模型支持情况说明
更新时间:2024-11-15
本文介绍了模型支持情况,在调用模型精调V2版本部分API时,需查看此文档各参数支持情况。
对话续写类
SFT
ERNIE系列
model | trainMode | parameterScale | hyperParameterConfig |
---|---|---|---|
ERNIE-Lite-8K-0308 | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate: FullFineTuning:[0.0000001,0.01],默认值0.00003,步长0.000001 LoRA:[0.000001,0.001],默认值0.0003,步长0.000001 (3)maxSeqLen: 单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)globalBatchSize:[1,10000],默认值16,步长4 (8)pseudoSamplingProb:[0,1],默认值0,步长0.1 (9)seed:[1,2147483647],默认值42 (10)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (11)numCycles:[0.1,0.5],默认值0.5,步长0.1 (12)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (13)power:[1,3],默认值1 (14)checkpointCount:[1,10],默认值1 (15)saveStep:[1,50000],默认为64 (16)validationStep:[0,1000000],默认值16,步长1 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank: 单选,2 或 4 或 8,默认为8 loraAllLinear: 单选,True 或 False,默认为True |
ERNIE-Lite-128K-0419 | SFT | FullFineTuning | (1)epoch:[1,50],默认值1 (2)maxSeqLen: 单选,16384 或 32768 或 65536 或 131072,默认值32768 (3)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (4)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (5)globalBatchSize:[1,10000],默认16,步长1 (6)pseudoSamplingProb:[0,1],默认值0,步长0.1 (7)seed:[1,2147483647],默认值42 (8)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (9)numCycles:[0.1,0.5],默认值0.5,步长0.1 (10)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (11)power:[1,3],默认值1 (12)checkpointCount:[1,10],默认值1 (13)saveStep:[1,50000],默认为64 (14)validationStep:[0,1000000],默认值16,步长1 (15)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 |
ERNIE-Lite-128K-0722 | SFT | FullFineTuning | (1)epoch:[1,50],默认值1 (2)learningRate: FullFineTuning:[0.0000001,0.01],默认值0.00003,步长0.000001 LoRA:[0.000001,0.001],默认值0.0003,步长0.000001 (3)maxSeqLen: 单选,16384 或 32768 或 65536 或 131072,默认值32768 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)globalBatchSize:[1,10000],默认16,步长1 (8)pseudoSamplingProb:[0,1],默认值0,步长0.1 (9)seed:[1,2147483647],默认值42 (10)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (11)numCycles:[0.1,0.5],默认值0.5,步长0.1 (12)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (13)power:[1,3],默认值1 (14)checkpointCount:[1,10],默认值1 (15)saveStep:[1,50000],默认为64 (16)validationStep:[0,1000000],默认值16,步长1 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 |
ERNIE-Speed-8K | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate: FullFineTuning:[0.0000001,0.01],默认值0.00003,步长0.000001 LoRA:[0.000001,0.001],默认值0.0003,步长0.000001 (3)maxSeqLen: 单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)globalBatchSize:[1,10000],默认值16,步长1 (7)pseudoSamplingProb:[0,1],默认值0,步长0.1 (8)seed:[1,2147483647],默认值42 (9)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (10)numCycles:[0.1,0.5],默认值0.5,步长0.1 (11)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (12)power:[1,3],默认值1 (13)checkpointCount:[1,10],默认值1 (14)saveStep:[1,50000],默认为64 (15)validationStep:[0,1000000],默认值16,步长1 (16)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (17)仅LoRA支持: loraRank: 单选,8 或 64,默认为64 loraAllLinear: 单选,True 或 False,默认为True |
ERNIE-Character-8K-0321 | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate: FullFineTuning:[0.0000001,0.01],默认值0.00003,步长0.000001 LoRA:[0.000001,0.001],默认值0.0003,步长0.000001 (3)maxSeqLen: 单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)globalBatchSize: FullFineTuning:[1,10000],默认值16,步长1 LoRA:[1,10000],默认值16,步长2 (8)pseudoSamplingProb:[0,1],默认值0,步长0.1 (9)seed:[1,2147483647],默认值42 (10)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (11)numCycles:[0.1,0.5],默认值0.5,步长0.1 (12)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (13)power:[1,3],默认值1 (14)checkpointCount:[1,10],默认值1 (15)saveStep:[1,50000],默认为64 (16)validationStep:[0,1000000],默认值16,步长1 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank: 单选,2 或 4 或 8,默认为8 loraAllLinear: 单选,True 或 False,默认为True |
ERNIE-Tiny-8K | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate: FullFineTuning:[0.0000001,0.01],默认值0.00003,步长0.000001 LoRA:[0.000001,0.001],默认值0.0003,步长0.000001 (3)maxSeqLen: 单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)globalBatchSize: FullFineTuning:[1,10000],默认值32,步长8 LoRA:[1,10000],默认值16,步长4 (8)pseudoSamplingProb:[0,1],默认值0,步长0.1 (9)seed:[1,2147483647],默认值42 (10)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (11)numCycles:[0.1,0.5],默认值0.5,步长0.1 (12)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (13)power:[1,3],默认值1 (14)checkpointCount:[1,10],默认值1 (15)saveStep:[1,50000],默认为64 (16)validationStep:[0,1000000],默认值16,步长1 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank: 单选,2 或 4 或 8,默认为8 loraAllLinear: 单选,True 或 False,默认为True |
ERNIE-4.0-Turbo-8K | SFT | LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.001],默认0.000001,步长0.000001 (3)maxSeqLen: 单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)pseudoSamplingProb:[0,1],默认值0,步长0.1 (8)seed:[1,2147483647],默认值42 (9)lrSchedulerType:单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认constant (10)numCycles:[0.1,0.5],默认值0.5,步长0.1 (11)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (12)power:[1,3],默认值1 (13)checkpointCount:[1,10],默认值1 (14)saveStep:[1,50000],默认为64 (15)validationStep:[0,1000000],默认值16,步长1 (16)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)loraRank:单选,2、4、8、16、32 或 64,默认为64 (19)loraAllLinear: 单选,True 或 False,默认为True (20)globalBatchSize:[1,10000],默认值18,步长1 |
