TaoNet-mini-A2:

Benchmark | Score
--- | ---
MMLU | 0.2412
HellaSwag | 0.3162
ARC-Easy | 0.4331
ARC-Challenge | 0.2560
PIQA | 0.6137
WinoGrande | 0.5083

These numbers should be treated as a snapshot of the current checkpoint, not as a universal capability guarantee.

## Limitations

- This is a relatively small model, so it will not match larger frontier models on broad reasoning or long-horizon planning
- It may hallucinate or produce incorrect answers, especially on ambiguous prompts or tasks that require deep domain knowledge
- Outputs can be sensitive to prompt wording and generation parameters
- The model is not intended for safety-critical, legal, medical, or high-stakes decision-making without human review
- The reported benchmark scores are limited to the tasks listed above and do not describe full real-world quality

## Citation

If you use TaoNet-mini-A2 in your research or product work, please cite:

```bibtex
@software{taonet_mini_a2_2026,
  title={TaoNet-mini-A2},
  author={Felix Thian},
  year={2026},
  url={https://huggingface.co/TaoTern/TaoNet-mini-A2}
}
```

## License

This repository is released under the MIT License.

## Acknowledgments

- Hugging Face Transformers for the model-loading interface
- SentencePiece for tokenizer support
- The TaoTrain export pipeline used to package the checkpoint