# Effective Qwen2.5-0.5B-Instruct UltraChat AutoRound recipe
decoder_autoround:
  method: AutoRound
  targets: Linear excluding lm_head
  iters: 200
  batch_size: 1
  weights: {num_bits: 4, type: int, symmetric: true, strategy: group, group_size: 128}
  calibration:
    dataset: HuggingFaceH4/ultrachat_200k
    split: train_sft[:512]
    preprocessing: tokenizer.apply_chat_template(messages), deterministic packed full-token spans; add_special_tokens=false
    samples: 512
    sequence_length: 1024
    shuffle: false
lm_head_gptq:
  method: GPTQ
  target: lm_head
  block_size: 128
  dampening_frac: 0.01
  actorder: static
  weights: {num_bits: 4, type: int, symmetric: true, strategy: group, group_size: 128}
decoder_target_count: 168