ERNIE-Speed-Pro-128K | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate: FullFineTuning:[0.0000001,0.01],默认值0.00003,步长0.000001 LoRA:[0.000001,0.001],默认值0.0003,步长0.000001 (3)maxSeqLen:单选,16384 或 32768 或 65536 或 131072, 默认32768 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)globalBatchSize:[1,10000],默认值16,步长1 (8)pseudoSamplingProb:[0,1],默认值0,步长0.1 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[1,50000],默认值64 (11)seed:[1,2147483647],默认值42 (12)lrSchedulerType:单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (13)numCycles:[0.1,0.5],默认值0.5,步长0.1 (14)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (15)power:[1,3],默认值1 (16)validationStep:[0,1000000],默认值16,步长1 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank:单选,8 或 64,默认为64 loraAllLinear:单选,True 或 False,默认为True |
ERNIE-Tiny-128K-0929 | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate: FullFineTuning:[0.0000001,0.01],默认值0.00003,步长0.000001 LoRA:[0.000001,0.001],默认值0.0003,步长0.000001 (3)maxSeqLen:单选,16384 或 32768 或 65536 或 131072, 默认值32768 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7) globalBatchSize:[1,10000],默认值16,步长2 (8)pseudoSamplingProb:[0,1],默认值0,步长0.1 (9)seed:[1,2147483647],默认值42 (10)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (11)numCycles:[0.1,0.5],默认值0.5,步长0.1 (12)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (13)power:[1,3],默认值1 (14)checkpointCount:[1,10],默认值1 (15)saveStep:[1,50000],默认为64 (16)validationStep:[0,1000000],默认值16,步长1 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank: 单选,2 或 4 或 8,默认为8 loraAllLinear: 单选,True 或 False,默认为True |
ERNIE-3.5-8K | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate: FullFineTuning:[0.0000001,0.01],默认0.00003,步长0.000001 LoRA:[0.0000001,0.001],默认0.0003,步长0.000001 (3)maxSeqLen: 单选,512 或 1024 或 2048 或 4096 或 8192,默认值 4096 (4)globalBatchSize:[1,10000],默认64,步长1 (5)loggingSteps:1 (6)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (7)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (8)pseudoSamplingProb:[0,1],默认值0,步长0.1 ,仅FullFineTuning支持 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[1,50000],默认值64 (11)seed:[1,2147483647],默认值42 (12)lrSchedulerType: FullFineTuning:单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear LoRA:单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值constant (13)numCycles:[0.1,0.5],默认值0.5,步长0.1 (14)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (15)power:[1,3],默认值1 (16)validationStep:[0,1000000],默认值16,步长1 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank: 单选,2 或 4 或 8 或 16 或 32 或 64,默认为64 loraAllLinear: 单选,True 或 False,默认为True |
ERNIE-Character-Fiction-8K-1028 | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate: FullFineTuning:[0.0000001,0.01],默认值0.00003,步长0.000001 LoRA:[0.000001,0.001],默认值0.0003,步长0.000001 (3)maxSeqLen: 单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)globalBatchSize: FullFineTuning:[1,10000],默认值16,步长1 LoRA:[1,10000],默认值16,步长2 (8)pseudoSamplingProb:[0,1],默认值0,步长0.1 (9)seed:[1,2147483647],默认值42 (10)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (11)numCycles:[0.1,0.5],默认值0.5,步长0.1 (12)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (13)power:[1,3],默认值1 (14)checkpointCount:[1,10],默认值1 (15)saveStep:[1,50000],默认为64 (16)validationStep:[0,1000000],默认值16,步长1 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank: 单选,2 或 4 或 8,默认为8 loraAllLinear: 单选,True 或 False,默认为True |
开源系列
model | trainMode | parameterScale | hyperParameterConfig |
---|---|---|---|
Meta-Llama-3.1-8B | SFT | FullFineTuning | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (3) validationStep:[0,1000000],默认值16,步长1 (4)batchSize:[1,4],默认值1 (5)Packing:字符串,true 或 false 或 auto,默认值auto (6)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (7)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (8)weightDecay:[0.001,1],默认值0.01,步长0.001 (9)maxSeqLen: 单选,512 或 1024 或 2048 或 4096,默认值4096 (10)checkpointCount:[1,10],默认值1 (11)saveStep:[64,4096],默认值64 (12)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64,默认值32 loraAlpha:单选,8 或 16 或 32 或 64,默认值32 loraDropout:[0.01,0.5],默认值0.1,步长0.001 |
Meta-Llama-3-8B | SFT | FullFineTuning | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (3)batchSize:[1,2],默认值1 (4)validationStep:[0,1000000],默认值16, 步长1 |
Meta-Llama-3.2-1B-128K | SFT | FullFineTuning | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (3)validationStep:[0,1000000],默认值16,步长1 (4)Packing:字符串,true 或 false 或 auto,默认值auto (5)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (6)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (7)weightDecay:[0.001,1],默认值0.01,步长0.001 (8)maxSeqLen: 单选,8192 或 16384 或 32768 或 65536 或 131072,默认值8192 (9)batchSize:[1,N],默认值1,其中的 N 和 maxSeqLen 有关联,关联关系如下: maxSeqLen = 131072 时,N=1 maxSeqLen = 65536 时,N=2 maxSeqLen = 32768 时,N=4 maxSeqLen = 16384 时,N=8 maxSeqLen = 8192 时,N=16 (10)checkpointCount:[1,10],默认值1 (11)saveStep:[64,4096],默认值64 |
Qianfan-Chinese-Llama-2-1.3B | SFT | FullFineTuning | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (3)batchSize:[1,4],默认值1 (4)Packing:字符串,true 或 false 或 auto,默认值auto (5)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (6)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (7)weightDecay:[0.001,1],默认值0.01,步长0.001 (8)maxSeqLen: 单选,512 或 1024 或 2048 或 4096,默认值4096 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[64,4096],默认值64 (11)validationStep:[0,1000000],默认值16,步长1 (12)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64,默认值32 loraAlpha:单选,8 或 16 或 32 或 64,默认值32 loraDropout:[0.01,0.5],默认值0.1,步长0.001 |
Qianfan-Chinese-Llama-2-7B | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (3)batchSize:[1,64],默认值1 (4)Packing:字符串,true 或 false 或 auto,默认值auto (5)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (6)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (7)weightDecay:[0.001,1],默认值0.01,步长0.001 (8)maxSeqLen: 单选,512 或 1024 或 2048 或 4096,默认值4096 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[64,4096],默认值64 (11)validationStep:[0,1000000],默认值16,步长1 (12)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64,默认值32 loraAlpha:单选,8 或 16 或 32 或 64,默认值32 loraDropout:[0.01,0.5],默认值0.1,步长0.001 |
