Llama-3.1-8B-Instruct-KTO-300

This model is a fine-tuned version of meta-llama/Meta-Llama-3.1-8B-Instruct on the bct_non_cot_kto_300 dataset. It achieves the following results on the evaluation set:

• Loss: 0.2807
• Rewards/chosen: 0.7524
• Logps/chosen: -12.9007
• Logits/chosen: -3652706.0
• Rewards/rejected: -1.5384
• Logps/rejected: -37.9030
• Logits/rejected: -7997593.7778
• Rewards/margins: 2.2908
• Kl: 0.6633

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-06
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 16
• 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: 0.1
• num_epochs: 10.0

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Logps/chosen	Logits/chosen	Rewards/rejected	Logps/rejected	Logits/rejected	Rewards/margins
0.4973	1.4889	50	0.4995	0.0407	-20.0184	-5131056.3333	0.0319	-22.1998	-8269521.7778	0.0088	4.2481
0.4683	2.9778	100	0.4692	0.3686	-16.7392	-4839914.0	0.0514	-22.0054	-8194471.1111	0.3172	12.3862
0.3709	4.4741	150	0.3692	0.5681	-14.7444	-4248186.6667	-0.5877	-28.3963	-8021952.0	1.1558	2.0650
0.3219	5.9630	200	0.3044	0.6412	-14.0130	-3768132.0	-1.2363	-34.8816	-7958945.7778	1.8775	1.4612
0.2882	7.4593	250	0.2864	0.7339	-13.0862	-3685730.3333	-1.4520	-37.0385	-7991116.4444	2.1858	0.8725
0.2574	8.9481	300	0.2807	0.7524	-12.9007	-3652706.0	-1.5384	-37.9030	-7997593.7778	2.2908	0.6633

Framework versions

• PEFT 0.12.0
• Transformers 4.46.1
• Pytorch 2.5.1+cu124
• Datasets 3.1.0
• Tokenizers 0.20.3