A LLaMA-style transformer (~989.7M params, ~0.99B) trained from scratch for the Data Scientist role. Supports up to 1M token context via RoPE with gradient checkpointing.

Architecture

Component	Value
Architecture	LLaMA-style (RoPE + RMSNorm + SwiGLU)
Parameters	~989.7M (~0.99B)
Layers	32
Heads	20
Embedding	1600
Max Context	100,000,000,000 tokens
Max Output	1,000,000 tokens
Vocab	2,025 BPE
Model Size	~4 GB (fp32)

Training

Best eval loss: 1.4600185036659241
Trained with gradient checkpointing on Apple M4 (MPS)
3 epochs, batch_size=1, grad_accum=16

Usage

python

from huggingface_hub import hf_hub_download
from tokenizers import Tokenizer

model_path = hf_hub_download("sathishphdai/data-scientist-slm-1m", "model.safetensors")
tokenizer_path = hf_hub_download("sathishphdai/data-scientist-slm-1m", "data_scientist_tokenizer.json")
tokenizer = Tokenizer.from_file(tokenizer_path)