{
 "model": "Qwen/Qwen3.5-4B",
 "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
 "gpu": "NVIDIA B200",
 "rank": 16,
 "alpha": 32,
 "trainable_parameters": 15687680,
 "max_tokens": 4096,
 "accumulation": 4,
 "permutation_pairs": true,
 "consistency_weight": 0.1,
 "initial_checkpoint": null,
 "optimizer_state": "fresh Adam; zero-B LoRA and random pointer head",
 "epochs": 1,
 "seed": 814,
 "adapter_lr": 0.0001,
 "head_lr": 0.0005,
 "training": "uncompiled BF16 backbone; FP32 adapter and head parameters; fresh LoRA/head; frozen pretrained base",
 "dataset": "data_v3 train, <=4096 tokens",
 "state_format": "raw",
 "layout": "State:\\n{state}\\nQuestion: {instructions}\\nOptions:\\n{key}: {desc}\\n...Decision:",
 "load_seconds": 19.239903798999997,
 "temperature": 1.96
}