| { | |
| "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 | |
| } |