llama_8b_lima_11_kto

This model is a fine-tuned version of OpenLeecher/llama_8b_lima_11 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.1369
• Rewards/chosen: 9.4796
• Logps/chosen: -655.7220
• Rewards/rejected: -11.1496
• Logps/rejected: -294.5653
• Rewards/margins: 20.6292
• Kl: 4.1962

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: 2e-06
• train_batch_size: 1
• eval_batch_size: 1
• seed: 66
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 20
• total_train_batch_size: 40
• total_eval_batch_size: 2
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: polynomial
• lr_scheduler_warmup_steps: 30
• num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Logps/chosen	Rewards/rejected	Logps/rejected	Rewards/margins	Kl
0.3224	0.2973	90	0.2095	5.6903	-664.1426	-9.8058	-291.5791	15.4961	3.0351
0.2716	0.5945	180	0.1669	8.8699	-657.0769	-11.5150	-295.3773	20.3849	4.1776
0.2089	0.8918	270	0.1369	9.4796	-655.7220	-11.1496	-294.5653	20.6292	4.1962

Framework versions

• Transformers 4.45.0
• Pytorch 2.4.1+cu124
• Datasets 2.21.0
• Tokenizers 0.20.1