# EBFT validation config — no vLLM, uses HF generate for simplicity
# Run: CUDA_VISIBLE_DEVICES=0 axolotl train examples/ebft/llama-1b-ebft-opencode-novllm.yaml

base_model: meta-llama/Llama-3.2-1B
chat_template: llama3
rl: ebft

ebft:
  feature_layers: [0.25, 0.5, 0.75]
  embed_method: last_token
  use_whitening: false
  alignment_coef: 1.0
  diversity_coef: 1.0
  ce_coef: 0.0

trl:
  num_generations: 4
  max_completion_length: 128
  temperature: 1.0
  use_vllm: false
  scale_rewards: true
  loss_type: grpo
  epsilon: 0.2

datasets:
  - path: nvidia/OpenCodeInstruct
    type: ebft_opencode.transform
    split: train[:1%]

sequence_len: 512
micro_batch_size: 2
gradient_accumulation_steps: 2
num_epochs: 1
max_steps: 10

learning_rate: 1.0e-5
optimizer: adamw_torch_fused
lr_scheduler: cosine
warmup_steps: 2
weight_decay: 0.01

adapter: lora
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true

bf16: auto
flash_attention: true
gradient_checkpointing: true

special_tokens:
  pad_token: "<|end_of_text|>"

val_set_size: 0.0
output_dir: ./outputs/ebft-validation

wandb_project: ebft
wandb_run_id:
wandb_watch:
wandb_log_model:

logging_steps: 1
save_steps: 100