{
  "advbench_path": "/home/fs01/smw362/mars/distill/SOLD/experiments/alpaca_smollm/data_advbench/advbench.jsonl",
  "ce_target_source": "reference",
  "ce_weighting": "uniform",
  "cvar_quantile": 0.25,
  "dataset_path": "/home/fs01/smw362/mars/distill/SOLD/experiments/alpaca_smollm/data",
  "ema_beta": null,
  "ema_target_source": "frozen",
  "encoding_sampler": "mdd",
  "eval_holdout_count": 200,
  "eval_holdout_seed": 0,
  "grad_accum": 8,
  "harmful_fraction": 0.286,
  "harmful_mix": "mixed",
  "kl_direction": "forward",
  "learning_rate": 1e-05,
  "log_every": 10,
  "max_grad_norm": 1.0,
  "max_new_tokens": 128,
  "max_prompt_chars": 400,
  "max_steps": 700,
  "mode": "reference",
  "model_kind": "instruct",
  "model_name": "allenai/OLMo-2-1124-7B-Instruct",
  "mse_layers": [
    8,
    16,
    24,
    32
  ],
  "mse_positions": "continuation",
  "num_encodings": 8,
  "objective": "mse_lastpos",
  "optim": "paged_adamw_8bit",
  "output_dir": "/home/fs01/smw362/mars/distill/SOLD/outputs/alpaca_smollm_olmo/instruct/inv_mse_ref_multi",
  "prefix_tokens": 8,
  "reference_device": null,
  "reference_model": "allenai/OLMo-2-1124-7B-Instruct",
  "save_steps": 175,
  "seed": 42,
  "stochastok_p": 0.3,
  "vocab_cache_file": null,
  "warmup_steps": 20
}