Files
M3PO-TriviaQA-bhattacharyya…/training_config.json
ModelHub XC b82a517223 初始化项目,由ModelHub XC社区提供模型
Model: Alienpenguin10/M3PO-TriviaQA-bhattacharyya-trial1-seed42
Source: Original Platform
2026-04-11 17:09:59 +08:00

30 lines
654 B
JSON

{
"num_iterations": 1,
"num_steps": 500,
"batch_size": 16,
"num_generations": 8,
"max_completion_length": 400,
"beta": 0.005,
"learning_rate": 5e-06,
"mu": 1,
"epsilon": 0.1,
"lambda_blend": 0.1,
"temperature_m3po": 0.1,
"use_m3po": true,
"gating_type": "bhattacharyya",
"gating_config": {
"temperature": 0.1,
"rank": 128,
"attn_dim": 256,
"init_strategy": "xavier",
"debug": false
},
"gating_warmup_steps": 50,
"gating_lr": 0.0005,
"gating_grad_clip": 1.0,
"gradient_accumulation_steps": 4,
"warmup_ratio": 0.1,
"seed": 42,
"trial_number": 1,
"model_name": "Qwen/Qwen2.5-1.5B-Instruct"
}