Qianfan-Chinese-Llama-2-7B-32K | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001, 步长0.000001 (3)batchSize:1 (4)Packing:字符串,true 或 false 或 auto,默认值auto (5)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (6)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (7)weightDecay:[0.001,1],默认值0.01,步长0.001 (8)maxSeqLen:单选,4096 或 8192 或 16384 或 32768,默认值32768 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[64,4096],默认值64 (11)validationStep:[0,1000000],默认值16,步长1 (12)saveStep:[64,4096],默认值64 (13)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64,默认值32 loraAlpha:单选,8 或 16 或 32 或 64,默认值32 loraDropout:[0.01,0.5],默认值0.1,步长0.001 |
Qianfan-Chinese-Llama-2-13B-v1 | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (3)batchSize:[1,8],默认1 (4)Packing:字符串,true 或 false 或 auto,默认值auto (5)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (6)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (7)weightDecay:[0.001,1],默认值0.01,步长0.001 (8)maxSeqLen: 单选,512 或 1024 或 2048 或 4096,默认值4096 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[64,4096],默认值64 (11)validationStep:[0,1000000],默认值16,步长1 (12)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64,默认值32 loraAlpha:单选,8 或 16 或 32 或 64,默认值32 loraDropout:[0.01,0.5],默认值0.1,步长0.001 |
Qianfan-Chinese-Llama-2-13B-v2 | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (3)batchSize:[1,8],默认1 (4)Packing:字符串,true 或 false 或 auto,默认值auto (5)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (6)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (7)weightDecay:[0.001,1],默认值0.01,步长0.001 (8)maxSeqLen: 单选,512 或 1024 或 2048 或 4096,默认值4096 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[64,4096],默认值64 (11)validationStep:[0,1000000],默认值16,步长1 (12)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64,默认值32 loraAlpha:单选,8 或 16 或 32 或 64,默认值32 loraDropout:[0.01,0.5],默认值0.1,步长0.001 |
Mixtral-8x7B | SFT | FullFineTuning | (1)epoch:[1,20],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.00001,步长0.000001 (3)batchSize:[1,4],默认1 (4)Packing:字符串,true 或 false 或 auto,默认值auto (5)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (6)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (7)weightDecay:[0.001,1],默认值0.01,步长0.001 (8)maxSeqLen: 单选,512 或 1024 或 2048 或 4096,默认值4096 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[64,4096],默认值64 (11)validationStep:[0,1000000],默认值16,步长1 (12)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64,默认值32 loraAlpha:单选,8 或 16 或 32 或 64,默认值32 loraDropout:[0.01,0.5],默认值0.1,步长0.001 |
SQLCoder-7B | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (3)batchSize:[1,4],默认值1 (4)Packing:字符串,true 或 false 或 auto,默认值auto (5)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (6)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (7)weightDecay:[0.001,1],默认值1,步长0.001 (8)maxSeqLen: 单选,512 或 1024 或 2048 或 4096,默认值4096 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[64,4096],默认值64 (11)validationStep:[0,1000000],默认值16,步长1 (12)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64,默认值32 loraAlpha:单选,8 或 16 或 32 或 64,默认值32 loraDropout:[0.01,0.5],默认值0.1,步长0.001 |
ChatGLM2-6B-32K | SFT | FullFineTuning | (1)epoch:[1,50],默认值1 (2)maxSeqLen:单选,4096 或 8192 或 16384 或 32768,默认值32768 (3)batchSize32k:1,前置条件maxSeqLen=32768 (4)batchSize16k:[1,2],默认值1,前置条件maxSeqLen=16384 (5)batchSize8k:[1,6],默认值1,前置条件maxSeqLen=8192 (6)batchSize4k:[1,12],默认值 1,前置条件:maxSeqLen=4096 (7)Packing:字符串,true 或 false 或 auto,默认值auto (8)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (9)schedulerName:单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (10)warmupRatio:[0.01,0.1],默认值0.03,,步长0.001 (11)weightDecay:[0.001,1],默认值0.01, 步长0.001 (12)checkpointCount:[1,10],默认值1 (13)validationStep:[0,1000000],默认值16,,步长1 (14)saveStep:[64,4096],默认值64 |
ChatGLM2-6B | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (3)batchSize:[1,2],默认值1 (4)Packing:字符串,true 或 false 或 auto,默认值auto (5)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (6)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (7)weightDecay:[0.001,1],默认值0.01,步长0.001 (8)maxSeqLen: 单选,512 或 1024 或 2048 或 4096,默认值4096 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[64,4096],默认值64 (11)validationStep:[0,1000000],默认值16,步长1 (12)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64,默认值32 loraAlpha:单选,8 或 16 或 32 或 64,默认值32 loraDropout:[0.01,0.5],默认值0.1,步长0.001 |
ChatGLM3-6B | SFT | FullFineTuning、LoRA | (1)epoch: FullFineTuning:[1,50],默认值3 LoRA:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (3)batchSize:16 或 32 或 64,默认值16 (4)Packing:字符串,true 或 false 或 auto,默认值auto (5)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (6)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (7)weightDecay:[0.001,1],默认值0.01,步长0.001 (8)maxSeqLen: FullFineTuning:单选,4096 或 8192,默认值4096 LoRA:单选,512 或 1024 或 2048 或 4096,默认值4096 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[64,4096],默认值64 (11)validationStep:[0,1000000],默认值16,步长1 (12)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64,默认值32 loraAlpha:单选,8 或 16 或 32 或 64,默认值32 loraDropout:[0.01,0.5],默认值0.1,步长0.001 |
Baichuan2-7B-Chat | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (3)batchSize:[1,4],默认值1 (4)Packing:字符串,true 或 false 或 auto,默认值auto (5)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (6)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (7)weightDecay:[0.001,1],默认0.01,步长0.001 (8)maxSeqLen: 单选,512 或 1024 或 2048 或 4096,默认值4096 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[64,4096],默认值64 (11)validationStep:[0,1000000],默认值16,步长1 (12)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64,默认值32 loraAlpha:单选,8 或 16 或 32 或 64,默认值32 loraDropout:[0.01,0.5],默认值0.1,步长0.001 |
