A 48M-parameter, GPT-2-style decoder-only transformer trained from scratch as part of the Leopard AI Model Suite. Small enough to run on CPU or any GPU; built as a learning/research model, not a production assistant.
Run it instantly with Ollama:ollama run rafi-dev/rb-nano
Or grab the quantized build from the GGUF repo (rb-nano-GGUF) for llama.cpp.
It loads as a standard GPT2LMHeadModel — no trust_remote_code required.
Prompt format
Trained on a simple user: / ai: turn format, prefixed with the <sos> token:
<sos>user: hello
ai: Hi there! How can I help you today?
user: what is python?
ai:
Architecture
Type
Decoder-only transformer (GPT-2 family)
Parameters
~48M
Embedding dim (n_embd)
512
Layers
10
Attention heads
8
Context length
1024 tokens
Position embeddings
Learned
Norm / activation
LayerNorm, GELU-tanh (gelu_new)
Head
Weight-tied to token embeddings
Tokenizer
ByteLevel BPE, 32k vocab
Format
safetensors (fp32)
Training
Pretrain — FineWeb-Edu (sample-10BT), ~50M tokens. Final val loss ≈ 3.44.
Finetune — Alpaca, Alpaca-cleaned, CodeAlpaca-20k, Dolly-15k, and ShareGPT (full multi-turn threads, loss masked to assistant turns only). Final val loss ≈ 2.67.
Recommended parameters
temperature 0.7
top_k 40
top_p 0.9
repeat_penalty 1.3
Limitations
Knowledge. At 48M params the model has very limited factual knowledge and will confidently hallucinate (made-up libraries, wrong dates, etc.). It cannot be a reliable source of facts.
Coherence. Good for short exchanges; longer or more technical answers drift.
Scope. English-centric, 1024-token context. Best for demos, experimentation, and edge/CPU inference — not production use.
License / attribution
Released under CC BY-NC 4.0 (non-commercial, attribution required). The finetune mixes datasets with non-commercial terms (Alpaca, CodeAlpaca, ShareGPT — OpenAI-derived), so commercial use is not granted. Trained on publicly available datasets (FineWeb-Edu, Alpaca, Dolly, CodeAlpaca, ShareGPT); review each dataset's license before redistributing derived outputs.
Made with care
rb-nano was built by Rafi (12 years old) and Buddi (10 years old) — pretrained and finetuned from scratch on a single RTX 4070 (8 GB VRAM). It's a passion project: proof that a coherent little chat model can be trained end-to-end on consumer hardware.
If you enjoy it and want to support more experiments like this, you can buy us a coffee ☕. Thank you for trying rb-nano — we hope you like it.
