{
 "base_model": "Qwen/Qwen2.5-3B",
 "base_revision": "3aab1f1954e9cc14eb9509a215f9e5ca08227a9b",
 "recipe": "tinker_cookbook.recipes.math_rl.train (cli_main, stock) @f79fdb2b",
 "recipe_args": {
  "env": "gsm8k",
  "group_size": 8,
  "groups_per_batch": 8,
  "max_tokens": 512,
  "max_steps": 40,
  "seed": 0,
  "loss_fn": "importance_sampling",
  "renderer_name": "role_colon",
  "eval_every": 0,
  "save_every": 20,
  "temperature": 1.0,
  "num_substeps": 1,
  "kl_penalty_coef": 0.0
 },
 "optimizer": "recipe AdamParams(lr, beta1=0.9, beta2=0.95, eps=1e-8), constant LR",
 "arms": {
  "rank1": [
   "3e-5 (grid-edge extension)",
   "1e-4",
   "3e-4",
   "1e-3"
  ],
  "rank32": [
   "1e-4",
   "3e-4",
   "1e-3"
  ]
 },
 "notes": [
  "cookbook model_info has no Qwen2.5-3B entry, so renderer_name=role_colon (cookbook base-model renderer) was passed explicitly",
  "cookbook CLI default loss_fn is now stable_reinforce; loss_fn=importance_sampling passed to match the spec's IS loss",
  "eval_every=0 disables the recipe's in-training test-set eval (logging only; training unchanged)",
  "rank32_lr1e-4 and rank32_lr3e-4 died with 410 model_lost at steps 38/36 and were resumed by the stock recipe from their step-20 state (with optimizer); steps 20-39 re-ran",
  "wrapper train_arm.py only injects Modal proxy headers and records read-only provenance"
 ],
 "n_eval": {
  "selection": 200,
  "confirmation": 1119
 }
}