Baichuan2-13B-Chat | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (3)batchSize:[1,2],默认值1 (4)Packing:字符串,true 或 false 或 auto,默认值auto (5)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (6)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (7)weightDecay:[0.001,1],默认0.01,步长0.001 (8)maxSeqLen: 单选,512 或 1024 或 2048 或 4096,默认值4096 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[64,4096],默认值64 (11)validationStep:[0,1000000],默认值16,步长1 (12)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64,默认值32 loraAlpha:单选,8 或 16 或 32 或 64,默认值32 loraDropout:[0.01,0.5],默认值0.1,步长0.001 |
BLOOMZ-7B | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (3)batchSize:[1,4],默认值1 (4)Packing:字符串,true 或 false 或 auto,默认值auto (5)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (6)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (7)weightDecay:[0.001,1],默认值0.01,步长0.001 (8)checkpointCount:[1,10],默认值1 (9)saveStep:[64,4096],默认值256 (10)validationStep:[0,1000000],默认值16,步长1 (11)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64,默认值32 loraAlpha:单选,8 或 16 或 32 或 64,默认值32 loraDropout:[0.01,0.5],默认值0.1,步长0.001 |
CodeLlama-7B | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (3)batchSize:[1,4],默认值1 (4)Packing:字符串,true 或 false 或 auto,默认值auto (5)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (6)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (7)weightDecay:[0.001,1],默认值0.01,步长0.001 (8)maxSeqLen: 单选,512 或 1024 或 2048 或 4096,默认值4096 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[64,4096],默认值64 (11)validationStep:[0,1000000],默认值16,步长1 loraAlpha:单选,8 或 16 或 32 或 64,默认值32 loraDropout:[0.01,0.5],默认值0.1 (12) loraTargetModules:多选,self_attn.q_proj、self_attn.k_proj、self_attn.v_proj、self_attn.o_proj、mlp.gate_proj、mlp.up_proj、mlp.down_proj,默认值self_attn.q_proj + self_attn.v_proj |
Custom-Model(自定义模型) | SFT | FullFineTuning | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000000001,0.0002],默认值0.000001,步长0.000001 (3)schedulerName: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值cosine (4)warmupRatio:[0.01,0.1],默认值0.03,步长0.001 (5)weightDecay:[0.001,1],默认值0.01,步长0.001 |
PostPretrain
model | trainMode | parameterScale | hyperParameterConfig |
---|---|---|---|
ERNIE-Speed-8K | PostPretrain | - | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.00003,步长0.000001 (3)maxSeqLen: 单选,4096 或 8192,默认值4096 (4)globalBatchSize:[1,10000],默认值32,步长1 (5)checkpointCount:[1,10],默认值1 (6)saveStep:[1,50000],默认值64,saveStep需要是validationStep的整数倍 (7)seed:[1,2147483647],默认值42 (8)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (9)numCycles:[0.1,0.5],默认值0.5,步长0.1 (10)lrEnd:[0.00000001,0.000001],默认值0.00000001 (11)power:[1,3],默认值1 (12)validationStep:[0,1000000],默认值16,步长1,saveStep需要是validationStep的整数倍 (13)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 |
ERNIE-Tiny-8K | PostPretrain | - | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.00003,步长0.000001 (3)maxSeqLen: 单选,4096 或 8192,默认值4096 (4)globalBatchSize:[1,10000],默认值32,步长8 (5)checkpointCount:[1,10],默认值1 (6)saveStep:[1,50000],默认值64,saveStep需要是validationStep的整数倍 (7)seed:[1,2147483647],默认值42 (8)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (9)numCycles:[0.1,0.5],默认值0.5,步长0.1 (10)lrEnd:[0.00000001,0.000001],默认值0.00000001 (11)power:[1,3],默认值1 (12)validationStep:[0,1000000],默认值16,步长1,saveStep需要是validationStep的整数倍 (13)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 |
Qianfan-Chinese-Llama-2-13B-v1 | PostPretrain | - | (1)epoch:1 (2)learningRate:[0.0000002,0.0002],默认值0.00002,步长0.000001 (3)batchSize:[48,960],默认值192,步长48 (4)weightDecay:[0.0001,0.05],默认值0.01,步长0.001 (5)checkpointCount:[1,10],默认值1 (6)saveStep:[64,8192],默认值64 (7)validationStep:[0, 1000000],默认值16,步长1 |
ERNIE-Speed-Pro-128K | PostPretrain | - | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.00003,步长0.000001 (3)maxSeqLen: 单选,8192 或 16384 或 32768 或 65536 或 131072,默认值32768 (4)globalBatchSize:[1,10000],默认值16,步长1 (5)checkpointCount:[1,10],默认值1 (6)saveStep:[1,50000],默认值64,saveStep需要是validationStep的整数倍 (7)seed:[1,2147483647],默认值42 (8)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (9)numCycles:[0.1,0.5],默认值0.5,步长0.1 (10)lrEnd:[0.00000001,0.000001],默认值0.00000001 (11)power:[1,3],默认值1 (12)validationStep:[0,1000000],默认值16,步长1,saveStep需要是validationStep的整数倍 (13)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 |
ERNIE-Tiny-128K-0929 | PostPretrain | - | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.00003,步长0.000001 (3)maxSeqLen: 单选,16384 或 32768 或 65536 或 131072,默认值32768 (4)globalBatchSize:[1,10000],默认值16,步长2 (5)checkpointCount:[1,10],默认值1 (6)saveStep:[1,50000],默认值64,saveStep需要是validationStep的整数倍 (7)seed:[1,2147483647],默认值42 (8)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (9)numCycles:[0.1,0.5],默认值0.5,步长0.1 (10)lrEnd:[0.00000001,0.000001],默认值0.00000001 (11)power:[1,3],默认值1 (12)validationStep:[0,1000000],默认值16,步长1,saveStep需要是validationStep的整数倍 (13)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 |
ERNIE-Lite-128K-0722 | PostPretrain | - | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.00003,步长0.000001 (3)maxSeqLen: 单选,8192 或 16384 或 32768 或 65536 或 131072, 默认值32768 (4)globalBatchSize:[1,10000], 默认值16, 步长2 (5)checkpointCount:[1,10],默认值1 (6)saveStep:[1,50000],默认值64,saveStep需要是validationStep的整数倍 (7)seed:[1,2147483647],默认值42 (8)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (9)numCycles:[0.1,0.5],默认值0.5,步长0.1 (10)lrEnd:[0.00000001,0.000001],默认值0.00000001 (11)power:[1,3],默认值1 (12)validationStep:[0,1000000],默认值16,步长1,saveStep需要是validationStep的整数倍 (13)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (14)tensorParallelDegree:[1,8],默认值4 (15)shardingParallelDegree:[1,64],默认值2 (16)sharding:stage1 或 stage2 或 stage3,默认值stage2 (17)recompute:0 或 1,默认值1 |
