llama3-8b-dpo-full-beta0.1-lr5e-7

aaaalongaa/llama3-8b-dpo-full-beta0.1-lr5e-7

llama3-8b-dpo-full-beta0.1-lr5e-7 is a machine learning model from aaaalongaa. Use it for the machine learning task on the model card, and read the license before you ship it in a product.

should probably proofread and complete it, then remove this comment. --

Downloads · 30 days

8% of all-time downloads

All-time downloads

Public

Parameters

16.1 GB on disk

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors16.1 GB · 100%

At a glance

Model type: llama
Access: Public
Created: May 11, 2025
Updated: May 11, 2025
SHA: 9dd956b0

Base models

princeton-nlp/Llama-3-Base-8B-SFT

Datasets

HuggingFaceH4/ultrafeedback_binarized

Model card

From the Hugging Face model README

llama3-8b-dpo-full-beta0.1-lr5e-7

This model is a fine-tuned version of princeton-nlp/Llama-3-Base-8B-SFT on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

Loss: 0.4888
Rewards/chosen: -39.1514
Rewards/rejected: -50.6944
Rewards/accuracies: 0.7698
Rewards/margins: 11.5430
Logps/rejected: -50.6944
Logps/chosen: -39.1514
Logits/rejected: 8.3371
Logits/chosen: 7.6560

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: 4
gradient_accumulation_steps: 2
total_train_batch_size: 32
total_eval_batch_size: 32
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
0.6923	0.0523	100	0.6921	-1.0925	-1.3138	0.5952	0.2214	-1.3138	-1.0925	-0.4698	-0.4884
0.6853	0.1047	200	0.6791	-1.2975	-1.7618	0.6389	0.4643	-1.7618	-1.2975	0.8598	0.7327
0.6644	0.1570	300	0.6830	-2.5291	-3.5897	0.6448	1.0606	-3.5897	-2.5291	3.7570	3.4297
0.6271	0.2094	400	0.6433	-3.6518	-5.2447	0.6548	1.5929	-5.2447	-3.6518	3.0076	2.6364
0.6277	0.2617	500	0.5910	-11.3918	-15.1541	0.6984	3.7622	-15.1541	-11.3918	4.5531	4.1724
0.5653	0.3141	600	0.5415	-23.4490	-30.1647	0.7242	6.7157	-30.1647	-23.4490	6.4639	5.9104
0.4946	0.3664	700	0.5268	-37.3159	-46.0423	0.7302	8.7264	-46.0423	-37.3159	7.5803	7.0493
0.504	0.4187	800	0.5158	-30.4668	-39.6835	0.7480	9.2168	-39.6835	-30.4668	7.4862	6.8119
0.5189	0.4711	900	0.5084	-32.7423	-43.1027	0.7421	10.3603	-43.1027	-32.7423	7.7962	7.1047
0.5565	0.5234	1000	0.5024	-37.7911	-48.0086	0.7401	10.2175	-48.0086	-37.7911	8.0538	7.4013
0.4591	0.5758	1100	0.4991	-33.0557	-43.8822	0.7560	10.8265	-43.8822	-33.0557	8.1008	7.4012
0.5096	0.6281	1200	0.4976	-36.8253	-47.5916	0.7599	10.7664	-47.5916	-36.8253	8.3503	7.6805
0.4908	0.6805	1300	0.4928	-35.6927	-47.3111	0.7659	11.6184	-47.3111	-35.6927	8.2122	7.5054
0.4607	0.7328	1400	0.4918	-37.7078	-48.8120	0.7679	11.1043	-48.8120	-37.7078	8.2679	7.5943
0.5121	0.7851	1500	0.4907	-38.0209	-49.3997	0.7718	11.3788	-49.3997	-38.0209	8.2755	7.5918
0.5183	0.8375	1600	0.4902	-38.2788	-49.8110	0.7758	11.5322	-49.8110	-38.2788	8.3217	7.6376
0.4572	0.8898	1700	0.4891	-39.6605	-51.1928	0.7679	11.5323	-51.1928	-39.6605	8.3700	7.6904
0.4925	0.9422	1800	0.4890	-39.0767	-50.6356	0.7698	11.5589	-50.6356	-39.0767	8.3316	7.6491
0.5166	0.9945	1900	0.4889	-39.1533	-50.6911	0.7679	11.5378	-50.6911	-39.1533	8.3381	7.6570

Framework versions

Transformers 4.43.1
Pytorch 2.1.2+cu121
Datasets 2.18.0
Tokenizers 0.19.1

Type: llama
Created: May 11, 2025
Updated: May 11, 2025