TaoNet-pico-A2-instruct

TaoNet Pico A2 Instruct

Model summary

TaoNet Pico A2 Instruct is an instruction-tuned TaoNet A2 variant built on the same hardware and TaoTrain framework as the earlier checkpoints. The model was first pretrained on TaoData, then supervised fine-tuned on TaoChat to improve instruction following and conversational behavior. It uses the TaoNet MLA + RoPE architecture and keeps the small-model design goal of the original project.

Generation

Transformers

import torch
from transformers import AutoModelForCausalLM

model_id = "Lobakkang/TaoNet-pico-A2-instruct"
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.bfloat16 if device == "cuda" else torch.float32

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype=dtype,
).to(device)
model.eval()

Sample inference

import sys
from pathlib import Path

import torch
from transformers import AutoModelForCausalLM

ROOT = Path(__file__).resolve().parent
SRC = ROOT / "src"
if str(SRC) not in sys.path:
    sys.path.insert(0, str(SRC))

from taoTrain.inference.loading import load_tokenizer

MODEL_ID = "Lobakkang/TaoNet-pico-A2-instruct"
TOKENIZER_PATH = ROOT / "tokenizer" / "tokenizer.model"

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

tokenizer = load_tokenizer(TOKENIZER_PATH)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    trust_remote_code=True,
    torch_dtype=dtype,
).to(device)
model.eval()

prompt = "<user>Hello world<assistant>"
input_ids = tokenizer.encode(prompt, return_tensors="pt").to(device)

with torch.no_grad():
    output_ids = model.generate(
        input_ids=input_ids,
        max_new_tokens=32,
        do_sample=True,
        temperature=0.7,
        top_p=0.95,
    )

print(tokenizer.decode(output_ids[0], skip_special_tokens=True))

Benchmarks

These results are from the instruction-tuned checkpoint:

Limitations

This checkpoint is best treated as a small instruction-tuned research model rather than a production assistant.

• It can still miss constraints, follow ambiguous prompts poorly, or produce incomplete answers.
• It is not a substitute for a safety-reviewed assistant or a domain-specific system.
• Output quality is expected to vary outside the data and evaluation distribution used during pretraining and SFT.
• As with other small language models, it can produce plausible but incorrect text.
• It is not guaranteed to handle long-context, tool use, or safety-critical prompts reliably.
• The model should be validated before any downstream deployment or product use.

Training

Model

• Architecture: TaoNet / MLA + RoPE
• Tokenizer: SentencePiece
• Framework: TaoTrain
• Pretraining dataset: TaoData
• SFT dataset: TaoChat

Hardware

• GPU: 1 x RTX 5090

Software

• Training framework: TaoTrain

License

MIT License.

Citation

@software{taonet_pico_a2_instruct,
  title={TaoNet Pico A2 Instruct},
  author={Felix Thian},
  year={2026},
  url={https://huggingface.co/TaoTern/TaoNet-pico-A2-instruct}
}

If you use TaoTrain, TaoData, or TaoChat in a writeup, cite the corresponding project references in addition to this checkpoint.