# Tiny demo config - fits RTX 3050 4GB. Scale to 1B+ only on cloud 8xH100.
model:
  arch: llama
  hidden_size: 512
  num_layers: 8
  num_heads: 8
  num_kv_heads: 4
  intermediate_size: 1376
  vocab_size: 32000
  max_seq_len: 2048
  rope_theta: 10000
tokenizer:
  type: bpe
  vocab_size: 32000
  train_files: data/processed/train.jsonl
multimodal:
  vision_encoder: openai/clip-vit-base-patch32
  projector_hidden: 512
  video_frames: 4
training:
  batch_size: 2
  grad_accum: 8
  lr: 3e-4
  epochs: 1
  output_dir: outputs/tiny-50M
gguf:
  quant: Q4_K_M
  out_file: outputs/tiny-50M.gguf