ap-normistral-7b-align-scan

This model is a fine-tuned version of data/ap-normistral-7b-sft-qlora on the hugodk-sch/aftonposten_title_prefs dataset. It achieves the following results on the evaluation set:

• Loss: 0.9524
• Rewards/chosen: -0.0816
• Rewards/rejected: -0.1313
• Rewards/accuracies: 0.5303
• Rewards/margins: 0.0497
• Logps/rejected: -36.6233
• Logps/chosen: -32.8513
• Logits/rejected: 98.1886
• Logits/chosen: 98.2171

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: 5e-06
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.927	0.26	100	0.9734	0.0205	-0.0061	0.5598	0.0265	-35.9968	-32.3408	98.7115	98.7190
0.7448	0.52	200	0.9482	-0.0840	-0.1367	0.5307	0.0527	-36.6501	-32.8631	98.2057	98.2271
0.7402	0.78	300	0.9400	-0.0802	-0.1425	0.5623	0.0624	-36.6792	-32.8440	98.1900	98.2178

Framework versions

• PEFT 0.10.0
• Transformers 4.39.0.dev0
• Pytorch 2.1.2+cu121
• Datasets 2.14.6
• Tokenizers 0.15.1