Upload config.py with huggingface_hub

config.py

@@ -0,0 +1,31 @@
+"""GPT-2 style model configuration for RTX 4070 (12GB VRAM)."""
+
+import os
+
+PROJECT_DIR = os.path.dirname(os.path.abspath(__file__))
+DATA_DIR = os.path.join(PROJECT_DIR, "data")
+CHECKPOINT_DIR = os.path.join(PROJECT_DIR, "checkpoints")
+LOG_DIR = os.path.join(PROJECT_DIR, "logs")
+TOKENIZER_PATH = os.path.join(DATA_DIR, "tokenizer.json")
+
+DATA_TEXT_PATH = os.path.join(DATA_DIR, "russian_dialogues.txt")
+MAX_SEQ_LEN = 64
+TRAIN_SPLIT = 0.9
+
+VOCAB_SIZE = 16000
+
+# Model
+EMBED_DIM = 256
+NUM_HEADS = 4
+NUM_LAYERS = 6
+FF_DIM = 1024
+DROPOUT = 0.15
+
+BATCH_SIZE = 32
+LEARNING_RATE = 5e-4
+WEIGHT_DECAY = 0.01
+WARMUP_STEPS = 50
+MAX_EPOCHS = 100
+PATIENCE = 20
+MIXED_PRECISION = False       # float32 — more stable for small models
+MAX_GRAD_NORM = 1.0