SciWorld-MPO

SciWorld-MPO

This model is a fine-tuned version of Llama-3.1-8B-Instruct on the sciworld-metaplan-preference-pairs dataset. It achieves the following results on the evaluation set:

• Loss: 1.5017
• Rewards/chosen: -3.8774
• Rewards/rejected: -5.1594
• Rewards/accuracies: 0.6419
• Rewards/margins: 1.2820
• Logps/chosen: -92.4593
• Logps/rejected: -109.6343
• Logits/chosen: 0.5212
• Logits/rejected: 0.5151

See the original paper for more details: MPO: Boosting LLM Agents with Meta Plan Optimization.

Code: https://github.com/WeiminXiong/MPO

Model description

This model uses Meta Plan Optimization (MPO) to improve the planning capabilities of LLM agents. It leverages high-level general guidance through meta plans and enables continuous optimization based on feedback from the agent's task execution. It achieves state-of-the-art performance on ALFWorld and SciWorld, with an average accuracy of 83.1.

Intended uses & limitations

More information needed

Training and evaluation data

The model was trained on the sciworld-metaplan-preference-pairs dataset, part of the Meta_Plan_Optimization dataset.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 2
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• total_eval_batch_size: 4
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 3.0

Training results

Framework versions

• Transformers 4.46.1
• Pytorch 2.5.1+cu124
• Datasets 3.1.0
• Tokenizers 0.20.3