so this is a terrible MoE based on smollm 8x360m, I don't think it even performs better than the original smollm2 360m, don't know, haven't benchmarked it.
dataset: openorca
adapter_config.json:
{
"alpha_pattern": {},
"auto_mapping": null,
"base_model_name_or_path": "/path/here/lol",
"bias": "none",
"fan_in_fan_out": false,
"inference_mode": true,
"init_lora_weights": true,
"layer_replication": null,
"layers_pattern": null,
"layers_to_transform": null,
"loftq_config": {},
"lora_alpha": 16,
"lora_dropout": 0,
"megatron_config": null,
"megatron_core": "megatron.core",
"modules_to_save": [
"gate"
],
"peft_type": "LORA",
"r": 8,
"rank_pattern": {},
"revision": null,
"target_modules": [
"w3",
"q_proj",
"w1",
"k_proj",
"gate",
"o_proj",
"v_proj",
"w2"
],
"task_type": "CAUSAL_LM",
"use_dora": false,
"use_rslora": false
}
