G1-nano-instruct

A 60M-parameter GPT pretrained and instruction-tuned entirely on a single 8GB-RAM NVIDIA Jetson device, without cloud infrastructure or multi-GPU setups. Chat-oriented checkpoint with a native 2048-token context and real multi-turn conversation support.

Overview

G1-nano-instruct is the instruction-tuned version of G1 Nano. Both pretraining and supervised fine-tuning were performed under an 8GB unified-memory constraint.

It is not a larger or universally better successor to G0 Nano. The model keeps roughly the same parameter count and compute budget while prioritizing a native 2048-token context and multi-turn conversation support.

Model variants

The raw pretrained version of the same model is available as G1-nano-base.

What this version adds

Compared with G1-nano-base, this checkpoint adds supervised instruction fine-tuning, a chat format and real multi-turn conversation support.

Compared with the G0 Nano design, the main differences are the native 2048-token context and the multi-turn fine-tuning recipe, while keeping a similar model size and hardware budget.

Architecture

Llama-style decoder-only Transformer:

Property	Value
Parameters	60.0M, with embeddings shared with the language-model head
Layers	14
Hidden size	576
Attention	Grouped-Query Attention, 9 query heads / 1 key-value head, head dimension 64
Position encoding	RoPE, θ=10000
Feed-forward network	SwiGLU, hidden dimension 1664
Normalization	RMSNorm
Context length	2048 tokens, used natively in pretraining and fine-tuning
Vocabulary	16,388 tokens: 16,384 SentencePiece tokens plus 4 chat tokens

Training

Pretraining data: approximately 1.5B tokens of English web and book text
Sources: FineWeb-Edu, BookCorpus, OpenWebText, PG-19 and WikiHow
Instruction tuning: approximately 39K conversations from UltraChat, OpenAssistant/oasst1 and cleaned Alpaca
Conversation format: multi-turn conversations, packed without cross-conversation leakage
Objective: causal next-token prediction followed by supervised instruction fine-tuning
Training hardware: a single NVIDIA Jetson with 8GB of unified memory

Usage

Hugging Face Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "AZERDSQ/G1-nano-instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)

messages = [
    {"role": "user", "content": "What is the capital of France?"},
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_new_tokens=50,
    do_sample=True,
    top_k=50,
    temperature=0.8,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

trust_remote_code=True is required because this repository uses a custom Transformer implementation.

Ollama

ollama run azerdsq/g1-nano-instruct

The chat format is built into the model. Conversation history can be carried across turns.

Benchmarks

Limitations

60M parameters impose a hard limit on factual knowledge; expect fluent but frequently incorrect answers on knowledge-intensive prompts.
Maximum context length is 2048 tokens, which remains short compared with modern language models.
English-only training data.
Multi-turn support does not imply long-term memory or reliable conversation consistency.
Single-sequence generation only; padded batched inference is not supported by the custom model code.

This model should not be used for high-stakes decisions, factual verification, medical advice, legal advice or autonomous actions.

License

Apache 2.0. This release contains model weights and the code required to load them; it does not include the training data or private training infrastructure.

AZERDSQ
/

G1-nano-instruct