ERNIE-Character-Fiction-8K | PostPretrain | - | (1)epoch:[1,50],默认值1 (2)learningRate:[0.00000010, 0.01],默认值 0.00003,步长0.0000010 (3)maxSeqLen: 单选, 可选项4096、8192, 默认值 4096 (4)globalBatchSize:[1,10000],默认值32,步长2 (5)checkpointCount:[1,10],默认值1 (6)saveStep:[1,50000],默认值64 (7)seed:[1, 2147483647],默认值 42 (8)lrSchedulerType:单选,可选项linear、cosine、polynomial、constant、constant_with_warmup,默认值 linear (9)numCycles:[0.1,0.5],默认值0.5,步长0.1 (10)lrEnd:[0.00000001, 0.0000010],默认值 0.00000010,步长0.00000001 (11)power:[1,3],默认值1 (12)validationStep:[0,1000000],默认值16,步长1 (13)早停策略相关参数: earlyStopping:单选, 可选项False、True,默认值 False earlyStopMetric:单选,可选项validationLoss,默认值 validationLoss earlyStoppingThreshold:[0,5],默认值 0.01,步长0.01 earlyStoppingPatience:[1,50],默认值 3,步长1 |
DPO
model | trainMode | parameterScale | hyperParameterConfig |
---|---|---|---|
ERNIE-Lite-8K-0308 | DPO | FullFineTuning | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.000001,步长0.0000001 (3)maxSeqLen: 单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)dpoBeta:[0.01,1],默认值0.1,步长0.001 (8)seed:[1,2147483647],默认值42 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[1,50000],默认值64,saveStep需要是validationStep的整数倍 (11)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (12)numCycles:[0.1,0.5],默认值0.5,步长0.1 (13)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (14)power:[1,3],默认值1 (15)validationStep:[0, 1000000],默认值16,步长1 (16)globalBatchSize:[1,10000],默认值16,步长4 (17)lossType:sigmoid 或 ipo 或 kto_pair,默认值sigmoid (18)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,当参数earlyStopping为True时,此参数有效 |
ERNIE-Lite-128K-0722 | DPO | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.000001,步长0.0000001 (3)maxSeqLen: 单选,16384 或 32768 或 65536 或 131072,默认值32768 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)dpoBeta:[0.01,1],默认值0.1,步长0.001 (8)seed:[1,2147483647],默认值42 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[1,50000],默认值64,saveStep需要是validationStep的整数倍 (11)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (12)numCycles:[0.1,0.5],默认值0.5,步长0.1 (13)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (14)power:[1,3],默认值1 (15)validationStep:[0, 1000000],默认值16,步长1 (16)globalBatchSize:[1,10000],默认值16,步长1 (17)lossType:sigmoid 或 ipo 或 kto_pair,默认值sigmoid (18)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,当参数earlyStopping为True时,此参数有效 |
ERNIE-Lite-128K-0419 | DPO | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.000001,步长0.0000001 (3)maxSeqLen: 单选,16384 或 32768 或 65536 或 131072,默认值32768 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)dpoBeta:[0.01,1],默认值0.1,步长0.001 (8)seed:[1,2147483647],默认值42 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[1,50000],默认值64,saveStep需要是validationStep的整数倍 (11)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (12)numCycles:[0.1,0.5],默认值0.5,步长0.1 (13)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (14)power:[1,3],默认值1 (15)validationStep:[0, 1000000],默认值16,步长1 (16)globalBatchSize:[1,10000],默认值16,步长1 (17)lossType:sigmoid 或 ipo 或 kto_pair,默认值sigmoid (18)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,当参数earlyStopping为True时,此参数有效 |
ERNIE-Speed-8K | DPO | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.000001,步长0.0000001 (3)maxSeqLen: 单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)dpoBeta:[0.01,1],默认值0.1,步长0.001 (8)seed:[1,2147483647],默认值42 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[1,50000],默认值64,saveStep需要是validationStep的整数倍 (11)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (12)numCycles:[0.1,0.5],默认值0.5,步长0.1 (13)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (14)power:[1,3],默认值1 (15)validationStep:[0, 1000000],默认值16,步长1 (16)globalBatchSize:[1,10000],默认值16,步长1 (17)lossType:sigmoid 或 ipo 或 kto_pair,默认值sigmoid (18)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,当参数earlyStopping为True时,此参数有效 |
ERNIE-Tiny-8K | DPO | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.000001,步长0.0000001 (3)maxSeqLen: 单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)dpoBeta:[0.01,1],默认值0.1,步长0.001 (8)seed:[1,2147483647],默认值42 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[1,50000],默认值64,saveStep需要是validationStep的整数倍 (11)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (12)numCycles:[0.1,0.5],默认值0.5,步长0.1 (13)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (14)power:[1,3],默认值1 (15)validationStep:[0, 1000000],默认值16,步长1 (16)globalBatchSize:[1,10000],默认值32,步长8 (17)lossType:sigmoid 或 ipo 或 kto_pair,默认值sigmoid (18)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,当参数earlyStopping为True时,此参数有效 (19)仅LoRA支持: loraRank:单选,2 或 4 或 8 ,默认值8 |
ERNIE-Speed-Pro-128K | DPO | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.000001,步长0.0000001 (3)maxSeqLen: 单选,16384 或 32768 或 65536 或 131072,默认值32768 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)dpoBeta:[0.01,1],默认值0.1,步长0.001 (8)seed:[1,2147483647],默认值42 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[1,50000],默认值64,saveStep需要是validationStep的整数倍 (11)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (12)numCycles:[0.1,0.5],默认值0.5,步长0.1 (13)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (14)power:[1,3],默认值1 (15)validationStep:[0, 1000000],默认值16,步长1 (16)globalBatchSize:[1,10000],默认值16,步长1 (17)lossType:sigmoid 或 ipo 或 kto_pair,默认值sigmoid (18)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,当参数earlyStopping为True时,此参数有效 (19)仅LoRA支持: loraRank:单选,8 或 64 ,默认值64 |
