Files
llama2_7b-chat-gsm8k-lisa-l…/finetune_config.json
ModelHub XC acd9b57949 初始化项目,由ModelHub XC社区提供模型
Model: kmseong/llama2_7b-chat-gsm8k-lisa-lr5e-5
Source: Original Platform
2026-09-01 17:16:21 +08:00

23 lines
610 B
JSON

{
"base_model": "kmseong/llama2_7b-chat-Safety-FT-lr5e-5",
"method": "LISA (Bi-State Optimization)",
"fine_tuning_type": "LoRA",
"dataset": "GSM8K (clean, no poison)",
"safety_data_path": "/root/Safety-WaRP-LLM/data/circuit_breakers_train.json",
"guide_data_num": 4994,
"rho": 1.0,
"alignment_step": 100,
"finetune_step": 900,
"num_train_samples": 7473,
"batch_size": 4,
"grad_accum": 4,
"learning_rate": 5e-05,
"weight_decay": 0.01,
"warmup_ratio": 0.1,
"epochs": 3,
"max_length": 1024,
"lr_scheduler_type": "cosine",
"lora_r": 16,
"lora_alpha": 32,
"dtype": "bf16"
}