This model is a fine-tuned version of JunxiongWang/llama3_mamba_0_5_sft on the HuggingFaceH4/ultrafeedback_binarized, the HuggingFaceH4/orca_dpo_pairs and the JunxiongWang/llama3-ultrafeedback-armorm datasets. It achieves the following results on the evaluation set:

Loss: 0.4002
Rewards/chosen: -2.2460
Rewards/rejected: -5.4992
Rewards/accuracies: 0.8536
Rewards/margins: 3.2532
Logps/rejected: -796.0059
Logps/chosen: -463.1195
Logits/rejected: -1.1906
Logits/chosen: -1.2034

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-07
train_batch_size: 4
eval_batch_size: 8
seed: 42
distributed_type: multi-GPU
num_devices: 8
total_train_batch_size: 32
total_eval_batch_size: 64
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_ratio: 0.1
num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.4244	0.4798	2000	0.4296	-2.2555	-4.8626	0.8250	2.6071	-732.3422	-464.0680	-1.2867	-1.2865
0.4311	0.9597	4000	0.4002	-2.2460	-5.4992	0.8536	3.2532	-796.0059	-463.1195	-1.1906	-1.2034

Framework versions

Transformers 4.43.1
Pytorch 2.1.1+cu118
Datasets 2.20.0
Tokenizers 0.19.1

MambaInLlama

@article{junxiongdaniele2024mambainllama,
  title   = {The Mamba in the Llama: Distilling and Accelerating Hybrid Models},
  author  = {Junxiong Wang and Daniele Paliotta and Avner May and Alexander M. Rush and Tri Dao},
  journal = {arXiv preprint arXiv:2408.15237},
  year    = {2024}
}