ERNIE-Tiny-128K-0929 | DPO | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.000001,步长0.0000001 (3)maxSeqLen: 单选,16384 或 32768 或 65536 或 131072, 默认:32768 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)dpoBeta:[0.01,1],默认值0.1,步长0.001 (8)seed:[1,2147483647],默认值42 (9)checkpointCount:[1,10],默认值1 (10)saveStep:[1,50000],默认值64,saveStep需要是validationStep的整数倍 (11)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (12)numCycles:[0.1,0.5],默认值0.5,步长0.1 (13)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (14)power:[1,3],默认值1 (15)validationStep:[0, 1000000],默认值16,步长1 (16)globalBatchSize:[1,10000],默认值16,步长2 (17)lossType:sigmoid 或 ipo 或 kto_pair,默认值sigmoid (18)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank:单选,2 或 4 或 8 ,默认值8 |
KTO
model | trainMode | parameterScale | hyperParameterConfig |
---|---|---|---|
ERNIE-Speed-8K | KTO | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.000001,步长0.0000001 (3)maxSeqLen:单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)ktoBeta:[0.01,1],默认值0.1,步长0.001 (8)checkpointCount:[1,10],默认值1 (9)saveStep:[1,50000],默认值64 (10)seed:[1,2147483647],默认值42 (11)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (12)numCycles:[0.1,0.5],默认值0.5,步长0.1 (13)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (14)validationStep:[0,1000000],默认值16,步长1 (15)power:[1,3],默认值1 (16)globalBatchSize: FullFineTuning:[1,10000],默认值16,步长1 LoRA:[1,10000],默认值16,步长2 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank:单选,8 或 64 ,默认值64 |
ERNIE-Lite-128K-0419 | KTO | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.000001,步长0.0000001 (3)maxSeqLen:单选,16384 或 32768 或 65536 或 131072,默认值32768 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)ktoBeta:[0.01,1],默认值0.1,步长0.001 (8)checkpointCount:[1,10],默认值1 (9)saveStep:[1,50000],默认值64 (10)seed:[1,2147483647],默认值42 (11)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (12)numCycles:[0.1,0.5],默认值0.5,步长0.1 (13)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (14)validationStep:[0,1000000],默认值16,步长1 (15)power:[1,3],默认值1 (16)globalBatchSize:[1,10000],默认值16,步长1 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank:单选,2 或 4 或 8 ,默认值8 |
ERNIE-Lite-8K-0308 | KTO | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.000001,步长0.0000001 (3)maxSeqLen:单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)ktoBeta:[0.01,1],默认值0.1,步长0.001 (8)checkpointCount:[1,10],默认值1 (9)saveStep:[1,50000],默认值64 (10)seed:[1,2147483647],默认值42 (11)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (12)numCycles:[0.1,0.5],默认值0.5,步长0.1 (13)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (14)validationStep:[0,1000000],默认值16,步长1 (15)power:[1,3],默认值1 (16)globalBatchSize:[1,10000],默认值16,步长4 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank: 单选,2 或 4 或 8 ,默认值8 |
ERNIE-Character-Fiction-8K | KTO | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.000001,步长0.0000001 (3)maxSeqLen:单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)ktoBeta:[0.01,1],默认值0.1,步长0.001 (8)checkpointCount:[1,10],默认值1 (9)saveStep:[1,50000],默认值64 (10)seed:[1,2147483647],默认值42 (11)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (12)numCycles:[0.1,0.5],默认值0.5,步长0.1 (13)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (14)validationStep:[0,1000000],默认值16,步长1 (15)power:[1,3],默认值1 (16)globalBatchSize: FullFineTuning:[1,10000],默认值16,步长1 LoRA:[1,10000],默认值16,步长2 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank:单选,2 或 4 或 8 ,默认值8 |
ERNIE-Character-8K-0321 | KTO | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.000001,步长0.0000001 (3)maxSeqLen:单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)ktoBeta:[0.01,1],默认值0.1,步长0.001 (8)checkpointCount:[1,10],默认值1 (9)saveStep:[1,50000],默认值64 (10)seed:[1,2147483647],默认值42 (11)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (12)numCycles:[0.1,0.5],默认值0.5,步长0.1 (13)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (14)validationStep:[0,1000000],默认值16,步长1 (15)power:[1,3],默认值1 (16)globalBatchSize: FullFineTuning:[1,10000],默认值16,步长1 LoRA:[1,10000],默认值16,步长2 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank:单选,2 或 4 或 8 ,默认值8 |
ERNIE-Tiny-8K | KTO | FullFineTuning、LoRA | ((1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.000001,步长0.0000001 (3)maxSeqLen:单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)ktoBeta:[0.01,1],默认值0.1,步长0.001 (8)checkpointCount:[1,10],默认值1 (9)saveStep:[1,50000],默认值64 (10)seed:[1,2147483647],默认值42 (11)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (12)numCycles:[0.1,0.5],默认值0.5,步长0.1 (13)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (14)validationStep:[0,1000000],默认值16,步长1 (15)power:[1,3],默认值1 (16)globalBatchSize:[1,10000],默认值32,步长8 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank:单选,2 或 4 或 8 ,默认值8 |
ERNIE-Tiny-128K-0929 | KTO | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.000001,步长0.0000001 (3)maxSeqLen:单选,16384 或 32768 或 65536 或 131072, 默认值32768 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)ktoBeta:[0.01,1],默认值0.1,步长0.001 (8)checkpointCount:[1,10],默认值1 (9)saveStep:[1,50000],默认值64 (10)seed:[1,2147483647],默认值42 (11)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (12)numCycles:[0.1,0.5],默认值0.5,步长0.1 (13)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (14)validationStep:[0,1000000],默认值16,步长1 (15)power:[1,3],默认值1 (16)globalBatchSize:[1,10000],默认值16,步长2 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank: 单选,2 或 4 或 8 ,默认值8 |
ERNIE-Speed-Pro-128K | KTO | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.000001,步长0.0000001 (3)maxSeqLen:单选,16384 或 32768 或 65536 或 131072, 默认值32768 (4)loggingSteps:1 (5)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (6)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (7)ktoBeta:[0.01,1],默认值0.1,步长0.001 (8)checkpointCount:[1,10],默认值1 (9)saveStep:[1,50000],默认值64 (10)seed:[1,2147483647],默认值42 (11)lrSchedulerType: 单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (12)numCycles:[0.1,0.5],默认值0.5,步长0.1 (13)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (14)validationStep:[0,1000000],默认值16,步长1 (15)power:[1,3],默认值1 (16)globalBatchSize:[1,10000],默认值16,步长1 (17)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (18)仅LoRA支持: loraRank:单选,8 或 64 ,默认值64 |
