saiteja33/qwen2.5-1.5B-sample-hdpo-en-final

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

qwen25-15b-sample-hdpo

This model is a fine-tuned version of saves/qwen2.5-1.5B-sample-dist-sft-en on the polyalignhdpotrain dataset. It achieves the following results on the evaluation set:

• —Loss: 0.3407
• —Rewards/chosen: -1.5331
• —Rewards/rejected: -5.1943
• —Rewards/accuracies: 0.9172
• —Rewards/margins: 3.6612
• —Logps/chosen: -273.2844
• —Logps/rejected: -162.0241
• —Logits/chosen: 0.7203
• —Logits/rejected: 0.7014
• —Hdpo/weight Mean: 1.0001
• —Hdpo/chosen Dist Score: 0.0313
• —Hdpo/rejected Dist Score: 0.0313
• —Hdpo/dist Gap: 0.0
• —Hdpo/dist Reg: 0.0313
• —Hdpo/online Reg: 0.0
• —Hdpo/sft Anchor: 1.9628

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
• —trainbatchsize: 1
• —evalbatchsize: 1
• —seed: 42
• —distributed_type: multi-GPU
• —num_devices: 4
• —gradientaccumulationsteps: 8
• —totaltrainbatch_size: 32
• —totalevalbatch_size: 4
• —optimizer: Use OptimizerNames.ADAMWTORCHFUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• —lrschedulertype: cosine
• —lrschedulerwarmup_ratio: 0.1
• —num_epochs: 1.0
• —mixedprecisiontraining: Native AMP

Training results

Framework versions

• —Transformers 4.57.6
• —Pytorch 2.10.0+cu128
• —Datasets 4.0.0
• —Tokenizers 0.22.2