r-zephyr-7b-beta-qlora

The 'r' means replicate. This model is a model replicated by using https://github.com/huggingface/alignment-handbook.

This model is a fine-tuned version on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.5232
• Rewards/chosen: -0.9374
• Rewards/rejected: -1.7181
• Rewards/accuracies: 0.7734
• Rewards/margins: 0.7807
• Logps/rejected: -420.1122
• Logps/chosen: -341.2448
• Logits/rejected: 0.6190
• Logits/chosen: 0.6345

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
• num_devices: 8
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• total_eval_batch_size: 64
• 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
0.5917	0.21	100	0.5950	-0.3904	-0.7775	0.7109	0.3872	-326.0618	-286.5451	-1.9790	-1.9769
0.5281	0.42	200	0.5492	-0.8657	-1.6137	0.7617	0.7479	-409.6739	-334.0814	-0.2289	-0.2367
0.5321	0.63	300	0.5321	-0.7444	-1.4427	0.7734	0.6983	-392.5731	-321.9463	0.3829	0.3741
0.5149	0.84	400	0.5233	-0.9570	-1.7432	0.7617	0.7862	-422.6298	-343.2071	0.6479	0.6688

Framework versions

• PEFT 0.7.1
• Transformers 4.36.2
• Pytorch 2.1.2+cu121
• Datasets 2.14.6
• Tokenizers 0.15.2

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	62.70
AI2 Reasoning Challenge (25-Shot)	63.05
HellaSwag (10-Shot)	85.38
MMLU (5-Shot)	63.10
TruthfulQA (0-shot)	46.32
Winogrande (5-shot)	79.32
GSM8k (5-shot)	39.04