Novi-Nano-Instruct
Novi-Nano-Instruct is a tiny instruction-tuned causal language model developed by Novi-AI.
It is based on Novi-Nano-Base and fine-tuned on a small instruction dataset to experiment with instruction following and conversational behavior at an extremely small scale.
⚡ 1.26M parameters · 500 training examples · 256-token context
Model Details
Architecture
| Property | Value |
|---|---|
| Model type | Causal Language Model |
| Base model | Novi-AI/Novi-Nano-Base |
| Parameters | 1,258,848 |
| Vocabulary size | 8,195 |
| Context length | 256 |
| Embedding size | 96 |
| Layers | 4 |
| Attention heads | 4 |
| FFN size | 384 |
| Tensor type | F32 |
Instruction Tuning
Novi-Nano-Instruct was trained from Novi-Nano-Base using a small instruction dataset containing 510 examples.
Dataset
| Split | Examples |
|---|---|
| Training | 500 |
| Validation | 10 |
The model uses a ChatML-style format with:
<|im_start|>
<|im_end|>
Training loss was applied specifically to the assistant responses, allowing the model to focus on learning how to respond to user instructions.
Training Configuration
| Property | Value |
|---|---|
| Epochs | 5 |
| Batch size | 16 |
| Gradient accumulation | 2 |
| Effective batch size | 32 |
| Maximum sequence length | 256 |
| Learning rate | 2e-5 |
| Precision | FP32 |
| Device | CPU |
Training Statistics
The final training run produced:
| Metric | Result |
|---|---|
| Final validation loss | 5.153667 |
| Final validation perplexity | 173.0650 |
| Training examples | 500 |
| Validation examples | 10 |
| Training time | ~32 seconds |
Because the validation set contains only 10 examples, these metrics should be considered experimental rather than a comprehensive benchmark.
Tokenizer
Novi-Nano-Instruct uses the custom tokenizer developed for Novi-Nano.
The original tokenizer vocabulary was 8,192 tokens, with additional tokens already present in the tokenizer.
Two ChatML tokens were added for instruction tuning:
<|im_start|>— 8193<|im_end|>— 8194
The final tokenizer size is 8,195 tokens.
The tokenizer was originally trained using data from:
- FineWeb-Edu
- FineWeb-HQ
- SmolLM-Cosmopedia
Intended Use
Novi-Nano-Instruct is primarily intended for:
- 🔬 Research and experimentation
- 🧪 Small-model instruction-tuning experiments
- 🎓 Educational purposes
- 💬 Tiny conversational-model experiments
- 💻 Lightweight local inference
- 🛠️ Experimenting with extremely small instruction-tuned models
As an experimental 1.26M-parameter model, it is not intended to compete with modern billion-parameter language models.
Limitations
Novi-Nano-Instruct is an extremely small experimental language model trained on only 500 instruction examples.
Because of its size and limited training data, it may:
- Generate incoherent text
- Repeat phrases
- Produce unrelated responses
- Fail to follow instructions
- Produce factual errors
- Have very limited world knowledge
- Perform poorly on reasoning tasks
- Struggle with longer conversations
- Lose context beyond its 256-token window
- Produce malformed or unexpected responses
Generation quality is currently highly experimental. The model can generate text, but it does not yet consistently produce reliable assistant-style responses.
This model should be considered a research and experimentation model, rather than a production-ready conversational AI.
Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "Novi-AI/Novi-Nano-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
messages = [
{
"role": "system",
"content": "You are Novi-Nano, a helpful AI assistant."
},
{
"role": "user",
"content": "Give a synonym for 'quiet'."
}
]
prompt = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(
**inputs,
max_new_tokens=50,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Chat Template
Novi-Nano-Instruct uses a ChatML-style conversation format:
<|im_start|>system
You are Novi-Nano, a helpful AI assistant.<|im_end|>
<|im_start|>user
Give a synonym for 'quiet'.<|im_end|>
<|im_start|>assistant
A synonym is 'silent'.<|im_end|>
For generation, the assistant message is opened automatically by the chat template.
Project History
Novi AI follows the earlier AppleMind experiments, with Novi becoming the primary project for developing small language models.
AppleMind → Novi AI → Novi-Nano → Novi-Nano-Instruct 🚀
Acknowledgements
Novi-Nano was built using the open-source machine-learning ecosystem and datasets made available by the community.
Special thanks to:
- Hugging Face 🤗
- FineWeb
- SmolLM
- Cosmopedia
License
This model is released under the Apache 2.0 license.
🧠 Novi AI
Small models. Big experiments.
Novi-Nano-Instruct explores instruction tuning at an extremely small scale, with just 1.26 million parameters and 500 training examples.
It is intentionally tiny — exploring how far instruction following can go with a fraction of the parameters used by modern LLMs.
Novi AI 2026 — Project Kairo