deberta-v3-large-survey-cross_passage_consistency-rater-all

This model is a fine-tuned version of microsoft/deberta-v3-large on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.8117
• Krippendorff: 0.7331
• Spearman: 0.7616
• Absolute Agreement: 0.7177
• Agreement Within One: 0.9643

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 6e-06
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Krippendorff	Spearman	Absolute Agreement	Agreement Within One
No log	1.0	55	1.8836	-0.0604	nan	0.3472	0.7361
No log	2.0	110	1.8656	-0.0604	nan	0.3472	0.7361
No log	3.0	165	1.8291	-0.0604	nan	0.3472	0.7361
No log	4.0	220	1.7616	-0.2804	-0.2073	0.2917	0.6667
No log	5.0	275	1.7034	-0.2704	nan	0.3194	0.6667
No log	6.0	330	1.5944	-0.2704	nan	0.3194	0.6667
No log	7.0	385	1.7482	-0.2704	nan	0.3194	0.6667
No log	8.0	440	1.6018	-0.2704	nan	0.3194	0.6667
No log	9.0	495	1.6257	-0.2754	-0.1058	0.3194	0.6667
1.6154	10.0	550	1.5990	-0.2704	nan	0.3194	0.6667
1.6154	11.0	605	1.5314	-0.0526	0.4075	0.4306	0.6944
1.6154	12.0	660	1.4840	0.0189	0.5219	0.4722	0.7361
1.6154	13.0	715	1.3873	0.0165	0.5569	0.4861	0.7361
1.6154	14.0	770	1.3236	0.4564	0.6132	0.4722	0.875
1.6154	15.0	825	1.3843	0.3405	0.5305	0.4167	0.875
1.6154	16.0	880	1.2952	0.4904	0.6072	0.4861	0.8611
1.6154	17.0	935	1.3089	0.4748	0.6265	0.4583	0.875
1.6154	18.0	990	1.2532	0.4117	0.6346	0.4861	0.8056
0.9883	19.0	1045	1.2916	0.2859	0.6616	0.5	0.7778
0.9883	20.0	1100	1.2912	0.4686	0.6283	0.4861	0.8472

Framework versions

• Transformers 4.26.0
• Pytorch 1.13.1
• Datasets 2.10.1
• Tokenizers 0.12.1