anydef-orpo-v2

This model is a fine-tuned version of mistralai/Mistral-7B-v0.1 on the arynkiewicz/anydef-kilt-tasks-v2 dataset.

Find out about Model description, Intended uses & limitations and Training and evaluation data on our github.

This is an updated version of the anydef model. The primary goal was to use an improved dataset during fine-tuning, enabling the model to better understand nuances. Overall, anydef-v2 offers better performance in benchmarks, and manual inspection of the results suggests that the model has indeed improved.

Precision (%):

Dataset	anydef	anydef-v2
RSS-500	66.23	66.89
ISTEX-1000	86.72	85.82
Reuters-128	63.8	64.88
TweekiGold	75.23	75.93

Retrieval rate (%):

Dataset	anydef	anydef-v2
RSS-500	82.78	84.11
ISTEX-1000	97.91	97.76
Reuters-128	80.47	83.33
TweekiGold	89.93	91.67

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-06
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 64
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: inverse_sqrt
• lr_scheduler_warmup_steps: 100
• num_epochs: 3

Training results

Framework versions

• Transformers 4.43.3
• Pytorch 2.1.2+cu121
• Datasets 2.18.0
• Tokenizers 0.19.1