NanoDex-1M
A 1,062,272-parameter decoder-only language model pre-trained from scratch on fineweb-edu, using the NanoDex Trainer Space.
Architecture
A standard LlamaForCausalLM decoder-only transformer — SiLU MLP, RMSNorm,
rotary position embeddings, grouped-query attention, tied embeddings, no biases —
scaled down in width and depth to fit the parameter budget.
| Parameters | 1,062,272 |
| Hidden size | 128 |
| Layers | 5 |
| Attention heads | 8 (KV: 4) |
| FFN size | 288 |
| Context length | 512 |
| Vocab | 2,048 (custom BPE trained on fineweb-edu) |
Training
| Tokens seen | 999,817,216 |
| Steps | 3,814 |
| Tokens / step | 262,144 |
| Optimizer | AdamW(0.9, 0.95) wd=0.1 clip=1.0 |
| LR schedule | warmup 2% + cosine to 10% (peak 3e-03) |
| Final loss | 3.3092 (ppl 27.4) |
| Wall time | 43.8 min |
| Trained by | @DedeProGames |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("DedeProGames/NanoDex-1M")
model = AutoModelForCausalLM.from_pretrained("DedeProGames/NanoDex-1M")
ids = tok("The mitochondria is", return_tensors="pt").input_ids
print(tok.decode(model.generate(ids, max_new_tokens=60, do_sample=True,
temperature=0.8, top_k=50)[0]))
Caveats
This is a nano-scale research artifact. At this parameter count and token budget the model learns word shapes, common collocations and a little syntax — it is not a useful assistant and its output is not factual. It exists to make "pre-train a transformer from scratch" something you can actually watch happen.