orpo-llama-3b-iter-2-bird-planner-no-filter-seed106

This model is a fine-tuned version of ./output/reproduce/orpo-llama-3b-bird-planner-no-filter-seed100/ on the ../data/llm_alignment/bird-p2-planner-seed106/dpo-llama-3-end2end-bird_train_planner dataset. It achieves the following results on the evaluation set:

• Loss: 0.2745
• Rewards/chosen: -0.1037
• Rewards/rejected: -7.1776
• Rewards/accuracies: 0.9400
• Rewards/margins: 7.0738
• Logps/rejected: -7.1776
• Logps/chosen: -0.1037
• Logits/rejected: -0.2937
• Logits/chosen: -0.2329
• Nll Loss: 0.1037
• Log Odds Ratio: -0.1496
• Log Odds Chosen: 9.0892

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: 8e-06
• train_batch_size: 2
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• total_eval_batch_size: 2
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 1

Training results

Framework versions

• Transformers 4.45.0
• Pytorch 2.2.2
• Datasets 2.14.6
• Tokenizers 0.20.3