id: trigor/20251213-trigo-value-gpt2-l6-h64-251211
desc:
- 'data_dir: /root/data/trigoRL/selfplay/iteration1-sim40-20251211/'
paths:
  root: /home/claude/work/trigoRL
  data: /root/data
  output: /root/training
data:
  type: TGNValueDataset
  data_dir: /root/data/trigoRL/selfplay/iteration1-sim40-20251211
  max_length: 8192
  min_length: 10
  max_file_size: 10000
  tokenizer_config: {}
  train_split: '*1..48/50'
  val_split: 49/50
  loader:
    batch_size: 20
    shuffle: true
    num_workers: 0
    pin_memory: true
model:
  type: ValueCausalLoss
  config:
    model_config:
      type: GPT2CausalLM
      config:
        vocab_size: 128
        hidden_size: 64
        num_layers: 6
        num_heads: 8
        max_seq_len: 8192
        dropout: 0.1
        activation: gelu_new
        intermediate_size: 256
        layer_norm_eps: 1.0e-05
        use_cache: true
    value_head_config:
      hidden_dim: 64
      intermediate_dim: 128
      bottleneck_dim: 64
      dropout: 0.1
    lambda_policy: 1.0
    lambda_value: 0.06
    gamma: 0.99
    territory_value_factor: 1.0
    ignore_index: 0
    value_id: 3
    pad_id: 0
    label_smoothing: 0.1
training:
  epochs: 50
  learning_rate: 0.0001
  weight_decay: 0.01
  warmup_steps: 50
  max_grad_norm: 1.0
  gradient_accumulation_steps: 1
  epoch_size: null
  dtype: bfloat16
  scheduler:
    type: inverse_sqrt
    d_model: 64
    lr_mul: 500.0
  save_frequency: 1
  keep_n_checkpoints: 3
  save_mode: best
  monitor:
    field: val_loss
    mode: min
  log_frequency: 160
  wandb:
    enabled: true
    tags:
    - gpt2
    - value-head
    - dual-head
    - rl-discount
  tensorboard:
    enabled: true
  env:
    CUDA_VISIBLE_DEVICES: 6
eval:
  eval_frequency: 1
  eval_batches: 50
device: cuda
seed: 42
deterministic: true