name: flatbot-mini-35M

description: >
  Optimized 35M parameter conversational assistant.
  12 layers, 512 hidden, 12 heads (4 KV).
  Based on demo-chat training recipe that achieved 0.04 loss on 4M model.

output_dir: outputs

dataset:
  format: jsonl
  path: dataset.jsonl
  field_mapping: {}
  train_split: 0.95
  val_split: 0.05
  test_split: 0.0
  max_samples: 10000
  max_length: 384
  seed: 42
  sample_type: conversation

tokenizer:
  source: train
  vocab_size: 1024
  min_frequency: 2
  added_tokens: []

chat_template:
  system: You are Flatbot, a friendly conversational assistant. If you do not know something, say "I don't know" or "I'm not sure" — do not make things up. Be clear, helpful, and concise.
  user_prefix: "\n\n<|user|>\n"
  assistant_prefix: "\n\n<|assistant|>\n"
  end_of_turn: "<|endoftext|>"
  separator: ""

model:
  vocab_size: 1024
  n_layers: 12
  n_heads: 16
  n_kv_heads: 4
  hidden_dim: 512
  ffn_dim: 1408
  context_length: 512
  rope_theta: 10000.0
  norm: rmsnorm
  activation: swiglu
  tie_embeddings: true
  attention_dropout: 0.0
  residual_dropout: 0.0
  embedding_dropout: 0.0
  initializer_range: 0.02

optimizer:
  type: adamw
  lr: 1.0e-3
  betas: [0.9, 0.99]
  eps: 1.0e-8
  weight_decay: 0.0

scheduler:
  type: cosine
  warmup_steps: 50
  min_lr_ratio: 0.05

trainer:
  epochs: 10
  batch_size: 16
  gradient_accumulation: 2
  max_steps: null
  precision: fp32
  gradient_checkpointing: false
  seed: 42
  eval_every_n_steps: 50
  log_every_n_steps: 10
  max_grad_norm: 1.0
  early_stopping:
    enabled: false
    patience: 5
    min_delta: 0.0
    monitor: val_loss

checkpoint:
  every_n_steps: 500
  keep_last: 3
  save_final: true
  resume_from: null

export:
  format: safetensors

generate:
  prompt: "Hi! What can you do?"
  max_new_tokens: 128
  temperature: 0.7
  top_p: 0.9
  top_k: 40
  do_sample: true