RLHF
model | trainMode | parameterScale | hyperParameterConfig |
---|---|---|---|
ERNIE-Lite-8K-0308 | RM | FullFineTuning | (1)epoch:[1, 50],默认值 1 (2)learningRate:[0.00000010, 0.01],默认值0.0000010,步长0.00000010 (3)maxSeqLen:单选,可选项4096、8192,默认值4096 (4)globalBatchSize:[1, 10000],默认值16,步长4 (5)useCls:单选,可选项True、False, 默认值True (6)warmupRatio:[0.01, 0.5],默认值0.1,步长0.01 (7)weightDecay:[0.0001, 0.1],默认值0.01,步长0.0001 (8)pseudoSamplingProb:[0, 0.9],默认值0,步长0.1 (9)seed:[1, 2147483647],默认值42 (10)lrSchedulerType:单选,可选项linear、cosine、polynomial、constant、constant_with_warmup,默认值linear (11)numCycles:[0.1, 0.5],默认值0.5,步长0.1 (12)lrEnd:[0.00000001, 0.0000010],默认值0.00000010,步长0.00000001 (13)validationStep:[0, 1000000],默认值16,步长1 (14)power:[1, 3],默认值1 |
ERNIE-Lite-8K-0308 | PPO | FullFineTuning | (1)epoch:[1, 50],默认值 1 (2)critic_learning_rate:[0.00000010, 0.00001],默认值0.000002,步长0.00000010 (3)learningRate:[0.00000010, 0.00001],默认值0.0000010,步长0.00000010 (4)maxSeqLen:单选,可选项4096、8192,默认值4096 (5)globalBatchSize:[1, 10000],默认值16,步长4 (6)clip_range_score:[5, 50],默认值10 (7)clip_range_value:[5, 50],默认值5 (8)clip_range_ratio:[0.01, 0.3],默认值0.2 (9)loggingSteps:[1, 1],默认值1 (10)warmupRatio:[0.01, 0.5],默认值0.1,步长0.01 (11)weightDecay:[0.0001, 0.1],默认值0.01,步长0.0001 (12)top_p:[0, 1],默认值0.9 (13)validationStep:[0, 1000000],默认值16,步长1 (14)repetition_penalty:[1, 2],默认值1 (15)temperature:[0, 1],默认值1 (16)kl_coeff:[0.001, 0.1],默认值0.02 (17)checkpointCount:[1, 10],默认值1 (18)saveStep:单选,可选项64、128、256、512、1024、2048、4096,默认值256 (19)seed:[1, 2147483647],默认值42 (20)lrSchedulerType:单选,可选项linear、cosine、polynomial、constant、constant_with_warmup,默认值linear (21)numCycles:[0.1, 0.5],默认值0.5,步长0.1 (22)lrEnd:[0.00000001, 0.0000010],默认值0.00000010,步长0.00000001 (23)power:[1, 3],默认值1 |
SimPO
model | trainMode | parameterScale | hyperParameterConfig |
---|---|---|---|
ERNIE-Character-Fiction-8K-1028 | SimPO | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.001],默认值0.000001,步长0.0000001 (3)maxSeqLen:单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)globalBatchSize: FullFineTuning:[1,10000],默认值16,步长1 LoRA:[1,10000],默认值16,步长4 (5)loggingSteps:1 (6)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (7)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (8)simpoBeta:[2,2.5],默认值2,步长0.001 (9)simpoGamma:[0.01,1.5],默认值0.5,步长0.001 (10)checkpointCount:[1,10],默认值1 (11)saveStep:[1,50000],默认值64 (12)seed:[1,2147483647],默认值42,步长1 (13)lrSchedulerType:单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (14)numCycles:[0.1,0.5],默认值0.5,步长0.1 (15)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (16)power:[1,3],默认值1 (17)validationStep:[0, 1000000],默认值16,步长1 (18)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (19)仅LoRA支持: loraRank:单选,2 或 4 或 8 ,默认值8 |
ERNIE-Lite-128K-0722 | SimPO | FullFineTuning | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001, 0.01],默认值0.00003,步长0.000001 (3)maxSeqLen:单选,16384 或 32768 或 65536 或 131072, 默认值32768 (4)globalBatchSize:[1,10000],默认值16,步长1 (5)loggingSteps:1 (6)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (7)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (8)simpoBeta:[2,2.5],默认值2,步长0.001 (9)simpoGamma:[0.01,1.5],默认值0.5,步长0.001 (10)checkpointCount:[1,10],默认值1 (11)saveStep:[1,50000],默认值64 (12)seed:[1,2147483647],默认值42,步长1 (13)lrSchedulerType:单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (14)numCycles:[0.1,0.5],默认值0.5,步长0.1 (15)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (16)power:[1,3],默认值1 (17)validationStep:[0, 1000000],默认值16,步长1 (18)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (19)tensorParallelDegree:[1,8],默认值4 (20)shardingParallelDegree:[1,64],默认值2 (21)sharding:stage1 或 stage2 或 stage3,默认值stage2 (22)recompute:0 或 1,默认值1 |
ERNIE-Lite-8K-0308 | SimPO | FullFineTuning | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.001],默认值0.000001,步长0.0000001 (3)maxSeqLen:单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)globalBatchSize:[1,10000], 默认值16,步长4 (5)loggingSteps:1 (6)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (7)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (8)simpoBeta:[2,2.5],默认值2,步长0.001 (9)simpoGamma:[0.01,1.5],默认值0.5,步长0.001 (10)checkpointCount:[1,10],默认值1 (11)saveStep:[1,50000],默认值64 (12)seed:[1,2147483647],默认值42,步长1 (13)lrSchedulerType:单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (14)numCycles:[0.1,0.5],默认值0.5,步长0.1 (15)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (16)power:[1,3],默认值1 (17)validationStep:[0, 1000000],默认值16,步长1 (18)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (19)(19)tensorParallelDegree:[1,8],默认值2 (20)shardingParallelDegree:[1,64],默认值4 (21)sharding:stage1 或 stage2 或 stage3,默认值stage2 |
ERNIE-Speed-8K | SimPO | FullFineTuning | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.001],默认值0.000001,步长0.0000001 (3)maxSeqLen:单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)globalBatchSize:[1,10000],默认值16,步长1 (5)loggingSteps:1 (6)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (7)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (8)simpoBeta:[2,2.5],默认值2,步长0.001 (9)simpoGamma:[0.01,1.5],默认值0.5,步长0.001 (10)checkpointCount:[1,10],默认值1 (11)saveStep:[1,50000],默认值64 (12)seed:[1,2147483647],默认值42,步长1 (13)lrSchedulerType:单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (14)numCycles:[0.1,0.5],默认值0.5,步长0.1 (15)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (16)power:[1,3],默认值1 (17)validationStep:[0, 1000000],默认值16,步长1 (18)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 |
