{
 "base_model": "Qwen/Qwen3.5-2B",
 "spec_base_model": "Qwen/Qwen3.5-0.8B",
 "lora_rank": 32,
 "optimizer": {
  "requested": "adam lr 1e-4",
  "sent": "tinker AdamParams(learning_rate=1e-4, beta1=0.9, beta2=0.95, eps=1e-12, weight_decay=0.0, grad_clip_norm=0.0)",
  "server_reported_optimizer_config": {
   "type": "adamw"
  }
 },
 "arms": [
  "baseline_center0_seed0",
  "treatment_center1_seed0",
  "baseline_center0_seed1",
  "treatment_center1_seed1"
 ],
 "steps": 100,
 "groups_per_step": 32,
 "group_size": 8,
 "batch_size": 256,
 "max_tokens": 512,
 "temperature": 1.0,
 "loss": {
  "CENTER=0": "-sum_t A*exp(logp - samp_logp) over completion tokens (stock importance_sampling surrogate)",
  "CENTER=1": "-sum_t A*(sg(w)*logp - correction), w=exp(logp-samp_logp) (weight_fn=identity), correction=sum_k sg(q_k*w(p_k/q_k) - alpha*p_k)*log p_k over sampler top-128, alpha=rho*w(1/rho)"
 },
 "score_centering_topk": 128,
 "updates_per_batch": 1,
 "rollout_lag": 0,
 "eval": {
  "n": 500,
  "items": "GSM8K test first 500 (test-0..test-499)",
  "every_steps": 10,
  "evaluator_id": "7c9cd38084fa1c5ab0bf60eeb477cf29955e8d344a2cef7b38ad6b93cc439a36"
 },
 "data_order": "random.Random(seed).shuffle over 7473 train rows; pilot used seed 1000 (disjoint from outcome seeds 0/1 schedules' identity)",
 "full_training_launched": false
}