Ma7ee7/Meet25M-Base

Meet25M Base

A small GPT-style causal language model trained from scratch.

Model

• Architecture: GPT-style decoder-only Transformer
• Approx size: ~25M parameters
• Context length: 1024
• Tokenizer: custom byte-level BPE
• Positional encoding: RoPE
• Normalization: RMSNorm
• MLP: SwiGLU
• Embeddings: tied input/output embeddings

Training Data Mix

Target pretraining mix:

• FineWeb-Edu
• FineWeb general
• Wikipedia
• OpenWebMath
• Project Gutenberg
• StackOverflow / Stack Exchange style posts
• CodeSearchNet

Total target: ~250M training tokens.

Files

• model.safetensors — safetensors checkpoint
• config.json — model config
• tokenizer/ — tokenizer files
• safetensors_info.json — checkpoint metadata

Loading

This is not a standard Transformers AutoModelForCausalLM checkpoint.
Use the custom GPT class from the training script and load model.safetensors.