Apppublic

rob-x-ai/genesis-1b-run2-playground

sourceHugging Faceupdated 28d agoView on Hugging Face

1likes

App README

Genesis 1B

Base completion · Run 2

Genesis 1B is a 1B-parameter language model trained from scratch by Kroonen AI. This Space exposes raw Run 2 pre-training checkpoints so you can inspect how direct language completion changes across training snapshots.

Base model behavior: Genesis 1B continues text. It is not instruction tuned and is not intended to act as a chat assistant. Start a passage, function, article, or scene instead of giving it conversational instructions.

System status


Mode	Base text completion
Training series	Run 2
Checkpoint source	rob-x-ai/genesis-1b-v3-checkpoints
Checkpoint order	Newest discovered step first
Precision	bfloat16
Inference runtime	Hugging Face ZeroGPU, default large allocation: half of an NVIDIA RTX PRO 6000 Blackwell with 48 GB VRAM, dynamically allocated per request.

The Space discovers Run 2 step directories from the checkpoint repository, preloads the latest available snapshot at startup, refreshes the index when the page loads, and checks again before each generation. The status field beneath the output records the exact checkpoint used.

Explore the checkpoints

—Compare early and later checkpoints from the dropdown.
—Test prose, code, multilingual prompts, factual-style completions, and story starts.
—Adjust temperature, Top-K, Top-P, and repetition penalty to explore sampling behavior.
—Use Refresh checkpoints to discover a newly published Run 2 step.

Architecture

Specification	Value
Parameters	Approximately 1B
Architecture	Llama-style decoder-only transformer
Layers	32
Hidden / FFN dimension	1,536 / 4,736 with SwiGLU
Attention	12 query heads, 6 KV heads with grouped-query attention
Context	2,048 tokens
Tokenizer	Custom SentencePiece BPE, 49,152-token vocabulary
Position encoding	Rotary embeddings, theta 500,000
Training corpus	60B-token multilingual corpus
Run 2 training hardware	2× NVIDIA RTX 4090 over PCIe, without NVLink

Runtime

Inference uses the default large Hugging Face ZeroGPU allocation: half of an NVIDIA RTX PRO 6000 Blackwell with 48 GB VRAM, dynamically allocated per request. The model checkpoint is prepared on CPU before GPU capacity is requested for generation.