ERNIE-Speed-Pro-128K | SimPO | FullFineTuning | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.00003,步长0.000001 (3)maxSeqLen:单选,16384 或 32768 或 65536 或 131072, 默认值32768 (4)globalBatchSize:[1,10000],默认值16,步长1 (5)loggingSteps:1 (6)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (7)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (8)simpoBeta:[2,2.5],默认值2,步长0.001 (9)simpoGamma:[0.01,1.5],默认值0.5,步长0.001 (10)checkpointCount:[1,10],默认值1 (11)saveStep:[1,50000],默认值64 (12)seed:[1,2147483647],默认值42,步长1 (13)lrSchedulerType:单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (14)numCycles:[0.1,0.5],默认值0.5,步长0.1 (15)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (16)power:[1,3],默认值1 (17)validationStep:[0, 1000000],默认值16,步长1 (18)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (19)sharding:stage1 或 stage2 或 stage3,默认值stage2 (20)recompute:0 或 1,默认值 1 |
ERNIE-Tiny-8K | SimPO | FullFineTuning | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.001],默认值0.000001,步长0.0000001 (3)maxSeqLen:单选,512 或 1024 或 2048 或 4096 或 8192,默认值4096 (4)globalBatchSize:[1,10000],默认16,步长8 (5)loggingSteps:1 (6)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (7)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (8)simpoBeta:[2,2.5],默认值2,步长0.001 (9)simpoGamma:[0.01,1.5],默认值0.5,步长0.001 (10)checkpointCount:[1,10],默认值1 (11)saveStep:[1,50000],默认值64 (12)seed:[1,2147483647],默认值42,步长1 (13)lrSchedulerType:单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (14)numCycles:[0.1,0.5],默认值0.5,步长0.1 (15)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (16)power:[1,3],默认值1 (17)validationStep:[0, 1000000],默认值16,步长1 (18)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (19)tensorParallelDegree:[1,8],默认1 (20)shardingParallelDegree:[1,64],默认8 |
ERNIE-Tiny-128K-0929 | SimPO | FullFineTuning | (1)epoch:[1,50],默认值1 (2)learningRate:[0.0000001,0.01],默认值0.00003,步长0.000001 (3)maxSeqLen:单选,16384 或 32768 或 65536 或 131072, 默认值32768 (4)globalBatchSize: FullFineTuning:[1,10000],默认值16,步长1 LoRA:[1,10000],默认值16,步长4 (5)loggingSteps:1 (6)warmupRatio:[0.01,0.5],默认值0.1,步长0.01 (7)weightDecay:[0.0001,0.1],默认值0.01,步长0.0001 (8)simpoBeta:[2,2.5],默认值2,步长0.001 (9)simpoGamma:[0.01,1.5],默认值0.5,步长0.001 (10)checkpointCount:[1,10],默认值1 (11)saveStep:[1,50000],默认值64 (12)seed:[1,2147483647],默认值42,步长1 (13)lrSchedulerType:单选,linear 或 cosine 或 polynomial 或 constant 或 constant_with_warmup,默认值linear (14)numCycles:[0.1,0.5],默认值0.5,步长0.1 (15)lrEnd:[0.00000001,0.000001],默认值0.0000001,步长0.00000001 (16)power:[1,3],默认值1 (17)validationStep:[0,1000000],默认值16,步长1 (18)早停策略相关参数: earlyStopping:True 或 False,默认False earlyStopMetric:ValidationLoss,当参数earlyStopping为True时,此参数有效 earlyStoppingThreshold:[0,5] ,默认值 0.01,步长0.01,当参数earlyStopping为True时,此参数有效 earlyStoppingPatience:[1,50],默认值 3,步长1,当参数earlyStopping为True时,此参数有效 (19)tensorParallelDegree:[1,8],默认值1 (20)shardingParallelDegree:[1,64],默认值8 (21)sharding:stage1 或 stage2 或 stage3,默认值stage2 (22)recompute:0 或 1,默认值1 |
图像生成类
model | trainMode | parameterScale | hyperParameterConfig |
---|---|---|---|
WENXIN-YIGE | SFT | FullFineTuning | (1)epoch:[1,100],默认值20 (2)learningRate:[0.00000001,0.01],默认值0.00001 (3)batchSize:[1,8],默认值8 |
Stable-Diffusion-XL-Base-1.0 | SFT | LoRA | epoch:[1,100],默认值20 (2)learningRate:[0.00001,0.0001],默认值0.00005 (3)batchSize:[2,8],默认值8 |
图像理解类型
model | trainMode | parameterScale | hyperParameterConfig |
---|---|---|---|
LLAVA-V1.6-13B | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate: FullFineTuning:[0.0000000001,0.001],默认值0.0002,递增步长0.00001 LoRA:[0.0000000001,0.001],默认值0.0002,递增步长0.00001 (3)validationStep:[0,1000000],默认值16,递增步长1 (4)batchSize:[1,16],默认值16 (5)schedulerName:单选:linear、cosine、polynomial、constant、constant_with_warmup,默认值cosine (6)warmupRatio:[0.01,0.1],默认值0.03,递增步长0.001 (7)weightDecay:[0.001,1],默认值0.001,递增步长0.001 (8)maxSeqLen:单选:512、1024、2048、4096,默认值4096 (9)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64 或 128 或 256,默认值128 loraAlpha:单选,8 或 16 或 32 或 64 或 128 或 256,默认值256 |
InternVL2-2B | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate: FullFineTuning:[0.0000000001,0.001],默认值0.00001,递增步长0.000001 LoRA:[0.0000000001,0.001],默认值0.0001,递增步长0.00004 (3)validationStep:[0, 1000000], 默认值16,递增步长1 (4)batchSize:[1,16],默认值1 (5)checkpointCount:[1,10],默认值1 (6)saveStep:[1,50000],默认值64 (7)schedulerName:单选:linear、cosine、polynomial、constant、constant_with_warmup,默认值cosine (8)warmupRatio:[0.01, 0.1],默认值0.05,递增步长0.001 (9)weightDecay:[0.001, 1],默认值0.1,递增步长0.001 (10)maxSeqLen:单选:512、1024、2048、4096,默认值2048 (11)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64 或 128 或 256,默认值8 |
InternVL2-8B | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate: FullFineTuning:[0.0000000001,0.001],默认值0.00001,递增步长0.000001 LoRA: [0.0000000001,0.001],默认值0.0001,步长0.00004 (3)validationStep:[0, 1000000], 默认值16,递增步长1 (4)batchSize: FullFineTuning:[1,4],默认值1,步长1 LoRA:[1,5],默认值1 (5)checkpointCount:[1,10],默认值1 (6)saveStep:[1,50000],默认值64 (7)schedulerName:单选:linear、cosine、polynomial、constant、constant_with_warmup,默认值cosine (8)warmupRatio:[0.01, 0.1],默认值0.05,递增步长0.001 (9)weightDecay:[0.001, 1],默认值0.1,递增步长0.001 (10)maxSeqLen:单选:512、1024、2048、4096,默认值2048 (11)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64 或 128 或 256,默认值8 |
InternLM-XComposer2.5 | SFT | FullFineTuning、LoRA | (1)epoch:[1,50],默认值1 (2)learningRate: FullFineTuning:[0.0000000001,0.001],默认值0.00001,递增步长0.000001 LoRA:[0.0000000001,0.001],默认值0.0001,递增步长0.00004 (3)batchSize:[1,16],默认值1 (4)validationStep:[0, 1000000],默认值16,递增步长1 (5)checkpointCount:[1,10],默认值1 (6)saveStep:[1,50000],默认值64 (7)schedulerName:单选:linear、cosine、polynomial、constant、constant_with_warmup,默认值cosine (8)warmupRatio:[0.01,0.1],默认值0.05,递增步长0.001 (9)weightDecay:[0.001, 1],默认值0.1,递增步长0.001 (10)maxSeqLen:单选:512、1024、2048、4096,默认值2048 (11)仅LoRA支持: loraRank:单选,8 或 16 或 32 或 64 或 128 或 256,默认值8 |