22 lines
1.0 KiB
JSON
22 lines
1.0 KiB
JSON
|
{
|
|||
|
|
"job_type": "instruction_refinement_batch",
|
|||
|
|
"dataset": {
|
|||
|
|
"input_path": "./train.json",
|
|||
|
|
"output_path": "./train_refined.json",
|
|||
|
|
"template" : "./chat_template/chat_template_kd.jinja"
|
|||
|
|
},
|
|||
|
|
"models": {
|
|||
|
|
"teacher": "teacher/Qwen/Qwen2.5-7B-Instruct/"
|
|||
|
|
},
|
|||
|
|
"inference": {
|
|||
|
|
"prompt" : "Assume you are a prompt re-writing expert. Given an instruction as input, you should generate a new instruction semantically similar to the input to support the training of large language models. Transform the input raw prompt into a detailed prompt that comprehensively captures the user’s request. Make sure to maintain the original intent while significantly enhancing clarity and depth. You should place your answer enclosed within <answer></answer> tags. The input prompt is as follows:",
|
|||
|
|
"enable_chunked_prefill": true,
|
|||
|
|
"seed": 777,
|
|||
|
|
"gpu_memory_utilization": 0.9,
|
|||
|
|
"temperature": 0.8,
|
|||
|
|
"trust_remote_code": true,
|
|||
|
|
"enforce_eager": false,
|
|||
|
|
"max_model_len": 4096,
|
|||
|
|
"max_new_tokens": 512
|
|||
|
|
}
|
|||
|
|
}
|