Description

This model was trained as part of the Reinforcement Learning - 24 project at Peking University, focusing on [simpo].

Authors

• Ejafa Bassam
• Yaroslav Ponomarenko

qwen2-1.5b-instruct-simpo-lr-5e-07-gamma-1.5

This model is a fine-tuned version of Qwen/Qwen2-1.5B-Instruct on the princeton-nlp/llama3-ultrafeedback dataset. It achieves the following results on the evaluation set:

• Loss: 1.6346
• Rewards/chosen: -2.6152
• Rewards/rejected: -2.7999
• Rewards/accuracies: 0.5685
• Rewards/margins: 0.1847
• Logps/rejected: -1.1200
• Logps/chosen: -1.0461
• Logits/rejected: -1.5578
• Logits/chosen: -1.5356

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-07
• train_batch_size: 2
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 128
• total_eval_batch_size: 32
• 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
1.6402	0.8549	400	1.6353	-2.6155	-2.7990	0.5726	0.1835	-1.1196	-1.0462	-1.5085	-1.4841

Framework versions

• Transformers 4.41.2
• Pytorch 2.3.1+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1