qwen2.5-3b-expo-L2EXPO-25-3b-1

This model is a fine-tuned version of hZzy/qwen2.5-3b-sft3-25-2 on the hZzy/train_pairwise_all_new4 dataset. It achieves the following results on the evaluation set:

• Loss: 0.4745
• Objective: 0.4683
• Reward Accuracy: 0.6228
• Logp Accuracy: 0.6055
• Log Diff Policy: 6.6940
• Chosen Logps: -92.7974
• Rejected Logps: -99.4915
• Chosen Rewards: -0.2833
• Rejected Rewards: -0.3471
• Logits: -1.2886

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: 1e-07
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 6
• gradient_accumulation_steps: 12
• total_train_batch_size: 72
• total_eval_batch_size: 6
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant_with_warmup
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Objective	Reward Accuracy	Logp Accuracy	Log Diff Policy	Chosen Logps	Rejected Logps	Chosen Rewards	Rejected Rewards	Logits
0.5438	0.1577	200	0.5090	0.5094	0.5266	0.5210	0.9004	-73.2070	-74.1074	-0.0874	-0.0932	-0.8639
0.4885	0.3154	400	0.4999	0.5020	0.5579	0.5546	1.9614	-78.4734	-80.4348	-0.1400	-0.1565	-0.9730
0.4875	0.4731	600	0.4911	0.4929	0.5741	0.5641	3.4973	-92.2967	-95.7940	-0.2783	-0.3101	-1.0447
0.4575	0.6308	800	0.4835	0.4798	0.5976	0.5965	5.5421	-94.7561	-100.2982	-0.3029	-0.3551	-1.1891
0.4794	0.7885	1000	0.4771	0.4715	0.6139	0.5976	6.4893	-98.6883	-105.1776	-0.3422	-0.4039	-1.2019
0.4268	0.9462	1200	0.4757	0.4708	0.6161	0.5965	6.4884	-94.2826	-100.7710	-0.2981	-0.3599	-1.2718

Framework versions

• Transformers 4.42.0
• Pytorch 2.6.0+cu124
• Datasets 3.2.0
• Tokenizers 0.19.1