Smart Tree - ST

# --- Configuration for RustyFlow Training and Chatting --- # This file is sourced by train.sh and chat.sh. # You can create multiple config files and pass the path as an argument. # e.g., ./train.sh my_wikitext_config.env # --- Common Settings --- # Dataset to use: tinyshakespeare, wikitext-2, short, or a path to a text file. # The active device (CPU or GPU) is determined by which script you run # (e.g., train_cpu.sh vs train_gpu.sh). DATASET="tinyshakespeare" #DATASET="wikitext-2" # Sequence length for training and context for chat. SEQ_LEN=64 # --- Training Hyperparameters --- NUM_EPOCHS=10 BATCH_SIZE=16 LEARNING_RATE=0.01 # --- Model Architecture --- EMBED_DIM=128 NUM_HEADS=4 NUM_LAYERS=2 # Path to save the trained model to, or load from for chatting. # The training and chat scripts will automatically append '-cpu.bin' or '-gpu.bin' # to this path based on which script is run. MODEL_PATH="models/${DATASET}-L${NUM_LAYERS}-H${NUM_HEADS}-E${EMBED_DIM}.bin" # --- Chat Hyperparameters --- # (SEQ_LEN from above is used for chat context, but loaded from model file) # Temperature for sampling. Lower is more deterministic, higher is more random. (e.g., 0.8) TEMPERATURE=0.8 # Top-p (nucleus) sampling. The model samples from the smallest set of tokens whose # cumulative probability exceeds p. 1.0 disables it. (e.g., 0.9) TOP_P=0.9