This model is a fine-tuned version of /scratch/qu.yang1/dynamic-dpo-v4/outputs/qwen3-8b-base-sft-hh-harmless-4xh200-batch-64-20260417-214452 on the Anthropic/hh-rlhf dataset. It achieves the following results on the evaluation set:

Loss: 0.6201
Beta Dpo/beta: 0.1744
Beta Dpo/loss Margin Mean: 9.2097
Beta Dpo/beta Margin Mean: 1.8984
Beta Dpo/beta Margin Std: 2.6738
Beta Dpo/beta Margin Grad Mean: -0.3317
Beta Dpo/beta Margin Grad Std: 0.2463
Beta Dpo/gap Mean: 8.2231
Beta Dpo/gap Std: 14.9607
Beta Dpo/beta Used Raw: 0.1592
Beta Dpo/beta Used: 0.1744
Beta Dpo/mask Keep Frac: 1.0
Logits/chosen: 2.4073
Logits/rejected: 2.4748

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: 8
eval_batch_size: 8
seed: 42
distributed_type: multi-GPU
num_devices: 4
gradient_accumulation_steps: 2
total_train_batch_size: 64
total_eval_batch_size: 32
optimizer: Use OptimizerNames.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: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Beta Dpo/beta	Beta Dpo/loss Margin Mean	Beta Dpo/beta Margin Mean	Beta Dpo/beta Margin Std	Beta Dpo/beta Margin Grad Mean	Beta Dpo/beta Margin Grad Std	Beta Dpo/gap Mean	Beta Dpo/gap Std	Beta Dpo/beta Used Raw	Beta Dpo/beta Used	Beta Dpo/mask Keep Frac	Logits/chosen	Logits/rejected
1.3421	0.1468	100	0.6724	0.1051	0.4154	0.0454	0.1109	-0.4887	0.0276	0.3297	0.9661	0.1051	0.1051	1.0	1.9351	2.0227
1.1746	0.2937	200	0.5979	0.1463	3.4910	0.5960	1.1231	-0.3984	0.1967	2.7526	6.6589	0.1443	0.1463	1.0	2.4745	2.5521
1.0588	0.4405	300	0.6234	0.1568	7.0255	1.3113	2.0548	-0.3561	0.2381	6.2446	12.2876	0.1469	0.1568	1.0	2.5889	2.6667
1.0387	0.5874	400	0.5861	0.0864	8.3161	0.9103	1.2402	-0.3898	0.1540	9.0513	15.3829	0.0559	0.0864	1.0	2.2982	2.3580
1.3836	0.7342	500	0.6929	0.2278	9.2738	2.4511	3.5401	-0.3173	0.2780	7.2783	15.5833	0.2197	0.2278	1.0	2.3062	2.3664
1.1032	0.8811	600	0.6201	0.1744	9.2097	1.8984	2.6738	-0.3317	0.2463	8.2231	14.9607	0.1592	0.1744	1.0	2.4073	2.4748

Framework versions

Transformers 4.51.0
Pytorch 2.3.1+cu121
Datasets 2.21.0
Tokenizers 0.21.4

jackf857/qwen3-8b-base-beta-dpo-hh-helpful-4xh200-batch-64

qwen3-8b-base-beta-dpo-hh-helpful-4xh200-batch-64-20260418-012645

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

jackf857/
qwen3-8b-base-beta-dpo-hh-helpful-4xh200-batch-64